Init GraphKernel.

- It provides a unified style to express graph and kernel for user.
- It provides a unified IR to represent graph and kernel for developer.
- It breaks the boundary between graph and kernel.
- It provides more opportunities to do compile optimization.

.gitmodules

@@ -13,3 +13,6 @@
 [submodule "graphengine"]
 	path = graphengine
 	url = https://gitee.com/mindspore/graphengine.git
+[submodule "akg"]
+	path = akg
+	url = https://gitee.com/mindspore/akg.git

CMakeLists.txt

@@ -86,10 +86,14 @@ if (ENABLE_GE OR ENABLE_D OR ENABLE_TESTCASES)
     include_directories(${CMAKE_CURRENT_SOURCE_DIR}/graphengine/third_party/fwkacllib/inc/toolchain)
 endif()
 
+if (ENABLE_AKG AND ENABLE_D)
+    add_subdirectory("${CMAKE_SOURCE_DIR}/akg")
+endif()
+
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fvisibility=hidden")
 add_subdirectory(mindspore/ccsrc)
 if (ENABLE_TESTCASES)
     add_subdirectory(tests)
 endif()
 
-include(cmake/package.cmake)
+include(cmake/package.cmake)

akg

@@ -0,0 +1 @@
+Subproject commit c460176523d039c8995f1d71089753725ebc0792

build.sh

@@ -246,6 +246,9 @@ checkopts "$@"
 echo "---------------- mindspore: build start ----------------"
 mkdir -pv "${BUILD_PATH}/package/mindspore/lib"
 git submodule update --init graphengine
+if [[ "X$ENABLE_AKG" = "Xon" ]] && [[ "X$ENABLE_D" = "Xon" ]]; then
+    git submodule update --init --recursive akg
+fi
 
 build_exit()
 {
@@ -308,7 +311,7 @@ build_mindspore()
     if [[ "X$USE_GLOG" = "Xon" ]]; then
         CMAKE_ARGS="${CMAKE_ARGS} -DUSE_GLOG=ON"
     fi
-    if [[ "X$ENABLE_AKG" = "Xon" ]]; then
+    if [[ "X$ENABLE_AKG" = "Xon" ]] && [[ "X$ENABLE_D" = "Xon" ]]; then
         CMAKE_ARGS="${CMAKE_ARGS} -DENABLE_AKG=ON"
     fi
     echo "${CMAKE_ARGS}"

cmake/package.cmake

@@ -236,6 +236,16 @@ if (ENABLE_GPU)
     endif ()
 endif ()
 
+if (ENABLE_D AND ENABLE_AKG)
+    set (AKG_PATH ${CMAKE_SOURCE_DIR}/build/mindspore/akg)
+    install(
+        DIRECTORY
+            ${AKG_PATH}/akg
+        DESTINATION ${INSTALL_PY_DIR}/..
+        COMPONENT mindspore
+    )
+endif ()
+
 if (EXISTS ${CMAKE_SOURCE_DIR}/mindspore/dataset)
     install(
         DIRECTORY ${CMAKE_SOURCE_DIR}/mindspore/dataset

mindspore/_extends/parallel_compile/akg_compiler/__init__.py

@@ -0,0 +1,14 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================

mindspore/_extends/parallel_compile/akg_compiler/compiler.py

@@ -0,0 +1,35 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""Providing akg compile with json"""
+import sys
+def run_compiler(op_json):
+    """
+    Run AKG compiler to compile op with subprocess, if this process of
+    compilation failed, an exception will be raised
+
+    Args:
+        op_json (str): json string of the op
+
+    Returns:
+        None
+    """
+    p = __import__("akg", globals(), locals(), ['ms'], 0)
+    func = getattr(p.ms, "compilewithjson")
+    res = func(op_json)
+    if not res:
+        raise ValueError("Compile error")
+
+if __name__ == "__main__":
+    run_compiler(sys.argv[1])

mindspore/_extends/parallel_compile/akg_compiler/multi_process_compiler.py

@@ -0,0 +1,71 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""Providing multi process compile with json"""
+import os
+import subprocess
+import sys
+from multiprocessing import Pool, cpu_count
+
+
+def _compile_akg_task(*json_strs):
+    """
+    compile func called in single process
+
+    Parameters:
+        json_strs: list. List contains multiple kernel infos, suitable for json compile api.
+    """
+    akg_compiler = os.path.join(os.path.split(
+        os.path.realpath(__file__))[0], "compiler.py")
+    for json_str in json_strs:
+        res = subprocess.run(
+            [sys.executable, akg_compiler, json_str], text=True)
+        if res.returncode != 0:
+            raise ValueError("Failed, args: {}!".format(json_str))
+
+
+def compile_akg_kernel_parallel(json_infos, process, waitime):
+    """
+    compile kernel use multi processes
+
+    Parameters:
+        json_infos: list. list contain kernel info(task id and json str)
+        process: int. processes num
+        waittime: int. max time the function blocked
+
+    Returns:
+        True for all compile success, False for some failed.
+    """
+    if not isinstance(json_infos, list):
+        raise ValueError("json_infos must be a list")
+    if not isinstance(process, int):
+        raise ValueError("process must be a num")
+    if not isinstance(waitime, int):
+        raise ValueError("waittime must be a num")
+
+    if process == 0 and json_infos:
+        process = 1
+
+    cpu_proc_num = cpu_count()
+    max_proc_num = 16
+    process = min([cpu_proc_num, max_proc_num, process])
+
+    args = [[] for _ in range(process)]
+    for p, info in enumerate(json_infos):
+        args[p % process].append(info)
+
+    with Pool(processes=process) as pool:
+        res = pool.starmap_async(_compile_akg_task, args)
+        res.get(timeout=waitime)
+    return True

mindspore/_extends/parallel_compile/multi_compiler.py

@@ -1,107 +0,0 @@
-# Copyright 2020 Huawei Technologies Co., Ltd
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ============================================================================
-"""Providing multi process compile with json"""
-import json
-import math
-import os
-import subprocess
-import sys
-from multiprocessing import Pool
-
-
-def _compiletask(platform, *jsons):
-    """
-        compile func called in single process
-
-        Parameters:
-            platform: str. AKG platform or TBE platform
-            *jsons: str. json str contain kernel info, suitable for json compile
-                    api
-
-        """
-    if platform == "AKG":
-        p = __import__("_akg", globals(), locals(), ['ms'], 0)
-        func = getattr(p.ms, "compilewithjson")
-        for json_item in jsons:
-            res = func(json_item)
-            if not res:
-                raise ValueError("Compile error")
-    if platform == "TBE":
-        tbe_compiler = os.path.join(os.path.split(os.path.realpath(__file__))[0], "tbe_compiler", "compiler.py")
-        for json_item in jsons:
-            res = subprocess.run([sys.executable, tbe_compiler], input=json_item, text=True)
-            if res.returncode != 0:
-                raise ValueError("Tbe compile error")
-
-
-def compilekernelparallel(jsons, process, waitime):
-    """
-    compile kernel use multi processes
-
-    Parameters:
-        jsons: list. json str list contain kernel info
-        process: int. processes num
-        waittime: int. max time the function blocked
-    """
-    if not isinstance(jsons, list):
-        raise ValueError("jsons must be a list")
-    if not isinstance(process, int):
-        raise ValueError("process must be a num")
-    if not isinstance(waitime, int):
-        raise ValueError("waittime must be a num")
-
-    jsons_akg = []
-    jsons_tbe = []
-    for json_ in jsons:
-        j = json.loads(json_)
-        if j["platform"] == "TBE":
-            jsons_tbe.append(json_)
-            continue
-        if j["platform"] == "AKG":
-            jsons_akg.append(json_)
-            continue
-        raise RuntimeError(
-            "not support this platform {0}".format(j["platform"]))
-    if jsons_akg:
-        process_akg = math.floor(len(jsons)/len(jsons_akg)*process)
-    else:
-        process_akg = 0
-
-    if process_akg == 0 and jsons_akg:
-        process_akg = 1
-    process_tbe = process-process_akg
-    if process_tbe == 0 and jsons_tbe:
-        process_tbe = 1
-        raise RuntimeWarning("we add a process for compile more operator")
-
-    args = [[] for _ in range(process_akg+process_tbe)]
-    args_lens = len(args)
-    for p in range(args_lens):
-        if p < process_tbe:
-            args[p].append("TBE")
-        else:
-            args[p].append("AKG")
-    jsons_tbe_lens = len(jsons_tbe)
-    for p in range(jsons_tbe_lens):
-        args[p % process_tbe].append(jsons_tbe[p])
-    jsons_akg_lens = len(jsons_akg)
-    for p in range(jsons_akg_lens):
-        args[process-p % process_akg-1].append(jsons_akg[p])
-    for p in range(args_lens):
-        args[p] = tuple(args[p])
-    with Pool(processes=process) as pool:
-        res = pool.starmap_async(_compiletask, args)
-        res.get(timeout=waitime)
-    return True

mindspore/ccsrc/CMakeLists.txt

@@ -39,7 +39,7 @@ if(ENABLE_GPU)
             "device/gpu/*.cu"
             "kernel/gpu/*.cu"
             "kernel/akg/gpu/*.cc"
-            "kernel/akg/akgkernelbuild.cc"
+            "kernel/akg/akg_kernel_build.cc"
             "kernel/akg/akg_kernel_attrs_process.cc"
             )
 

mindspore/ccsrc/common/trans.cc

@@ -428,6 +428,10 @@ std::vector<size_t> TransShapeToDevice(const std::vector<size_t> &shape, const s
   auto temp_shape = shape;
   std::vector<size_t> device_shape;
   if (format == kOpFormat_FRAC_NZ) {
+    if (shape.size() == 1 && (shape[0] == 1 || shape[0] % kCubeSize == 0)) {
+      // For [1] and [1024] shape we can trait it as NZ shape
+      return shape;
+    }
     if (shape.size() < 2) {
       MS_LOG(EXCEPTION) << "Format" << format << " is not support shape " << shape.size();
     } else {

mindspore/ccsrc/debug/anf_ir_dump.cc

@@ -111,9 +111,15 @@ void DumpGlobalInfoEntry(const FuncGraphPtr &graph, std::ostringstream &buffer)
   }
 
   buffer << "#IR entry      : @" << graph->ToString() << "." << graph->debug_info()->get_id() << std::endl;
-  buffer << "#flags         :" << std::endl;
-  for (const auto &flag : graph->flags()) {
-    buffer << flag.first << " : " << flag.second << std::endl;
+  buffer << "#attrs         :" << std::endl;
+  for (const auto &attr : graph->attrs()) {
+    buffer << attr.first << " : ";
+    if (attr.second->isa<BoolImm>()) {
+      buffer << GetValue<bool>(attr.second);
+    } else if (attr.second->isa<StringImm>()) {
+      buffer << GetValue<std::string>(attr.second);
+    }
+    buffer << std::endl;
   }
 }
 
@@ -417,10 +423,16 @@ void DumpSubgraph(const OrderedMap<FuncGraphPtr, std::shared_ptr<SubGraphIRInfo>
   fout << std::endl;
 
   for (const auto &sg : *sub_graphs) {
-    fout << "subgraph flag:" << std::endl;
+    fout << "subgraph attr:" << std::endl;
     MS_EXCEPTION_IF_NULL(sg.first);
-    for (const auto &flag : sg.first->flags()) {
-      fout << flag.first << " : " << flag.second << std::endl;
+    for (const auto &attr : sg.first->attrs()) {
+      fout << attr.first << " : ";
+      if (attr.second->isa<BoolImm>()) {
+        fout << GetValue<bool>(attr.second);
+      } else if (attr.second->isa<StringImm>()) {
+        fout << GetValue<std::string>(attr.second);
+      }
+      fout << std::endl;
     }
     fout << "subgraph @" << sg.first->ToString() << ".";
     fout << sg.first->debug_info()->get_id() << "(";

mindspore/ccsrc/device/ascend/ascend_stream_assign.cc

@@ -548,9 +548,15 @@ void AscendStreamAssign::GetNeedActiveStreams(const shared_ptr<session::KernelGr
   for (size_t i = 0; i < cnode_ptr_list.size(); ++i) {
     cur_cnode_ptr = cnode_ptr_list[i];
     MS_EXCEPTION_IF_NULL(cur_cnode_ptr);
+    ValuePtr value_ptr = nullptr;
     auto primitive = AnfAlgo::GetCNodePrimitive(cur_cnode_ptr);
-    MS_EXCEPTION_IF_NULL(primitive);
-    auto value_ptr = primitive->GetAttr(kStreamNeedActivedFirst);
+    if (primitive != nullptr) {
+      value_ptr = primitive->GetAttr(kStreamNeedActivedFirst);
+    } else {
+      auto func_graph = AnfAlgo::GetCNodeFuncGraphPtr(cur_cnode_ptr);
+      MS_EXCEPTION_IF_NULL(func_graph);
+      value_ptr = func_graph->get_attr(kStreamNeedActivedFirst);
+    }
     if (value_ptr == nullptr) {
       continue;
     }

mindspore/ccsrc/device/ascend/kernel_build_ascend.cc

@@ -26,10 +26,12 @@
 #include "kernel/kernel.h"
 #include "kernel/tbe/tbe_kernel_build.h"
 #include "kernel/tbe/tbe_kernel_parallel_build.h"
+#include "kernel/akg/ascend/akg_ascend_kernel_build.h"
 #include "kernel/aicpu/aicpu_kernel_build.h"
 #include "kernel/hccl/hccl_kernel_build.h"
 #include "kernel/rts/rt_kernel_build.h"
 #include "kernel/tbe/tbe_utils.h"
+#include "kernel/common_utils.h"
 #include "operator/ops.h"
 #include "session/anf_runtime_algorithm.h"
 #include "./common.h"
@@ -91,6 +93,7 @@ static bool KernelPreBuildParallelCompile(const mindspore::session::KernelGraph
 static bool KernelBuildParallelCompile(const mindspore::session::KernelGraph *kernel_graph_ptr) {
   MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
   std::vector<AnfNodePtr> tbe_nodes;
+  std::vector<AnfNodePtr> akg_nodes;
   std::vector<AnfNodePtr> other_nodes;
   for (const auto &anf_node : kernel_graph_ptr->execution_order()) {
     MS_EXCEPTION_IF_NULL(anf_node);
@@ -105,19 +108,26 @@ static bool KernelBuildParallelCompile(const mindspore::session::KernelGraph *ke
         }
         break;
       }
+      case KernelType::AKG_KERNEL: {
+        akg_nodes.push_back(anf_node);
+        break;
+      }
       default: {
         other_nodes.push_back(anf_node);
         break;
       }
     }
   }
-  bool ret = kernel::TbeOpParallelBuild(tbe_nodes);
+  bool tbe_ret = kernel::TbeOpParallelBuild(tbe_nodes);
+  bool akg_ret = kernel::AkgAscendKernelParallelBuild(akg_nodes);
+  auto bin_map = kernel::tbe::KernelMeta::GetInstance();
+  (void)bin_map->ReadIndex(kernel::kCceKernelMeta);
   for (const auto &anf_node : other_nodes) {
     kernel::KernelModPtr kernel_mod_ptr = SerialCompileImpl(anf_node);
     MS_EXCEPTION_IF_NULL(kernel_mod_ptr);
     AnfAlgo::SetKernelMod(kernel_mod_ptr, anf_node.get());
   }
-  return ret;
+  return tbe_ret && akg_ret;
 }
 
 static std::vector<int> CalCleanZerosSize(const CNodePtr &pre_node) {
@@ -234,7 +244,7 @@ void KernelBuildPreprocess(mindspore::session::KernelGraph *kernel_graph) {
   for (const auto &anf_node : kernel_graph->execution_order()) {
     std::string apply_function_name = AnfAlgo::GetCNodeName(anf_node);
     if (apply_function_name == prim::kPrimMaxPoolGrad->name() &&
-        AnfAlgo::GetKernelType(anf_node) == KernelType::AUTO_DIFF_KERNEL) {
+        AnfAlgo::GetKernelType(anf_node) == KernelType::AKG_KERNEL) {
       auto clear_zero_prim = std::make_shared<Primitive>(kClearZeroOpName);
       MS_EXCEPTION_IF_NULL(clear_zero_prim);
       auto new_value_node = NewValueNode(clear_zero_prim);

mindspore/ccsrc/device/ascend/kernel_select_ascend.cc

@@ -15,16 +15,27 @@
  */
 
 #include "device/ascend/kernel_select_ascend.h"
+
 #include <string>
 #include <vector>
 #include <memory>
 #include <utility>
+#include <algorithm>
 #include <map>
-#include "kernel/oplib/oplib.h"
-#include "kernel/kernel_query.h"
-#include "session/anf_runtime_algorithm.h"
-#include "utils/context/ms_context.h"
+#include <unordered_map>
+#include <unordered_set>
+
+#include "common/utils.h"
 #include "debug/anf_ir_dump.h"
+#include "operator/ops.h"
+#include "ir/func_graph.h"
+#include "utils/context/ms_context.h"
+#include "session/anf_runtime_algorithm.h"
+#include "device/kernel_info.h"
+#include "kernel/common_utils.h"
+#include "kernel/kernel_query.h"
+#include "kernel/oplib/oplib.h"
+#include "kernel/kernel_build_info.h"
 
 namespace mindspore {
 namespace device {
@@ -121,12 +132,23 @@ void UpdateCurMatchCounts(const kernel::KernelBuildInfo &kernel_build_info, cons
   }
   auto pri_match_format = GetPriorityMatchFormat(kernel_node);
   for (size_t input_index = 0; input_index < AnfAlgo::GetInputTensorNum(kernel_node); ++input_index) {
+    auto input_anf_node = kernel_node->input(input_index + 1);
+    // we do not take ValueNode into consideration in graph kernel.
+    if (kernel_build_info.kernel_type() == KernelType::AKG_KERNEL) {
+      if (input_anf_node->isa<ValueNode>() && AnfAlgo::GetOutputDeviceDataType(input_anf_node, 0) == kTypeUnknown) {
+        continue;
+      }
+    }
     auto base_score = AnfAlgo::IsFeatureMapInput(kernel_node, input_index) ? kFeatureMapBaseScore : kWegihtBaseScore;
     if (kernel_build_info.GetInputFormat(input_index) == AnfAlgo::GetPrevNodeOutputFormat(kernel_node, input_index)) {
       (*cur_kernelinfo_match_counts)[MATCH_FORMAT_COUNT] += base_score;
     }
-    if (kernel_build_info.GetInputDeviceType(input_index) ==
-        AnfAlgo::GetPrevNodeOutputDeviceDataType(kernel_node, input_index)) {
+    // we match output fix precision first.
+    auto prev_device_type = AnfAlgo::GetPrevNodeOutputPrecision(kernel_node, input_index);
+    if (prev_device_type == kTypeUnknown) {
+      prev_device_type = AnfAlgo::GetPrevNodeOutputDeviceDataType(kernel_node, input_index);
+    }
+    if (kernel_build_info.GetInputDeviceType(input_index) == prev_device_type) {
       (*cur_kernelinfo_match_counts)[MATCH_DTYPE_COUNT] += base_score;
     }
     if (kernel_build_info.GetInputFormat(input_index) == pri_match_format) {
@@ -146,41 +168,6 @@ void UpdateCurMatchCounts(const kernel::KernelBuildInfo &kernel_build_info, cons
   }
 }
 
-void SetTensorDeviceInfo(const kernel::KernelBuildInfo &selected_kernel_info, const CNodePtr &kernel_node) {
-  MS_EXCEPTION_IF_NULL(kernel_node);
-  for (size_t input_index = 0; input_index < AnfAlgo::GetInputTensorNum(kernel_node); ++input_index) {
-    auto input_kernel_node = AnfAlgo::GetInputNode(kernel_node, input_index);
-    MS_EXCEPTION_IF_NULL(input_kernel_node);
-    auto input_with_index = AnfAlgo::VisitKernel(input_kernel_node, 0);
-    MS_EXCEPTION_IF_NULL(input_with_index.first);
-    auto real_input_node = input_with_index.first;
-    if (real_input_node->isa<CNode>()) {
-      continue;
-    }
-    std::shared_ptr<kernel::KernelBuildInfo::KernelBuildInfoBuilder> builder =
-      std::make_shared<kernel::KernelBuildInfo::KernelBuildInfoBuilder>();
-    bool is_ref = false;
-    auto op_info = mindspore::kernel::OpLib::FindOp(AnfAlgo::GetCNodeName(kernel_node), kernel::kTBE);
-    if (op_info != nullptr) {
-      is_ref = op_info->is_ref();
-    }
-    auto ms_context = MsContext::GetInstance();
-    MS_EXCEPTION_IF_NULL(ms_context);
-    if (ms_context->execution_mode() == kPynativeMode &&
-        AnfAlgo::GetOutputDeviceDataType(real_input_node, 0) != kTypeUnknown) {
-      continue;
-    }
-    // we set special device info of a input tensor.
-    if (AnfAlgo::GetOutputDeviceDataType(real_input_node, 0) == kTypeUnknown || is_ref) {
-      std::vector<std::string> output_format = {selected_kernel_info.GetInputFormat(input_index)};
-      builder->SetOutputsFormat(output_format);
-      std::vector<TypeId> output_type = {AnfAlgo::GetInputDeviceDataType(kernel_node, input_index)};
-      builder->SetOutputsDeviceType(output_type);
-      AnfAlgo::SetSelectKernelBuildInfo(builder->Build(), real_input_node.get());
-    }
-  }
-}
-
 void AddSupportMixedPrecisionDataTypeIndex(TypeId data_type, std::vector<int> *support_index) {
   MS_EXCEPTION_IF_NULL(support_index);
   int index = kUnSupportMixedDataTypeIndex;
@@ -467,6 +454,51 @@ std::vector<std::shared_ptr<kernel::KernelBuildInfo>> FilterRaisedOrReducePrecis
 }
 }  // namespace
 
+void SetTensorDeviceInfo(const kernel::KernelBuildInfo &selected_kernel_info, const CNodePtr &kernel_node) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  for (size_t input_index = 0; input_index < AnfAlgo::GetInputTensorNum(kernel_node); ++input_index) {
+    auto input_kernel_node = AnfAlgo::GetInputNode(kernel_node, input_index);
+    MS_EXCEPTION_IF_NULL(input_kernel_node);
+    auto input_with_index = AnfAlgo::VisitKernel(input_kernel_node, 0);
+    MS_EXCEPTION_IF_NULL(input_with_index.first);
+    auto real_input_node = input_with_index.first;
+    if (real_input_node->isa<CNode>()) {
+      continue;
+    }
+    if (real_input_node->isa<Parameter>() && !AnfAlgo::IsParameterWeight(real_input_node->cast<ParameterPtr>())) {
+      continue;
+    }
+    auto builder = std::make_shared<kernel::KernelBuildInfo::KernelBuildInfoBuilder>();
+    if (IsValueNode<tensor::Tensor>(input_kernel_node) &&
+        AnfAlgo::GetOutputDeviceDataType(input_kernel_node, 0) == kTypeUnknown) {
+      std::vector<std::string> output_format = {selected_kernel_info.GetInputFormat(input_index)};
+      builder->SetOutputsFormat(output_format);
+      std::vector<TypeId> output_type = {selected_kernel_info.GetInputDeviceType(input_index)};
+      builder->SetOutputsDeviceType(output_type);
+      AnfAlgo::SetSelectKernelBuildInfo(builder->Build(), input_kernel_node.get());
+      continue;
+    }
+    // we set special device info of a input tensor.
+    bool is_ref = false;
+    auto op_info = kernel::OpLib::FindOp(AnfAlgo::GetCNodeName(kernel_node), kernel::kTBE);
+    if (op_info != nullptr) {
+      is_ref = op_info->is_ref();
+    }
+    MS_EXCEPTION_IF_NULL(MsContext::GetInstance());
+    if (MsContext::GetInstance()->execution_mode() == kPynativeMode &&
+        AnfAlgo::GetOutputDeviceDataType(real_input_node, 0) != kTypeUnknown) {
+      continue;
+    }
+    if (AnfAlgo::GetOutputDeviceDataType(real_input_node, 0) == kTypeUnknown || is_ref) {
+      std::vector<std::string> output_format = {selected_kernel_info.GetInputFormat(input_index)};
+      builder->SetOutputsFormat(output_format);
+      std::vector<TypeId> output_type = {selected_kernel_info.GetInputDeviceType(input_index)};
+      builder->SetOutputsDeviceType(output_type);
+      AnfAlgo::SetSelectKernelBuildInfo(builder->Build(), real_input_node.get());
+    }
+  }
+}
+
 KernelSelectStatus SetMatchedKernelInfo(const CNodePtr &kernel_node,
                                         const std::vector<std::shared_ptr<kernel::KernelBuildInfo>> &kernel_info_list) {
   MS_EXCEPTION_IF_NULL(kernel_node);
@@ -498,11 +530,17 @@ KernelSelectStatus SetMatchedKernelInfo(const CNodePtr &kernel_node,
   return select_status;
 }
 
-KernelSelectStatus SelectKernelInfo(const CNodePtr &kernel_node) {
+KernelSelectStatus SelectKernelInfo(const CNodePtr &kernel_node, KernelType kernel_type) {
   std::vector<std::shared_ptr<kernel::KernelBuildInfo>> kernel_info_list;
   std::vector<std::shared_ptr<kernel::KernelBuildInfo>> aicpu_kernel_info_list;
   MS_EXCEPTION_IF_NULL(kernel_node);
-  kernel::KernelQuery(kernel_node, &kernel_info_list);
+  if (AnfAlgo::IsGraphKernel(kernel_node)) {
+    auto func_graph = GetValueNode<FuncGraphPtr>(kernel_node->input(kAnfPrimitiveIndex));
+    MS_EXCEPTION_IF_NULL(func_graph);
+    SelectGraphKernelInfo(kernel_node, func_graph);
+    return kStatusAllMatched;
+  }
+  kernel::KernelQuery(kernel_node, &kernel_info_list, kernel_type);
   auto select_status = SetMatchedKernelInfo(kernel_node, kernel_info_list);
   // If aicore not find valid kernel info reloading aicpu kernel info list to find it
   if (select_status == kNoMatched) {

mindspore/ccsrc/device/ascend/kernel_select_ascend.h

@@ -27,7 +27,10 @@ enum KernelSelectStatus {
   kStatusReducePrecision = 1,
   kStatusRaisePrecision = 2,
 };
-KernelSelectStatus SelectKernelInfo(const CNodePtr &kernel_node);
+KernelSelectStatus SelectKernelInfo(const CNodePtr &kernel_node,
+                                    KernelType kernel_type = KernelType::UNKNOWN_KERNEL_TYPE);
+void SetTensorDeviceInfo(const kernel::KernelBuildInfo &selected_kernel_info, const CNodePtr &kernel_node);
+void SelectGraphKernelInfo(const CNodePtr &kernel_node, const FuncGraphPtr &func_graph);
 }  // namespace ascend
 }  // namespace device
 }  // namespace mindspore

mindspore/ccsrc/device/ascend/kernel_select_graph_kernel.cc

@@ -0,0 +1,516 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "device/ascend/kernel_select_ascend.h"
+#include "session/anf_runtime_algorithm.h"
+#include "device/kernel_info.h"
+#include "ir/func_graph.h"
+#include "kernel/common_utils.h"
+#include "kernel/kernel_query.h"
+#include "kernel/kernel_build_info.h"
+
+namespace mindspore {
+namespace device {
+namespace ascend {
+
+TypeId GetPrimitivePrecision(const CNodePtr &cnode) {
+  auto primitive = AnfAlgo::GetCNodePrimitive(cnode);
+  MS_EXCEPTION_IF_NULL(primitive);
+
+  TypeId except_type = kTypeUnknown;
+  if (primitive->GetAttr(kAttrFixPrecision) != nullptr) {
+    auto strExceptDtype = GetValue<std::string>(primitive->GetAttr(kAttrFixPrecision));
+    if (strExceptDtype == "float16") {
+      except_type = kNumberTypeFloat16;
+    } else if (strExceptDtype == "float32") {
+      except_type = kNumberTypeFloat32;
+    } else {
+      MS_LOG(EXCEPTION) << "The fix precision must be float16 or float32, but got" << strExceptDtype;
+    }
+  }
+
+  return except_type;
+}
+
+void ResetKernelBuildInfo(const CNodePtr &kernel_node) {
+  size_t input_num = AnfAlgo::GetInputTensorNum(kernel_node);
+  for (size_t input_index = 0; input_index < input_num; ++input_index) {
+    auto input_kernel_node = AnfAlgo::GetInputNode(kernel_node, input_index);
+    MS_EXCEPTION_IF_NULL(input_kernel_node);
+    auto kernel_with_index = AnfAlgo::VisitKernel(input_kernel_node, 0);
+    if (!kernel::IsWeightBoundary(kernel_with_index.first)) {
+      continue;
+    }
+    // reset format and dtype.
+    kernel::KernelBuildInfo::KernelBuildInfoBuilder builder;
+    builder.SetOutputsFormat(std::vector<std::string>{kOpFormat_DEFAULT});
+    builder.SetOutputsDeviceType(std::vector<TypeId>{kTypeUnknown});
+    AnfAlgo::SetSelectKernelBuildInfo(builder.Build(), input_kernel_node.get());
+  }
+}
+
+void UpdateKernelInfo(const std::vector<AnfNodePtr> &node_list) {
+  for (size_t i = 0; i < node_list.size(); ++i) {
+    // select nodes in subgraph.
+    auto anf_node = node_list[i];
+    MS_EXCEPTION_IF_NULL(anf_node);
+    auto cnode = anf_node->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    auto fix_precision_type = GetPrimitivePrecision(cnode);
+    if (fix_precision_type != kTypeUnknown) {
+      std::vector<std::shared_ptr<kernel::KernelBuildInfo>> kernel_info_list;
+      kernel::KernelQuery(cnode, &kernel_info_list, KernelType::AKG_KERNEL);
+
+      for (size_t index = 0; index < kernel_info_list.size(); ++index)
+        // only math the first input
+        if (kernel_info_list[index]->GetInputDeviceType(0) == fix_precision_type &&
+            kernel_info_list[index]->GetInputFormat(0) == AnfAlgo::GetPrevNodeOutputFormat(cnode, 0) &&
+            AnfAlgo::GetInputDeviceDataType(cnode, 0) != fix_precision_type) {
+          auto selected_kernel_info_ptr = kernel_info_list[index];
+          ResetKernelBuildInfo(cnode);
+          AnfAlgo::SetSelectKernelBuildInfo(selected_kernel_info_ptr, cnode.get());
+          SetTensorDeviceInfo(*selected_kernel_info_ptr, cnode);
+          break;
+        }
+    }
+  }
+}
+
+bool CanConvertDefaultShapeToNZ(const std::vector<size_t> &shape) {
+  for (size_t i = 1; i <= shape.size(); ++i) {
+    if (i > 2) {
+      break;
+    }
+    if (shape[shape.size() - i] != 1 && shape[shape.size() - i] % kCubeSize != 0) {
+      return false;
+    }
+  }
+  return true;
+}
+
+std::vector<int> DefaultToFracNZAxis(const std::vector<size_t> &ori_shape, const std::vector<int> &axis) {
+  std::vector<int> frac_nz_axis = axis;
+  auto shape_len = ori_shape.size();
+  for (size_t i = 0; i < axis.size(); ++i) {
+    auto axis_idx = (frac_nz_axis[i] + shape_len) % shape_len;
+    if (axis_idx == shape_len - 1) {
+      frac_nz_axis[i] = axis_idx - 1;
+      frac_nz_axis.push_back(axis_idx + 2);
+    } else if (axis_idx == shape_len - 2) {
+      frac_nz_axis[i] = axis_idx + 1;
+      frac_nz_axis.push_back(axis_idx + 2);
+    } else {
+      frac_nz_axis[i] = axis_idx;
+    }
+  }
+  return frac_nz_axis;
+}
+
+std::vector<size_t> GetReducedFracNZShape(const std::vector<size_t> &ori_shape, const std::vector<int> &axis,
+                                          bool keep_dims) {
+  std::vector<size_t> result;
+  std::set<size_t> positive_idx;
+  for (const auto &a : axis) {
+    positive_idx.insert(a >= 0 ? a : ori_shape.size() + a);
+  }
+  for (size_t i = 0; i < ori_shape.size(); ++i) {
+    if (positive_idx.count(i) == 0) {
+      result.push_back(ori_shape[i]);
+    } else if (keep_dims) {
+      result.push_back(1);
+    }
+  }
+  return result;
+}
+
+void UpdateFracNZReduceOp(const CNodePtr &cnode) {
+  MS_EXCEPTION_IF_NULL(cnode);
+  auto input_format = AnfAlgo::GetPrevNodeOutputFormat(cnode, 0);
+  if (input_format == kOpFormat_FRAC_NZ) {
+    // Clone primitive to modify it
+    auto prim = GetCNodePrimitive(cnode);
+    auto new_prim = std::make_shared<Primitive>(*prim);
+    auto new_prim_node = NewValueNode(new_prim);
+    cnode->set_input(0, new_prim_node);
+
+    auto axis_value = new_prim->GetAttr(kAttrAxis);
+    std::vector<int> default_axis;
+    if (axis_value->isa<ValueList>()) {
+      auto value_list = dyn_cast<ValueList>(axis_value);
+      for (const auto &item : value_list->value()) {
+        if (item->isa<Int32Imm>()) {
+          default_axis.push_back(GetValue<int32_t>(item));
+        }
+      }
+    } else if (axis_value->isa<ValueTuple>()) {
+      auto value_tuple = dyn_cast<ValueTuple>(axis_value);
+      for (const auto &item : value_tuple->value()) {
+        if (item->isa<Int32Imm>()) {
+          default_axis.push_back(GetValue<int32_t>(item));
+        }
+      }
+    } else {
+      MS_LOG(ERROR) << "Axis attr type is not correct!";
+    }
+    auto infer_shape = AnfAlgo::GetPrevNodeOutputInferShape(cnode, 0);
+    std::vector<int> frac_nz_axis = DefaultToFracNZAxis(infer_shape, default_axis);
+    AnfAlgo::SetNodeAttr(kAttrAxis, MakeValue<std::vector<int>>(frac_nz_axis), cnode);
+    auto output_shape = AnfAlgo::GetOutputInferShape(cnode, 0);
+    if (output_shape.size() == 1) {
+      AnfAlgo::SetNodeAttr(kAttrOutputDefault, MakeValue<bool>(true), cnode);
+    }
+  }
+}
+
+void GetDefaultFormat(const CNodePtr &kernel_node, std::string *default_format, bool *use_same_format) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  MS_EXCEPTION_IF_NULL(default_format);
+  MS_EXCEPTION_IF_NULL(use_same_format);
+  std::unordered_map<std::string, size_t> all_input_formats;
+  size_t input_num = AnfAlgo::GetInputTensorNum(kernel_node);
+  for (size_t i = 0; i < input_num; ++i) {
+    auto input_kernel_node = AnfAlgo::VisitKernel(kernel_node->input(i + 1), 0).first;
+    MS_EXCEPTION_IF_NULL(input_kernel_node);
+    if (!input_kernel_node->isa<Parameter>()) {
+      auto pre_format = AnfAlgo::GetPrevNodeOutputFormat(kernel_node, i);
+      ++all_input_formats[pre_format];
+      continue;
+    }
+    auto para = input_kernel_node->cast<ParameterPtr>();
+    MS_EXCEPTION_IF_NULL(para);
+    if (AnfAlgo::GetOutputDeviceDataType(para, 0) != kTypeUnknown) {
+      auto pre_format = AnfAlgo::GetOutputFormat(para, 0);
+      ++all_input_formats[pre_format];
+      continue;
+    }
+    *use_same_format = false;
+  }
+
+  if (all_input_formats.empty()) {
+    // all inputs are parameter.
+    *default_format = kOpFormat_NC1HWC0;
+  } else {
+    std::vector<std::pair<std::string, size_t>> pairs;
+    for (auto iter = all_input_formats.begin(); iter != all_input_formats.end(); ++iter) {
+      pairs.push_back(std::make_pair(iter->first, iter->second));
+    }
+    auto cmp_func = [](const std::pair<std::string, size_t> &a, const std::pair<std::string, size_t> &b) {
+      if (a.second != b.second) {
+        return a.second > b.second;
+      } else if (a.first == kOpFormat_DEFAULT) {
+        return a.second + 1 > b.second;
+      } else if (b.first == kOpFormat_DEFAULT) {
+        return a.second > b.second + 1;
+      }
+      return a.second > b.second;
+    };
+    std::sort(pairs.begin(), pairs.end(), cmp_func);
+    *default_format = pairs.begin()->first;
+  }
+
+  for (size_t i = 0; i < input_num; ++i) {
+    auto input_kernel_node = AnfAlgo::VisitKernel(kernel_node->input(i + 1), 0).first;
+    MS_EXCEPTION_IF_NULL(input_kernel_node);
+    if (!input_kernel_node->isa<Parameter>() ||
+        AnfAlgo::GetOutputDeviceDataType(input_kernel_node, 0) != kTypeUnknown) {
+      continue;
+    }
+    auto weight_infer_shape = AnfAlgo::GetOutputInferShape(input_kernel_node, 0);
+    if (weight_infer_shape.size() < 2 && *default_format == kOpFormat_FRAC_NZ) {
+      *default_format = kOpFormat_DEFAULT;
+      *use_same_format = true;
+      break;
+    }
+  }
+}
+
+void UpdateGraphKernelInputsKernelInfo(const CNodePtr &kernel_node, const std::vector<AnfNodePtr> &input_list,
+                                       const std::string &default_format, bool use_same_format,
+                                       std::vector<std::string> *graph_input_format,
+                                       std::vector<TypeId> *graph_input_type) {
+  MS_EXCEPTION_IF_NULL(graph_input_format);
+  MS_EXCEPTION_IF_NULL(graph_input_type);
+  // We set same format to all inputs of graph kernel subgraph, and process this latter.
+  // We set dtype to inputs of graph kernel subgraph same as infer dtypes.
+  size_t input_num = AnfAlgo::GetInputTensorNum(kernel_node);
+  for (size_t i = 0; i < input_num; ++i) {
+    auto input_kernel_node = AnfAlgo::VisitKernel(kernel_node->input(i + 1), 0).first;
+    MS_EXCEPTION_IF_NULL(input_kernel_node);
+    if (use_same_format) {
+      bool can_convert = true;
+      if (default_format == kOpFormat_FRAC_NZ) {
+        auto infer_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, i);
+        if (!CanConvertDefaultShapeToNZ(infer_shape)) {
+          MS_LOG(WARNING) << "Shape can't be converted to frac nz shape, so use default format instead";
+          can_convert = false;
+        }
+      }
+      if (can_convert) {
+        graph_input_format->push_back(default_format);
+      } else {
+        graph_input_format->push_back(kOpFormat_DEFAULT);
+      }
+      graph_input_type->push_back(AnfAlgo::GetPrevNodeOutputDeviceDataType(kernel_node, i));
+      continue;
+    }
+
+    if (!input_kernel_node->isa<Parameter>()) {
+      // subgraph parameter from output of other nodes.
+      graph_input_format->push_back(AnfAlgo::GetPrevNodeOutputFormat(kernel_node, i));
+      graph_input_type->push_back(AnfAlgo::GetPrevNodeOutputDeviceDataType(kernel_node, i));
+      continue;
+    }
+
+    auto para = input_kernel_node->cast<ParameterPtr>();
+    MS_EXCEPTION_IF_NULL(para);
+    if (AnfAlgo::GetOutputDeviceDataType(para, 0) != kTypeUnknown) {
+      // parameter already selected.
+      graph_input_format->push_back(AnfAlgo::GetOutputFormat(para, 0));
+      graph_input_type->push_back(AnfAlgo::GetOutputDeviceDataType(para, 0));
+      continue;
+    }
+
+    // weight parameter.
+    graph_input_format->push_back(default_format);
+    graph_input_type->push_back(AnfAlgo::GetOutputInferDataType(input_kernel_node, 0));
+  }
+
+  for (size_t i = 0; i < input_num; ++i) {
+    kernel::KernelBuildInfo::KernelBuildInfoBuilder builder;
+    std::vector<std::string> outputs_format = {(*graph_input_format)[i]};
+    std::vector<TypeId> outputs_device_type = {(*graph_input_type)[i]};
+    builder.SetOutputsFormat(outputs_format);
+    builder.SetOutputsDeviceType(outputs_device_type);
+    AnfAlgo::SetSelectKernelBuildInfo(builder.Build(), input_list[i].get());
+  }
+}
+
+void UpdateEquivFormat(const std::vector<std::pair<AnfNodePtr, size_t>> &output_index,
+                       const std::vector<AnfNodePtr> &node_list, const FuncGraphPtr &func_graph,
+                       const FuncGraphManagerPtr &mng) {
+  MS_EXCEPTION_IF_NULL(mng);
+  for (size_t i = 0; i < node_list.size(); ++i) {
+    // select nodes in subgraph.
+    auto anf_node = node_list[i];
+    MS_EXCEPTION_IF_NULL(anf_node);
+    auto cnode = anf_node->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    cnode->set_kernel_info(std::make_shared<device::KernelInfo>());
+    SelectKernelInfo(cnode, KernelType::AKG_KERNEL);
+    // Update ReduceSum
+    if (!IsPrimitiveCNode(cnode, prim::kPrimReduceSum)) {
+      continue;
+    }
+    UpdateFracNZReduceOp(cnode);
+    // If ReduceSum's output is 1d and not Default format, convert it to Default format
+    auto out_format = AnfAlgo::GetOutputFormat(cnode, 0);
+    if (out_format == kOpFormat_DEFAULT || !AnfAlgo::HasNodeAttr(kAttrOutputDefault, cnode)) {
+      continue;
+    }
+    auto infer_shape = AnfAlgo::GetOutputInferShape(cnode, 0);
+    // Insert EquivFormat node, then select kernel info again
+    std::vector<AnfNodePtr> trans_inputs;
+    trans_inputs.push_back(NewValueNode(prim::kPrimEquivFormat));
+    trans_inputs.push_back(cnode);
+    CNodePtr trans_node = func_graph->NewCNode(trans_inputs);
+    AnfAlgo::SetOutputInferTypeAndShape({AnfAlgo::GetPrevNodeOutputInferDataType(cnode, 0)},
+                                        {AnfAlgo::GetOutputInferShape(cnode, 0)}, trans_node.get());
+    AnfAlgo::SetNodeAttr(kAttrInputNames, MakeValue<std::vector<std::string>>({"x"}), trans_node);
+
+    if (trans_node->kernel_info() == nullptr) {
+      trans_node->set_kernel_info(std::make_shared<device::KernelInfo>());
+    }
+    SelectKernelInfo(trans_node, KernelType::AKG_KERNEL);
+    mng->Replace(cnode, trans_node);
+  }
+}
+
+void UpdateFormatsAndDtypes(const CNodePtr &kernel_node, const std::vector<AnfNodePtr> &node_list,
+                            const std::vector<AnfNodePtr> &input_list, const FuncGraphManagerPtr &mng,
+                            const std::string &default_format, std::vector<std::string> *graph_input_format,
+                            std::vector<TypeId> *graph_input_type) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  MS_EXCEPTION_IF_NULL(mng);
+  MS_EXCEPTION_IF_NULL(graph_input_format);
+  MS_EXCEPTION_IF_NULL(graph_input_type);
+  // update graph input format and dtype use inner ops.
+  size_t input_num = AnfAlgo::GetInputTensorNum(kernel_node);
+  if (graph_input_format->size() != input_num) {
+    MS_LOG(EXCEPTION) << "Graph input format size is not equal to input num of cnode[" << kernel_node->DebugString()
+                      << "], [%" << graph_input_format->size() << "] != [%" << input_num << "]";
+  }
+  std::vector<bool> need_update(input_num, false);
+  auto &node_users = mng->node_users();
+  for (size_t i = 0; i < input_num; ++i) {
+    auto &input = input_list[i];
+    auto iter = node_users.find(input);
+    if (iter == node_users.end() || iter->second.empty()) {
+      continue;
+    }
+    for (auto &node_user : iter->second) {
+      if (node_user.first->kernel_info() == nullptr ||
+          node_user.first->kernel_info()->select_kernel_build_info() == nullptr) {
+        // maybe not a real kernel.
+        continue;
+      }
+      auto user_format = AnfAlgo::GetInputFormat(node_user.first, IntToSize(node_user.second - 1));
+      if (user_format != (*graph_input_format)[i]) {
+        MS_LOG(WARNING) << "Users of input: [" << i << "][" << input->DebugString(2) << " of ["
+                        << kernel_node->DebugString()
+                        << "] selected different format. we use defult: " << default_format;
+        (*graph_input_format)[i] = default_format;
+        need_update[i] = true;
+      }
+
+      if (kernel_node->input(i + 1)->isa<Parameter>()) {
+        auto user_dtype = AnfAlgo::GetInputDeviceDataType(node_user.first, IntToSize(node_user.second - 1));
+        if (user_dtype != (*graph_input_type)[i]) {
+          TypeId default_dtype = AnfAlgo::GetOutputInferDataType(input, 0);
+          MS_LOG(WARNING) << "Users of input: [" << i << "][" << input->DebugString(2) << " of ["
+                          << kernel_node->DebugString()
+                          << "] selected different dtype. we use default: " << TypeIdLabel(default_dtype);
+          (*graph_input_type)[i] = default_dtype;
+          need_update[i] = true;
+        }
+      }
+    }
+  }
+
+  for (size_t i = 0; i < input_num; ++i) {
+    if (!need_update[i]) {
+      continue;
+    }
+    need_update[i] = false;
+
+    MS_LOG(DEBUG) << "Update input format: " << i << " of: [" << kernel_node->DebugString()
+                  << "] to: " << (*graph_input_format)[i];
+    MS_LOG(DEBUG) << "Update input dtype: " << i << " of: [" << kernel_node->DebugString()
+                  << "] to: " << TypeIdLabel((*graph_input_type)[i]);
+    kernel::KernelBuildInfo::KernelBuildInfoBuilder builder;
+    std::vector<std::string> outputs_format = {(*graph_input_format)[i]};
+    std::vector<TypeId> outputs_device_type = {(*graph_input_type)[i]};
+    builder.SetOutputsFormat(outputs_format);
+    builder.SetOutputsDeviceType(outputs_device_type);
+    AnfAlgo::SetSelectKernelBuildInfo(builder.Build(), input_list[i].get());
+  }
+
+  ResetKernelBuildInfo(kernel_node);
+  // select nodes in subgraph again.
+  for (size_t i = 0; i < node_list.size(); ++i) {
+    auto anf_node = node_list[i];
+    MS_EXCEPTION_IF_NULL(anf_node);
+    auto cnode = anf_node->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    kernel::KernelBuildInfo::KernelBuildInfoBuilder builder;
+    size_t cnode_input_num = AnfAlgo::GetInputTensorNum(cnode);
+    for (size_t j = 0; j < cnode_input_num; ++j) {
+      auto input_node = cnode->input(j + 1);
+      MS_EXCEPTION_IF_NULL(input_node);
+      if (!IsValueNode<tensor::Tensor>(input_node)) {
+        continue;
+      }
+      // reset format and dtype of const tensor.
+      builder.SetOutputsFormat(std::vector<std::string>{kOpFormat_DEFAULT});
+      builder.SetOutputsDeviceType(std::vector<TypeId>{kTypeUnknown});
+      AnfAlgo::SetSelectKernelBuildInfo(builder.Build(), input_node.get());
+    }
+    SelectKernelInfo(node_list[i]->cast<CNodePtr>(), KernelType::AKG_KERNEL);
+  }
+}
+
+void SetGraphKernelInfo(const CNodePtr &kernel_node, const std::vector<std::pair<AnfNodePtr, size_t>> &output_index,
+                        const std::vector<std::string> &graph_input_format,
+                        const std::vector<TypeId> &graph_input_type) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  std::vector<std::string> graph_output_format;
+  std::vector<TypeId> graph_output_type;
+  for (size_t i = 0; i < output_index.size(); ++i) {
+    auto const &output = output_index[i];
+    graph_output_format.push_back(AnfAlgo::GetOutputFormat(output.first, output.second));
+    TypeId output_type(kTypeUnknown);
+    if (output.first->isa<CNode>()) {
+      output_type = AnfAlgo::GetCNodeOutputPrecision(output.first);
+    }
+    if (output_type == kTypeUnknown) {
+      output_type = AnfAlgo::GetOutputDeviceDataType(output.first, output.second);
+    }
+    graph_output_type.push_back(output_type);
+  }
+
+  kernel::KernelBuildInfo::KernelBuildInfoBuilder graph_info_builder;
+  graph_info_builder.SetInputsFormat(graph_input_format);
+  graph_info_builder.SetInputsDeviceType(graph_input_type);
+  graph_info_builder.SetOutputsFormat(graph_output_format);
+  graph_info_builder.SetOutputsDeviceType(graph_output_type);
+  graph_info_builder.SetProcessor(kernel::Processor::AICORE);
+  graph_info_builder.SetKernelType(KernelType::AKG_KERNEL);
+  graph_info_builder.SetFusionType(kernel::FusionType::OPAQUE);
+  auto graph_selected_info = graph_info_builder.Build();
+  MS_EXCEPTION_IF_NULL(graph_selected_info);
+  AnfAlgo::SetSelectKernelBuildInfo(graph_selected_info, kernel_node.get());
+  SetTensorDeviceInfo(*graph_selected_info, kernel_node);
+}
+
+void SelectGraphKernelInfo(const CNodePtr &kernel_node, const FuncGraphPtr &func_graph) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  MS_EXCEPTION_IF_NULL(func_graph);
+
+  // collect input info of funcgraph
+  std::vector<AnfNodePtr> node_list;
+  std::vector<AnfNodePtr> input_list;
+  std::vector<AnfNodePtr> output_list;
+  kernel::GetValidKernelNodes(func_graph, &node_list, &input_list, &output_list);
+  if (input_list.size() != kernel_node->inputs().size() - 1) {
+    MS_EXCEPTION(ArgumentError) << "Input num of funcgraph[" << func_graph->ToString() << "] not equal input of cnode["
+                                << kernel_node->DebugString() << "], [%" << input_list.size() << "] != ["
+                                << kernel_node->inputs().size() << "]";
+  }
+
+  std::string default_format;
+  bool use_same_format = true;
+  GetDefaultFormat(kernel_node, &default_format, &use_same_format);
+  MS_LOG(DEBUG) << "GraphKernel[" << func_graph->ToString() << "] use same input format[" << default_format
+                << "] for ParameterWeight.";
+
+  std::vector<std::string> graph_input_format;
+  std::vector<TypeId> graph_input_type;
+  UpdateGraphKernelInputsKernelInfo(kernel_node, input_list, default_format, use_same_format, &graph_input_format,
+                                    &graph_input_type);
+
+  auto mng = func_graph->manager();
+  if (mng == nullptr) {
+    mng = Manage(func_graph, true);
+  }
+  auto output_index = kernel::GetOutputIndex(node_list, input_list, output_list);
+  UpdateEquivFormat(output_index, node_list, func_graph, mng);
+  node_list.clear();
+  input_list.clear();
+  output_list.clear();
+  kernel::GetValidKernelNodes(func_graph, &node_list, &input_list, &output_list);
+
+  // update graph input format and dtype use inner ops.
+  UpdateFormatsAndDtypes(kernel_node, node_list, input_list, mng, default_format, &graph_input_format,
+                         &graph_input_type);
+
+  // set fix_precision for kernel when the me prim has fix_precision attr
+  UpdateKernelInfo(node_list);
+
+  output_index = kernel::GetOutputIndex(node_list, input_list, output_list);
+  SetGraphKernelInfo(kernel_node, output_index, graph_input_format, graph_input_type);
+}
+}  // namespace ascend
+}  // namespace device
+}  // namespace mindspore

mindspore/ccsrc/device/ascend/profiling/reporter/graph_desc_reporter.cc

@@ -24,7 +24,7 @@ namespace device {
 namespace ascend {
 void GraphDescReporter::ReportData() {
   for (const auto &node : cnode_list_) {
-    if (AnfAlgo::GetKernelType(node) != TBE_KERNEL && AnfAlgo::GetKernelType(node) != AUTO_DIFF_KERNEL) {
+    if (AnfAlgo::GetKernelType(node) != TBE_KERNEL && AnfAlgo::GetKernelType(node) != AKG_KERNEL) {
       MS_LOG(WARNING) << "Skip non tbe kernel";
       continue;
     }

mindspore/ccsrc/device/ascend/profiling/reporter/task_desc_reporter.cc

@@ -31,7 +31,7 @@ void TaskDescReporter::ReportData() {
 
   size_t task_index = 0;
   for (const auto &node : cnode_list_) {
-    if (AnfAlgo::GetKernelType(node) != TBE_KERNEL && AnfAlgo::GetKernelType(node) != AUTO_DIFF_KERNEL) {
+    if (AnfAlgo::GetKernelType(node) != TBE_KERNEL && AnfAlgo::GetKernelType(node) != AKG_KERNEL) {
       MS_LOG(WARNING) << "Skip non tbe kernel";
       ++task_index;
       continue;

mindspore/ccsrc/device/ascend/tasksink/task_generator.cc

@@ -43,7 +43,37 @@ bool TaskGenerator::GenTasks(const std::vector<CNodePtr> &anf_node_list, std::ve
 void TaskGenerator::LaunchAddrCleanKernel(const CNodePtr &anf_node_ptr, AddressPtrList *kernel_inputs) {
   MS_EXCEPTION_IF_NULL(anf_node_ptr);
   if (anf_node_ptr->inputs().size() != 2) {
-    MS_LOG(EXCEPTION) << "atomic Addr clean Node Input nodes not equal 2.";
+    // akg process
+    // set atomic clean addr
+    if (AnfAlgo::HasNodeAttr(kAttrAutomicOutputIndexs, anf_node_ptr)) {
+      auto clean_output_indexs = AnfAlgo::GetNodeAttr<std::vector<size_t>>(anf_node_ptr, kAttrAutomicOutputIndexs);
+      auto graph = anf_node_ptr->func_graph();
+      MS_EXCEPTION_IF_NULL(graph);
+      auto manager = graph->manager();
+      MS_EXCEPTION_IF_NULL(manager);
+      auto node_users = manager->node_users();
+      if (node_users[anf_node_ptr].empty()) {
+        MS_LOG(EXCEPTION) << "Node users of " << anf_node_ptr->ToString() << " is empty.";
+      }
+      auto depend_node = node_users[anf_node_ptr].pop().first;
+      if (!IsPrimitiveCNode(depend_node, prim::kPrimDepend)) {
+        MS_LOG(EXCEPTION) << "Checking Depend node failed";
+      }
+      if (node_users[depend_node].empty()) {
+        MS_LOG(EXCEPTION) << "Node users of " << depend_node->ToString() << " is empty.";
+      }
+      auto post_node = node_users[depend_node].pop().first;
+      for (auto index : clean_output_indexs) {
+        auto device_address = AnfAlgo::GetOutputAddr(post_node, index);
+        kernel::AddressPtr input = std::make_shared<kernel::Address>();
+        input->addr = device_address->ptr_;
+        MS_EXCEPTION_IF_NULL(input->addr);
+        input->size = device_address->size_;
+        kernel_inputs->push_back(input);
+      }
+      MS_LOG(DEBUG) << "AtomicAddClean clean output size: " << clean_output_indexs.size();
+    }
+    return;
   }
   MS_EXCEPTION_IF_NULL(anf_node_ptr->inputs()[1]);
   auto pre_node = (anf_node_ptr->inputs()[1])->cast<CNodePtr>();
@@ -59,7 +89,7 @@ void TaskGenerator::LaunchAddrCleanKernel(const CNodePtr &anf_node_ptr, AddressP
       input->size = device_address->size_;
       kernel_inputs->push_back(input);
     }
-    MS_LOG(INFO) << "AtomicAddClean clean output size:" << clean_output_indexs.size();
+    MS_LOG(DEBUG) << "AtomicAddClean clean output size:" << clean_output_indexs.size();
   }
   // set clean workspace address
   if (AnfAlgo::HasNodeAttr(kAttrAutomicWorkspaceSize, pre_node)) {

mindspore/ccsrc/device/gpu/gpu_kernel_build.cc

@@ -16,7 +16,7 @@
 #include "device/gpu/gpu_kernel_build.h"
 #include <string>
 #include "kernel/kernel.h"
-#include "kernel/akg/akgkernelbuild.h"
+#include "kernel/akg/akg_kernel_build.h"
 #include "kernel/akg/gpu/akg_gpu_kernel_build.h"
 #include "kernel/gpu/gpu_kernel_factory.h"
 #include "operator/ops.h"
@@ -37,7 +37,7 @@ void GpuBuild(const KernelGraphPtr &kernel_graph) {
       continue;
     }
 
-    if (session::AnfRuntimeAlgorithm::GetKernelType(kernel) == KernelType::AUTO_DIFF_KERNEL) {
+    if (session::AnfRuntimeAlgorithm::GetKernelType(kernel) == KernelType::AKG_KERNEL) {
       auto gpu_kernel_ptr = kernel::AkgGpuKernelBuild(kernel);
       if (!gpu_kernel_ptr) {
         MS_LOG(EXCEPTION) << "Build akg kernel op[" << kernel_name << "] failed";

mindspore/ccsrc/device/gpu/kernel_info_setter.cc

@@ -184,7 +184,7 @@ void SetKernelInfo(const CNodePtr &kernel_node) {
 
   if (!result) {
     result = SelectAkgKernel(kernel_node, builder->Build());
-    kernel_type = AUTO_DIFF_KERNEL;
+    kernel_type = AKG_KERNEL;
   }
 
   if (!result) {

mindspore/ccsrc/ir/anf.cc

@@ -26,6 +26,8 @@
 #include "ir/func_graph.h"
 #include "ir/primitive_base.h"
 
+#include "operator/ops.h"
+
 namespace mindspore {
 // namespace to support intermediate representation definition
 CNode::CNode(const std::vector<AnfNodePtr> &inputs, const FuncGraphPtr &func_graph)
@@ -106,10 +108,14 @@ std::string ValueNode::fullname_with_scope() {
 bool IsPrimitiveCNode(const AnfNodePtr &node, const PrimitivePtr &value) {
   MS_EXCEPTION_IF_NULL(node);
   auto cnode = node->cast<CNodePtr>();
-  if (cnode != nullptr) {
+  if (cnode == nullptr) {
+    return false;
+  }
+  if (value != nullptr) {
     return cnode->IsApply(value);
   }
-  return false;
+  const auto &prim = GetValueNode<PrimitivePtr>(cnode->input(0));
+  return prim != nullptr;
 }
 
 PrimitivePtr GetCNodePrimitive(const AnfNodePtr &node) {

mindspore/ccsrc/ir/anf.h

@@ -124,6 +124,7 @@ class AnfNode : public Base {
 
   const KernelInfoDevice *kernel_info() const { return kernel_info_.get(); }
   KernelInfoDevice *kernel_info() { return kernel_info_.get(); }
+  const KernelInfoDevicePtr &kernel_info_ptr() { return kernel_info_; }
   void set_kernel_info(const KernelInfoDevicePtr &kernel_info) { kernel_info_ = kernel_info; }
 
   AbstractBasePtr abstract() const { return abstract_; }
@@ -395,9 +396,9 @@ static S GetValue(const ValuePtr &value) {
 std::string GetCNodeFuncName(CNodePtr cnode);
 
 // used to check whether an AnfNode is a cnode with a kind of Primitive as first input
-bool IsPrimitiveCNode(const AnfNodePtr &node, const PrimitivePtr &value);
+bool IsPrimitiveCNode(const AnfNodePtr &node, const PrimitivePtr &value = nullptr);
 
-// used to check whether an AnfNode is a cnode with a Primitive as first input
+// used to get PrimitivePtr from a cnode first input
 PrimitivePtr GetCNodePrimitive(const AnfNodePtr &node);
 
 // used to check whether an AnfNode is a valuenode having some Primitive value

mindspore/ccsrc/ir/anf_extends.cc

@@ -70,7 +70,7 @@ std::string CNode::fullname_with_scope() {
     }
     fullname_with_scope_ = name;
   } else {
-    // cnode input 0 should be primitive ptr
+    // cnode input 0 should be primitive ptr or funcgraph ptr
     auto value_ptr = input(0)->cast<ValueNodePtr>();
     if (value_ptr == nullptr) {
       MS_LOG(WARNING) << "Input 0 of cnode is not a value node, its type is " << input(0)->type_name() << ".";
@@ -84,11 +84,23 @@ std::string CNode::fullname_with_scope() {
       return fullname_with_scope_;
     }
 
-    PrimitivePtr prim = GetValue<PrimitivePtr>(input_value);
+    auto prim = input_value->cast<PrimitivePtr>();
     MS_EXCEPTION_IF_NULL(scope());
-    MS_EXCEPTION_IF_NULL(prim);
-    fullname_with_scope_ =
-      scope()->name() + "/" + prim->name() + "-op" + id_generator::get_id(shared_from_base<CNode>());
+    fullname_with_scope_ = scope()->name() + "/";
+    if (prim != nullptr) {
+      fullname_with_scope_ += prim->name();
+    } else {
+      auto func_graph = input_value->cast<FuncGraphPtr>();
+      MS_EXCEPTION_IF_NULL(func_graph);
+      auto fg_flag = func_graph->get_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL);
+      if (fg_flag != nullptr) {
+        auto fg_name = GetValue<std::string>(fg_flag);
+        fullname_with_scope_ += "GraphKernel_" + fg_name;
+      } else {
+        fullname_with_scope_ += func_graph->ToString();
+      }
+    }
+    fullname_with_scope_ += "-op" + id_generator::get_id(shared_from_base<CNode>());
   }
 
   return fullname_with_scope_;

mindspore/ccsrc/ir/dtype/number.h

@@ -77,9 +77,9 @@ class Bool : public Number {
 
   TypeId generic_type_id() const override { return kNumberTypeBool; }
   TypePtr DeepCopy() const override { return std::make_shared<Bool>(); }
-  std::string ToString() const override { return "Bool_"; }
-  std::string ToReprString() const override { return "bool_"; }
-  std::string DumpText() const override { return "Bool_"; }
+  std::string ToString() const override { return "Bool"; }
+  std::string ToReprString() const override { return "bool"; }
+  std::string DumpText() const override { return "Bool"; }
 };
 
 // Int

mindspore/ccsrc/ir/func_graph.cc

@@ -34,7 +34,7 @@ namespace mindspore {
  * Methods of Graph
  */
 FuncGraph::FuncGraph()
-    : flags_(),
+    : attrs_(),
       transforms_(),
       parameter_default_value_(),
       seen_(0),
@@ -95,13 +95,27 @@ ParameterPtr FuncGraph::AddWeightParameter(const std::string &name) {
   return p;
 }
 
-bool FuncGraph::has_flag(const std::string &flag) {
-  if (flags_.count(flag)) {
-    return flags_[flag];
+bool FuncGraph::has_flag(const std::string &key) {
+  auto iter = attrs_.find(key);
+  if (iter != attrs_.cend()) {
+    if (iter->second->isa<BoolImm>()) {
+      return GetValue<bool>(iter->second);
+    }
+    MS_LOG(WARNING) << "key " << key << " is not a flag, please use has_attr function.";
   }
   return false;
 }
 
+bool FuncGraph::has_attr(const std::string &key) {
+  auto iter = attrs_.find(key);
+  return !(iter == attrs_.cend());
+}
+
+ValuePtr FuncGraph::get_attr(const std::string &key) {
+  auto iter = attrs_.find(key);
+  return iter == attrs_.cend() ? nullptr : iter->second;
+}
+
 CNodePtr FuncGraph::NewCNode(const std::vector<AnfNodePtr> &inputs) {
   CNodePtr cnode = std::make_shared<CNode>(inputs, shared_from_base<FuncGraph>());
   if (has_flag(GRAPH_FLAG_HAS_EFFECT)) {

mindspore/ccsrc/ir/func_graph.h

@@ -74,6 +74,7 @@ using FuncGraphMap = OrderedMap<FuncGraphPtr, int>;
 const char FUNC_GRAPH_FLAG_IGNORE_VALUES[] = "ignore_values";
 const char FUNC_GRAPH_FLAG_DEFER_INLINE[] = "defer_inline";
 const char FUNC_GRAPH_FLAG_CORE[] = "core";
+const char FUNC_GRAPH_ATTR_GRAPH_KERNEL[] = "graph_kernel";
 const char FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER[] = "spec_param";
 
 namespace abstract {
@@ -195,10 +196,19 @@ class FuncGraph : public FuncGraphBase {
   void set_is_generate(bool generated) { is_generated_ = generated; }
   bool is_generated() const { return is_generated_; }
 
-  bool has_flag(const std::string &flag);
-  std::unordered_map<std::string, bool> &flags() { return flags_; }
-  void set_flags(const std::unordered_map<std::string, bool> &flags) { flags_ = flags; }
-  void set_flags(const std::string &key, const bool value) { flags_[key] = value; }
+  std::unordered_map<std::string, ValuePtr> &attrs() { return attrs_; }
+  void set_attrs(const std::unordered_map<std::string, ValuePtr> &attrs) {
+    for (auto &attr : attrs) {
+      attrs_[attr.first] = attr.second;
+    }
+  }
+  bool has_flag(const std::string &key);
+  void set_flag(const std::string &key, bool flag) { attrs_[key] = MakeValue(flag); }
+  void erase_flag(const std::string &key) { (void)attrs_.erase(key); }
+
+  bool has_attr(const std::string &key);
+  ValuePtr get_attr(const std::string &key);
+  void set_attr(const std::string &key, const ValuePtr &value) { attrs_[key] = value; }
 
   std::unordered_map<std::string, FuncGraphTransform> &transforms() { return transforms_; }
   void set_transforms(const std::unordered_map<std::string, FuncGraphTransform> &transforms) {
@@ -317,7 +327,7 @@ class FuncGraph : public FuncGraphBase {
 
   std::unordered_map<AnfNodePtr, AnfNodePtr> &make_ref_params() { return make_ref_params_; }
 
-  std::unordered_map<std::string, bool> flags_;
+  std::unordered_map<std::string, ValuePtr> attrs_;
   std::unordered_map<std::string, FuncGraphTransform> transforms_;
   // parameter default value
   std::map<std::string, AnfNodePtr> parameter_default_value_;

mindspore/ccsrc/ir/func_graph_cloner.cc

@@ -90,6 +90,7 @@ void Cloner::CloneCNode(const AnfNodePtr &node, const FuncGraphPtr &target) {
   new_node->set_abstract(old_node->abstract());
   ScopePtr scope = (node->scope() != kDefaultScope) ? node->scope() : this->scope();
   new_node->set_scope(scope);
+  new_node->set_kernel_info(old_node->kernel_info_ptr());
   repl_node_[old_node] = new_node;
   nodes_.emplace_back(old_node, new_node);
   TraceManager::EndTrace();
@@ -211,7 +212,7 @@ void Cloner::SetFuncGraphInfo(const FuncGraphPtr &func_graph, FuncGraphPtr *cons
   MS_EXCEPTION_IF_NULL(target_func_graph);
   TraceManager::DebugTrace(func_graph->debug_info(), target_relation_);
   *target_func_graph = std::make_shared<FuncGraph>();
-  (*target_func_graph)->set_flags(func_graph->flags());
+  (*target_func_graph)->set_attrs(func_graph->attrs());
   (*target_func_graph)->set_transforms(func_graph->transforms());
   (*target_func_graph)->set_has_vararg(func_graph->has_vararg());
   (*target_func_graph)->set_has_kwarg(func_graph->has_kwarg());
@@ -636,9 +637,14 @@ FuncGraphPtr TransformableClone(const FuncGraphPtr &func_graph, const TraceInfoP
 
   if (MsContext::GetInstance()->is_multi_graph_sink()) {
     if (func_graph->has_flag(FUNC_GRAPH_FLAG_IGNORE_VALUES)) {
-      new_func_graph->set_flags(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
+      new_func_graph->set_flag(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
     }
   }
+
+  if (func_graph->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+    new_func_graph->set_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL, func_graph->get_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL));
+  }
+
   return new_func_graph;
 }
 }  // namespace mindspore

mindspore/ccsrc/ir/func_graph_extends.cc

@@ -399,8 +399,8 @@ void FuncGraph::ReleaseFullOrderToEffectOrder() {
         depend_inputs.push_back(*iter);
       }
     }
-    set_flags(GRAPH_FLAG_HAS_EFFECT, false);
-    set_flags(GRAPH_FLAG_EFFECT_PATIAL_ORDER, true);
+    set_flag(GRAPH_FLAG_HAS_EFFECT, false);
+    set_flag(GRAPH_FLAG_EFFECT_PATIAL_ORDER, true);
     if (!depend_inputs.empty()) {
       SetEffectDepends(depend_inputs);
     }

mindspore/ccsrc/kernel/CMakeLists.txt

@@ -9,6 +9,10 @@ if (ENABLE_D)
 	file(GLOB_RECURSE D_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR}
 		"kernel_query.cc"
 		"kernel_fusion.cc"
+		"akg/ascend/*.cc"
+		"akg/akg_kernel_build.cc"
+		"akg/akg_kernel_attrs_process.cc"
+		"akg/akg_kernel_metadata.cc"
 		"tbe/*.cc"
 		"aicpu/*.cc"
 		"rts/*.cc"
@@ -33,7 +37,7 @@ if (ENABLE_GPU)
     file(GLOB_RECURSE CUDA_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR}
         "gpu/*.cu"
         "akg/gpu/*.cc"
-        "akg/akgkernelbuild.cc"
+        "akg/akg_kernel_build.cc"
         "akg/akg_kernel_attrs_process.cc"
 	)
 

mindspore/ccsrc/kernel/aicpu/aicpu_kernel_build.cc

@@ -24,7 +24,7 @@
 #include <map>
 #include "device/kernel_runtime.h"
 #include "kernel/aicpu/aicpu_kernel_mod.h"
-#include "kernel/akg/akgkernelbuild.h"
+#include "kernel/akg/akg_kernel_build.h"
 #include "proto/tensor.pb.h"
 #include "proto/tensor_shape.pb.h"
 #include "proto/attr.pb.h"

mindspore/ccsrc/kernel/akg/akg_kernel_attrs_process.cc

@@ -79,6 +79,10 @@ void SetAkgAttrsForCast(const AnfNodePtr &anf_node) {
     dst_type = "float32";
   } else if (output_type == kFloat16->type_id()) {
     dst_type = "float16";
+  } else if (output_type == kInt32->type_id()) {
+    dst_type = "int32";
+  } else {
+    MS_LOG(WARNING) << "Unknown cast_to type: " << TypeIdToType(output_type)->ToString();
   }
   AnfAlgo::SetNodeAttr("dst_type", MakeValue(dst_type), anf_node);
 }

mindspore/ccsrc/kernel/akg/akg_kernel_build.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "kernel/akg/akgkernelbuild.h"
+#include "kernel/akg/akg_kernel_build.h"
 #include <Python.h>
 #include <sys/types.h>
 #include <signal.h>
@@ -43,7 +43,9 @@ namespace kernel {
 constexpr int ME_MAX_KERNEL_NAME_LENGTH = 200;
 constexpr int32_t ARGS_SIZE = 1;
 constexpr auto kCompileWithJsonFunc = "compilewithjson";
+
 // json key
+constexpr auto kOpDesc = "op_desc";
 constexpr auto kInputDesc = "input_desc";
 constexpr auto kShape = "shape";
 constexpr auto kDataType = "data_type";
@@ -51,13 +53,24 @@ constexpr auto kOutputDesc = "output_desc";
 constexpr auto kName = "name";
 constexpr auto kTensorName = "tensor_name";
 constexpr auto kValue = "value";
-constexpr auto KInpputNames = "input_names";
+constexpr auto KDynInputSizes = "dyn_input_sizes";
+constexpr auto KInputNames = "input_names";
 constexpr auto KInput = "input";
 constexpr auto KDtype = "dtype";
-int AkgKernelBuild::op_cnt_ = 0;
-std::mutex AkgKernelBuild::op_cnt_mtx_;
+namespace {
+template <typename T>
+std::string Vector2Str(const std::vector<T> &inputs) {
+  if (!inputs.empty()) {
+    std::ostringstream oss;
+    (void)std::copy(inputs.begin(), inputs.end() - 1, std::ostream_iterator<T>(oss, ", "));
+    oss << inputs.back();
+    return oss.str();
+  }
+  return "";
+}
+}  // namespace
 
-std::string PyObjectToStr(PyObject *const PyObj) {
+std::string AkgKernelBuild::PyObjectToStr(PyObject *const PyObj) {
   char *pChar = nullptr;
   std::string str_res;
   if (PyObj == nullptr) {
@@ -76,6 +89,72 @@ std::string PyObjectToStr(PyObject *const PyObj) {
   return str_res;
 }
 
+std::string GetTensorName(const nlohmann::json &node_json, const std::string &tag,
+                          const std::pair<size_t, size_t> &position) {
+  if (node_json.count(tag) == 0) {
+    MS_LOG(ERROR) << "Node [" << node_json.dump() << "] has no key [" << tag << "].";
+    return "";
+  }
+
+  auto const &tag_desc = node_json[tag];
+  nlohmann::json first_index;
+  if (tag == kOutputDesc) {
+    first_index = tag_desc;
+  } else if (!tag_desc.is_array() || tag_desc.size() <= position.first) {
+    MS_LOG(ERROR) << "Node [" << tag_desc.dump() << "] has no enough value [" << position.first << "].";
+    return "";
+  } else {
+    first_index = tag_desc[position.first];
+  }
+
+  if (!first_index.is_array() || first_index.size() <= position.second) {
+    MS_LOG(ERROR) << "Node [" << first_index.dump() << "] has no enough value [" << position.second << "].";
+    return "";
+  }
+  auto const &second_index = first_index[position.second];
+  if (second_index.count(kTensorName) == 0) {
+    MS_LOG(ERROR) << "Node [" << second_index.dump() << "] has no key [" << kTensorName << "].";
+    return "";
+  }
+
+  return second_index[kTensorName];
+}
+
+void SetTensorName(const std::string &tag, const std::string &new_name, const std::pair<size_t, size_t> &position,
+                   nlohmann::json *const node_json) {
+  MS_EXCEPTION_IF_NULL(node_json);
+  if (node_json->count(tag) == 0) {
+    MS_LOG(ERROR) << "Node [" << node_json->dump() << "] has no key [" << tag << "].";
+    return;
+  }
+
+  nlohmann::json *tag_desc = &((*node_json)[tag]);
+  nlohmann::json *first_index;
+  if (tag == kOutputDesc) {
+    first_index = tag_desc;
+  } else if (!tag_desc->is_array() || tag_desc->size() <= position.first) {
+    MS_LOG(ERROR) << "Node [" << tag_desc->dump() << "] has no enough value [" << position.first << "].";
+    return;
+  } else {
+    first_index = &((*tag_desc)[position.first]);
+  }
+
+  if (!first_index->is_array() || first_index->size() <= position.second) {
+    MS_LOG(ERROR) << "Node [" << first_index->dump() << "] has no enough value [" << position.second << "].";
+    return;
+  }
+  nlohmann::json *second_index = &((*first_index)[position.second]);
+  if (second_index->count(kTensorName) == 0) {
+    MS_LOG(ERROR) << "Node [" << second_index->dump() << "] has no key [" << kTensorName << "].";
+    return;
+  }
+  (*second_index)[kTensorName] = new_name;
+  return;
+}
+
+int AkgKernelBuild::op_cnt_ = 0;
+std::mutex AkgKernelBuild::op_cnt_mtx_;
+
 std::string AkgKernelBuild::GetProcessor(const AnfNodePtr &anf_node) {
   MS_EXCEPTION_IF_NULL(anf_node);
   std::string device;
@@ -187,10 +266,7 @@ bool AkgKernelBuild::CreateInputDescJson(const AnfNodePtr &anf_node, nlohmann::j
     for (size_t input_i = 0; input_i < input_tensor_num; input_i++) {
       // dtype : float16
       auto type_id = AnfAlgo::GetInputDeviceDataType(anf_node, real_input_index);
-      TypePtr type_ptr = TypeIdToType(type_id);
-      MS_EXCEPTION_IF_NULL(type_ptr);
-      std::string dtype = type_ptr->ToString();
-      dtype = Dtype2String(dtype);
+      std::string dtype = TypeId2String(type_id);
       if (dtype.empty()) {
         MS_LOG(ERROR) << "Op [" << op_name << "] input [" << input_i << "] data type is null. ";
         return false;
@@ -198,13 +274,23 @@ bool AkgKernelBuild::CreateInputDescJson(const AnfNodePtr &anf_node, nlohmann::j
       nlohmann::json input_desc_json;
       input_desc_json[kDataType] = dtype;
       input_desc_json[kName] = op_input_name;
-      input_desc_json[kTensorName] =
-        op_input_name + "_" + std::to_string(real_input_index) + "_" + std::to_string(input_i);
-      input_desc_json[kShape] = AnfAlgo::GetInputDeviceShape(anf_node, real_input_index);
+      input_desc_json[kTensorName] = "input_" + std::to_string(GetInputTensorIdxInc(anf_node, real_input_index));
+      auto input_shape = AnfAlgo::GetInputDeviceShape(anf_node, real_input_index);
+      if (GetInputTensorValue(anf_node, real_input_index, &input_desc_json)) {
+        MS_LOG(WARNING) << "we take input[" << real_input_index << "] of [" << anf_node->DebugString(2)
+                        << "] as const tensor, shape: [" << Vector2Str(input_shape)
+                        << "], value: " << input_desc_json[kValue];
+
+        input_shape.clear();
+      }
+      if (input_shape.empty()) {
+        input_shape.push_back(1);
+      }
+      input_desc_json[kShape] = input_shape;
       input_list.emplace_back(input_desc_json);
+      real_input_index++;
     }
     inputs_json->emplace_back(input_list);
-    real_input_index++;
   }
   return true;
 }
@@ -220,10 +306,7 @@ bool AkgKernelBuild::CreateOutputDescJson(const AnfNodePtr &anf_node, nlohmann::
   for (size_t i = 0; i < output_tensor_num; i++) {
     nlohmann::json output_json;
     auto type_id = AnfAlgo::GetOutputDeviceDataType(anf_node, i);
-    TypePtr type_ptr = TypeIdToType(type_id);
-    MS_EXCEPTION_IF_NULL(type_ptr);
-    std::string dtype = type_ptr->ToString();
-    dtype = Dtype2String(dtype);
+    std::string dtype = TypeId2String(type_id);
     if (dtype.empty()) {
       MS_LOG(ERROR) << "Op [" << op_name << "] output [" << i << "] data type is null. ";
       return false;
@@ -232,7 +315,7 @@ bool AkgKernelBuild::CreateOutputDescJson(const AnfNodePtr &anf_node, nlohmann::
     std::string output_name = outputs[i]->name();
     output_json[kDataType] = dtype;
     output_json[kName] = output_name;
-    output_json[kTensorName] = output_name + "_" + std::to_string(i);
+    output_json[kTensorName] = "output_" + std::to_string(i) + "_" + std::to_string(GetOutputTensorIdxInc());
     output_json[kShape] = AnfAlgo::GetOutputDeviceShape(anf_node, i);
     outputs_json->push_back(output_json);
   }
@@ -358,15 +441,14 @@ bool AkgKernelBuild::GenerateSingleKernelJson(const AnfNodePtr &anf_node, const
   MS_EXCEPTION_IF_NULL(op_info_ptr);
 
   // get basic params from currentNodeOpDesc
-  (*node_json)["platform"] = "AKG";
   (*node_json)[kName] = op_name;
-  (*node_json)["fusion_type"] = AnfAlgo::GetFusionType(anf_node);
   (*node_json)["impl_path"] = op_info_ptr->impl_path();
   (*node_json)["process"] = AkgKernelBuild::GetProcessor(anf_node);
+  (*node_json)["composite"] = false;
 
   auto primitive = AnfAlgo::GetCNodePrimitive(anf_node);
   MS_EXCEPTION_IF_NULL(primitive);
-  ValuePtr input_names_v = primitive->GetAttr(KInpputNames);
+  ValuePtr input_names_v = primitive->GetAttr(KInputNames);
   if (input_names_v == nullptr) {
     MS_LOG(ERROR) << "ApplyKernel has no input_names, op[" << op_name << "].";
     return false;
@@ -465,12 +547,12 @@ KernelPackPtr AkgKernelBuild::OpBuild(const std::string &node_json, const AnfNod
   (void)alarm(0);
   if (pRes == nullptr) {
     MS_LOG(ERROR) << "No ret got, failed to call function [" << kCompileWithJsonFunc << "], args:\n("
-                  << PyObjectToStr(pArg) << ").";
+                  << AkgKernelBuild::PyObjectToStr(pArg) << ").";
     return nullptr;
   }
   if (PyObject_IsTrue(pRes) != 1) {
     MS_LOG(ERROR) << "Illegal ret, failed to call function [" << kCompileWithJsonFunc << "], args:\n("
-                  << PyObjectToStr(pArg) << ").";
+                  << AkgKernelBuild::PyObjectToStr(pArg) << ").";
     return nullptr;
   }
 
@@ -513,5 +595,29 @@ KernelPackPtr AkgKernelBuild::BuildByJson(const AnfNodePtr &anf_node, std::vecto
                << "]";
   return kernel_pack;
 }
+
+size_t AkgKernelBuild::GetInputTensorIdxInc(const AnfNodePtr &anf_node, size_t input_idx) {
+  MS_EXCEPTION_IF_NULL(anf_node);
+  auto cnode = anf_node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(cnode);
+  if (input_idx + 1 >= cnode->inputs().size()) {
+    MS_EXCEPTION(ArgumentError) << "input_idx [" << input_idx << "] is out of index of inputs of ["
+                                << cnode->inputs().size() - 1 << "][" << cnode->DebugString() << "]";
+  }
+
+  auto input_node = cnode->input(input_idx + 1);
+  if (input_tensor_idx_.find(input_node) == input_tensor_idx_.end()) {
+    size_t index = input_tensor_idx_.size();
+    input_tensor_idx_[input_node] = index;
+  }
+
+  return input_tensor_idx_[input_node];
+}
+
+size_t AkgKernelBuild::GetOutputTensorIdxInc() {
+  size_t idx = output_tensor_idx_++;
+  return idx;
+}
+
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/akg/akg_kernel_build.h

@@ -32,29 +32,45 @@ namespace mindspore {
 namespace kernel {
 class AkgKernelBuild {
  public:
-  AkgKernelBuild() = default;
+  AkgKernelBuild() {
+    input_tensor_idx_ = {};
+    output_tensor_idx_ = 0;
+  }
   ~AkgKernelBuild() = default;
 
   KernelPackPtr BuildByJson(const AnfNodePtr &anf_node, std::vector<size_t> *const input_size,
                             std::vector<size_t> *const output_size);
+  static std::string GetProcessor(const AnfNodePtr &anf_node);
+  static std::string PyObjectToStr(PyObject *const PyObj);
 
- private:
+ protected:
   bool CreateInputDescJson(const AnfNodePtr &anf_node, nlohmann::json *const inputs_json);
   bool CreateOutputDescJson(const AnfNodePtr &anf_node, nlohmann::json *const outputs_json);
   bool CreateAttrDescJson(const AnfNodePtr &anf_node, const std::string &op_name,
                           const std::shared_ptr<OpInfo> &op_info, nlohmann::json *const attrs_json);
+  KernelPackPtr OpBuild(const std::string &node_json, const AnfNodePtr &anf_node);
+  int GetOpCntInc();
+  size_t GetInputTensorIdxInc(const AnfNodePtr &anf_node, size_t input_idx);
+  size_t GetOutputTensorIdxInc();
   bool GenerateSingleKernelJson(const AnfNodePtr &anf_node, const std::string &op_name,
                                 nlohmann::json *const node_json);
-  KernelPackPtr OpBuild(const std::string &node_json, const AnfNodePtr &anf_node);
 
-  int GetOpCntInc();
-  std::string GetProcessor(const AnfNodePtr &anf_node);
   static int op_cnt_;
   // lock for variable fusionOpCnt in singleton mode
   static std::mutex op_cnt_mtx_;
   std::string json_name_;
   std::string json_info_;
+  std::unordered_map<AnfNodePtr, size_t> input_tensor_idx_;
+  size_t output_tensor_idx_;
 };
+
+bool GetIOSize(const nlohmann::json &node_json, std::vector<size_t> *const input_size,
+               std::vector<size_t> *const output_size);
+void SetTensorName(const std::string &tag, const std::string &new_name, const std::pair<size_t, size_t> &position,
+                   nlohmann::json *const node_json);
+std::string GetTensorName(const nlohmann::json &node_json, const std::string &tag,
+                          const std::pair<size_t, size_t> &position);
+
 }  // namespace kernel
 }  // namespace mindspore
 

mindspore/ccsrc/kernel/akg/akg_kernel_metadata.cc

@@ -0,0 +1,50 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernel/akg/akg_kernel_metadata.h"
+#include <memory>
+#include "session/anf_runtime_algorithm.h"
+#include "kernel/oplib/oplib.h"
+#include "kernel/common_utils.h"
+
+namespace mindspore {
+namespace kernel {
+void AkgMetadataInfo(const CNodePtr &kernel_node,
+                     std::vector<std::shared_ptr<KernelBuildInfo>> *const kernel_info_list) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  MS_EXCEPTION_IF_NULL(kernel_info_list);
+
+  std::string op_name = AnfAlgo::GetCNodeName(kernel_node);
+  for (size_t i = 0; i < support_devices.size(); i++) {
+    auto op_info_ptr = mindspore::kernel::OpLib::FindOp(op_name, OpImplyType::kAKG);
+    if (op_info_ptr == nullptr) {
+      continue;
+    }
+
+    if (!ParseMetadata(kernel_node, op_info_ptr, Processor(i), kernel_info_list)) {
+      MS_LOG(WARNING) << "Akg parsed metadata of op[" << op_name << "], device[" << support_devices[i] << "] failed.";
+    } else {
+      MS_LOG(DEBUG) << "Akg parsed metadata of op[" << op_name << "], device[" << support_devices[i] << "].";
+      break;
+    }
+  }
+
+  if (kernel_info_list->empty()) {
+    MS_LOG(WARNING) << "Akg dose not has metadata of op[" << op_name << "].";
+  }
+}
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/kernel/akg/akg_kernel_metadata.h

@@ -0,0 +1,31 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_KERNEL_AKG_AKG_KERNEL_METADATA_H_
+#define MINDSPORE_CCSRC_KERNEL_AKG_AKG_KERNEL_METADATA_H_
+
+#include <string>
+#include <vector>
+#include <unordered_map>
+#include <memory>
+#include "kernel/kernel_build_info.h"
+
+namespace mindspore {
+namespace kernel {
+void AkgMetadataInfo(const CNodePtr &kernel_node, std::vector<std::shared_ptr<KernelBuildInfo>> *kernel_info_list);
+}  // namespace kernel
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_KERNEL_AKG_AKG_KERNEL_METADATA_H_

mindspore/ccsrc/kernel/akg/ascend/akg_ascend_kernel_build.cc

@@ -0,0 +1,385 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernel/akg/ascend/akg_ascend_kernel_build.h"
+
+#include <algorithm>
+#include <map>
+#include <memory>
+#include <string>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+#include <Python.h>
+#include "ir/dtype.h"
+#include "ir/func_graph.h"
+#include "kernel/kernel.h"
+#include "kernel/common_utils.h"
+#include "kernel/tbe/tbe_utils.h"
+#include "kernel/akg/ascend/akg_ascend_kernel_mod.h"
+#include "kernel/akg/akg_kernel_attrs_process.h"
+#include "session/anf_runtime_algorithm.h"
+
+namespace mindspore {
+namespace kernel {
+
+constexpr int32_t PARALLEL_ARGS_SIZE = 3;
+constexpr int32_t PROCESS_NUM = 16;
+constexpr int32_t TIME_OUT = 300;
+
+constexpr auto kOpDesc = "op_desc";
+constexpr auto kShape = "shape";
+constexpr auto kDataType = "data_type";
+constexpr auto kInputDesc = "input_desc";
+constexpr auto kOutputDesc = "output_desc";
+constexpr auto kTensorName = "tensor_name";
+constexpr auto kCompileAkgKernelParallelFunc = "compile_akg_kernel_parallel";
+constexpr auto kMultiProcModule = "mindspore._extends.parallel_compile.akg_compiler.multi_process_compiler";
+
+bool AkgAscendKernelBuilder::CollectJson(const AnfNodePtr &anf_node) {
+  MS_EXCEPTION_IF_NULL(anf_node);
+  std::string op_name = AnfAlgo::GetCNodeName(anf_node);
+  MS_LOG(INFO) << "AKG start compile, op[" << op_name << "], device[" << AkgKernelBuild::GetProcessor(anf_node) << "]";
+  auto it = kAkgKernelAttrsProcessMap.find(op_name);
+  if (it != kAkgKernelAttrsProcessMap.end()) {
+    it->second(anf_node);
+  }
+  MS_LOG(INFO) << "Akg start compile, op[" << op_name << "], device[" << AkgKernelBuild::GetProcessor(anf_node) << "]";
+  nlohmann::json node_json;
+  if (!GenerateSingleKernelJson(anf_node, op_name, &node_json)) {
+    MS_LOG(ERROR) << "Op[" << op_name << "] create single kernel json failed.";
+  }
+
+  kernel_json_ = node_json.dump();
+
+  if (!GetIOSize(node_json, &input_size_list_, &output_size_list_)) {
+    MS_LOG(ERROR) << "Cal mem size failed.";
+    return false;
+  }
+
+  return true;
+}
+
+bool AkgAscendKernelBuilder::CollectFusedJson(const std::vector<AnfNodePtr> &anf_nodes,
+                                              const std::vector<AnfNodePtr> &input_list,
+                                              const std::vector<AnfNodePtr> &output_list) {
+  if (anf_nodes.empty() || input_list.empty()) {
+    MS_LOG(ERROR) << "Invalid input size, anf_nodes [" << anf_nodes.size() << "], input_list [" << input_list.size()
+                  << "].";
+    return false;
+  }
+  MS_LOG(INFO) << "anf_nodes [" << output_list.size() << "], input_list [" << anf_nodes.size() << "], output_list ["
+               << input_list.size() << "].";
+
+  std::map<AnfNodePtr, nlohmann::json> node_json_map;
+
+  for (auto const &anf_node : anf_nodes) {
+    MS_EXCEPTION_IF_NULL(anf_node);
+    std::string op_name = AnfAlgo::GetCNodeName(anf_node);
+    if (!AnfAlgo::IsRealKernel(anf_node)) {
+      MS_LOG(ERROR) << "Invalid anf node to build [" << anf_node->fullname_with_scope() << "].";
+      return false;
+    }
+    auto it = kAkgKernelAttrsProcessMap.find(op_name);
+    if (it != kAkgKernelAttrsProcessMap.end()) {
+      it->second(anf_node);
+    }
+
+    nlohmann::json node_json;
+    if (!GenerateSingleKernelJson(anf_node, op_name, &node_json)) {
+      MS_LOG(ERROR) << "Op [" << op_name << "] create single kernel json failed.";
+      return false;
+    }
+    // No need for composite op.
+    node_json.erase("id");
+    node_json.erase("op");
+    node_json.erase("composite");
+
+    auto primitive = AnfAlgo::GetCNodePrimitive(anf_node);
+    MS_EXCEPTION_IF_NULL(primitive);
+
+    if (primitive->GetAttr("fusion") != nullptr) {
+      node_json["fusion"] = primitive->GetAttr("fusion")->ToString();
+    }
+
+    node_json_map[anf_node] = node_json;
+  }
+
+  for (auto const &anf_node : anf_nodes) {
+    std::vector<int> dyn_input_sizes;
+    auto primitive = AnfAlgo::GetCNodePrimitive(anf_node);
+    MS_EXCEPTION_IF_NULL(primitive);
+
+    if (primitive->GetAttr(kAttrDynInputSizes) != nullptr) {
+      dyn_input_sizes = GetValue<const std::vector<int>>(primitive->GetAttr(kAttrDynInputSizes));
+    }
+
+    bool is_dynamic_input = !dyn_input_sizes.empty();
+    size_t input_num = is_dynamic_input ? dyn_input_sizes.size() : AnfAlgo::GetInputTensorNum(anf_node);
+    size_t real_input_index = 0;
+    for (size_t i = 0; i < input_num; ++i) {
+      size_t input_tensor_num = is_dynamic_input ? IntToSize(dyn_input_sizes[i]) : 1;
+      for (size_t j = 0; j < input_tensor_num; ++j) {
+        auto tmp_input = GetKernelInput(anf_node, real_input_index);
+        std::string tensor_name = GetTensorName(node_json_map[anf_node], kInputDesc, std::make_pair(i, j));
+        if (node_json_map.find(tmp_input.first) != node_json_map.end()) {
+          std::string new_tensor_name =
+            GetTensorName(node_json_map[tmp_input.first], kOutputDesc, std::make_pair(0, tmp_input.second));
+          SetTensorName(kInputDesc, new_tensor_name, std::make_pair(i, j), &(node_json_map[anf_node]));
+          MS_LOG(DEBUG) << "Update [" << real_input_index << "] input [" << tensor_name << "] of ["
+                        << anf_node->fullname_with_scope() << "] to [" << tmp_input.second << "] output ["
+                        << new_tensor_name << "] of [" << tmp_input.first->fullname_with_scope() << "].";
+        } else {
+          MS_LOG(DEBUG) << "[" << real_input_index << "] input " << tensor_name << "] of ["
+                        << anf_node->fullname_with_scope() << "] is out input.";
+        }
+        real_input_index++;
+      }
+    }
+  }
+
+  nlohmann::json fused_node_json;
+  std::vector<nlohmann::json> node_json_desc;
+  std::transform(anf_nodes.begin(), anf_nodes.end(), std::back_inserter(node_json_desc),
+                 [&node_json_map](const AnfNodePtr &anf_node) { return node_json_map[anf_node]; });
+  fused_node_json[kOpDesc] = node_json_desc;
+
+  nlohmann::json inputs_json;
+  auto input_index = GetInputIndex(anf_nodes, input_list);
+  for (size_t i = 0; i < input_index.size(); ++i) {
+    auto tmp_input = input_index[i];
+    auto type_id = AnfAlgo::GetInputDeviceDataType(tmp_input.first, tmp_input.second.first);
+    std::string dtype = TypeId2String(type_id);
+    nlohmann::json input_desc_json;
+    input_desc_json[kTensorName] = GetTensorName(node_json_map[tmp_input.first], kInputDesc, tmp_input.second);
+    input_desc_json[kDataType] = dtype;
+    input_desc_json[kShape] = AnfAlgo::GetInputDeviceShape(tmp_input.first, tmp_input.second.first);
+    inputs_json.emplace_back(std::vector<nlohmann::json>{input_desc_json});
+  }
+  fused_node_json[kInputDesc] = inputs_json;
+
+  nlohmann::json outputs_json;
+  auto output_index = GetOutputIndex(anf_nodes, input_list, output_list);
+  for (size_t i = 0; i < output_index.size(); ++i) {
+    auto tmp_output = output_index[i];
+    bool found = false;
+    nlohmann::json output_desc_json;
+    for (size_t input_i = 0; input_i < input_list.size(); ++input_i) {
+      if (tmp_output.first == input_list[input_i]) {
+        output_desc_json = inputs_json[input_i][0];
+        found = true;
+        break;
+      }
+    }
+    if (!found) {
+      auto type_id = AnfAlgo::GetOutputDeviceDataType(tmp_output.first, tmp_output.second);
+      std::string dtype = TypeId2String(type_id);
+      output_desc_json[kTensorName] =
+        GetTensorName(node_json_map[tmp_output.first], kOutputDesc, std::make_pair(0, tmp_output.second));
+      output_desc_json[kDataType] = dtype;
+      auto output_shape = AnfAlgo::GetOutputDeviceShape(tmp_output.first, tmp_output.second);
+      if (output_shape.empty()) {
+        output_shape.push_back(1);
+      }
+      output_desc_json[kShape] = output_shape;
+    }
+    outputs_json.emplace_back(output_desc_json);
+  }
+  fused_node_json[kOutputDesc] = outputs_json;
+
+  size_t hash_id = std::hash<std::string>()(fused_node_json.dump());
+  json_name_ = "Fused_";
+  auto fg = anf_nodes[0]->func_graph();
+  MS_EXCEPTION_IF_NULL(fg);
+  auto attr_val = fg->get_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL);
+  if (attr_val != nullptr) {
+    auto fg_attr = GetValue<std::string>(attr_val);
+    (void)json_name_.append(fg_attr).append("_");
+  }
+  (void)json_name_.append(std::to_string(hash_id));
+  fused_node_json["composite_graph"] = fg->ToString();
+  fused_node_json["op"] = json_name_;
+  fused_node_json["platform"] = "AKG";
+  fused_node_json["process"] = "aicore";
+  fused_node_json["composite"] = true;
+
+  kernel_json_ = fused_node_json.dump();
+
+  if (!GetIOSize(fused_node_json, &input_size_list_, &output_size_list_)) {
+    MS_LOG(ERROR) << "Cal mem size failed.";
+    return false;
+  }
+
+  return true;
+}
+
+void GenParallelCompileFuncArgs(const std::vector<std::string> &kernel_jsons, PyObject **p_args) {
+  MS_EXCEPTION_IF_NULL(p_args);
+  *p_args = PyTuple_New(PARALLEL_ARGS_SIZE);
+
+  PyObject *arg1 = PyList_New(kernel_jsons.size());
+  for (int i = 0; i < PyList_Size(arg1); ++i) {
+    PyList_SetItem(arg1, i, Py_BuildValue("s", kernel_jsons[i].c_str()));
+  }
+  PyObject *arg2 = Py_BuildValue("i", PROCESS_NUM);
+  PyObject *arg3 = Py_BuildValue("i", TIME_OUT);
+
+  (void)PyTuple_SetItem(*p_args, 0, arg1);
+  (void)PyTuple_SetItem(*p_args, 1, arg2);
+  (void)PyTuple_SetItem(*p_args, 2, arg3);
+}
+
+bool AkgOpParallelBuild(const std::vector<std::pair<AkgAscendKernelBuilder, AnfNodePtr>> &build_args) {
+  // Remove cached nodes, gether unique nodes, and collect repeated nodes which need postprecess.
+  std::vector<std::string> jsons;
+  std::unordered_set<std::string> json_name_set;
+  std::vector<std::pair<AkgAscendKernelBuilder, AnfNodePtr>> repeat_nodes;
+  for (const auto &[builder, anf_node] : build_args) {
+    MS_EXCEPTION_IF_NULL(anf_node);
+    auto json_name = builder.json_name();
+    MS_LOG(DEBUG) << "Akg start compile op: " << json_name;
+    auto cached_kernel_pack = tbe::TbeUtils::SearchCache(json_name, AkgKernelBuild::GetProcessor(anf_node));
+    if (cached_kernel_pack != nullptr) {
+      MS_LOG(DEBUG) << "Use cached kernel, json_name_[" << json_name << "], fullname_with_scope["
+                    << anf_node->fullname_with_scope() << "].";
+      auto kernel_mod_ptr = std::make_shared<AkgKernelMod>(cached_kernel_pack);
+      kernel_mod_ptr->SetInputSizeList(builder.input_size_list());
+      kernel_mod_ptr->SetOutputSizeList(builder.output_size_list());
+      AnfAlgo::SetKernelMod(kernel_mod_ptr, anf_node.get());
+      continue;
+    }
+
+    if (json_name_set.count(json_name) != 0) {
+      repeat_nodes.push_back({builder, anf_node});
+      continue;
+    }
+    json_name_set.insert(json_name);
+    auto node_json = builder.kernel_json();
+    kernel::SaveJsonInfo(json_name, node_json);
+    jsons.push_back(node_json);
+  }
+
+  // No nodes need to be compiled!
+  if (jsons.empty()) {
+    return true;
+  }
+
+  // Try to call python method to compile nodes parallely.
+  PyObject *p_module = nullptr;
+  PyObject *p_func = nullptr;
+  PyObject *p_arg = nullptr;
+  PyObject *p_res = nullptr;
+
+  p_module = PyImport_ImportModule(kMultiProcModule);
+  if (p_module == nullptr) {
+    MS_LOG(ERROR) << "Failed to import [" << kMultiProcModule << "].";
+    return false;
+  }
+
+  p_func = PyObject_GetAttrString(p_module, kCompileAkgKernelParallelFunc);
+  GenParallelCompileFuncArgs(jsons, &p_arg);
+  MS_LOG(DEBUG) << "Call function [" << kCompileAkgKernelParallelFunc << "], try to compile " << jsons.size()
+                << " Akg kernels parallelly.";
+  p_res = PyEval_CallObject(p_func, p_arg);
+  if (p_res == nullptr) {
+    PyErr_Print();
+    MS_LOG(ERROR) << "No ret got, failed to call function [" << kCompileAkgKernelParallelFunc << "], args:\n("
+                  << AkgKernelBuild::PyObjectToStr(p_arg) << ").";
+    return false;
+  }
+  if (PyObject_IsTrue(p_res) != 1) {
+    PyErr_Print();
+    MS_LOG(ERROR) << "Illegal ret, failed to call function [" << kCompileAkgKernelParallelFunc << "], args:\n("
+                  << AkgKernelBuild::PyObjectToStr(p_arg) << ").";
+    return false;
+  }
+
+  // All unique done here, cache them and set kernel.
+  for (const auto &[builder, anf_node] : build_args) {
+    auto json_name = builder.json_name();
+    auto new_kernel_pack = tbe::TbeUtils::InsertCache(json_name, AkgKernelBuild::GetProcessor(anf_node));
+    if (new_kernel_pack == nullptr) {
+      MS_LOG(ERROR) << "Insert to cache failed, json_name_[" << json_name << "], fullname_with_scope["
+                    << anf_node->fullname_with_scope() << "].";
+      return false;
+    }
+    auto kernel_mod_ptr = std::make_shared<AkgKernelMod>(new_kernel_pack);
+    kernel_mod_ptr->SetInputSizeList(builder.input_size_list());
+    kernel_mod_ptr->SetOutputSizeList(builder.output_size_list());
+    AnfAlgo::SetKernelMod(kernel_mod_ptr, anf_node.get());
+    MS_LOG(DEBUG) << "Akg compile " << json_name << " kernel and insert cache successfully!";
+  }
+
+  // Handle repeated nodes.
+  for (const auto &[builder, anf_node] : repeat_nodes) {
+    auto node_json = builder.kernel_json();
+    auto json_name = builder.json_name();
+    auto cached_kernel_pack = tbe::TbeUtils::SearchCache(json_name, AkgKernelBuild::GetProcessor(anf_node));
+    if (cached_kernel_pack == nullptr) return false;
+    MS_LOG(INFO) << "Use just compiled kernel, json_name_[" << json_name << "], fullname_with_scope["
+                 << anf_node->fullname_with_scope() << "].";
+    auto kernel_mod_ptr = std::make_shared<AkgKernelMod>(cached_kernel_pack);
+    kernel_mod_ptr->SetInputSizeList(builder.input_size_list());
+    kernel_mod_ptr->SetOutputSizeList(builder.output_size_list());
+    AnfAlgo::SetKernelMod(kernel_mod_ptr, anf_node.get());
+  }
+
+  return true;
+}
+
+bool AkgAscendKernelParallelBuild(const std::vector<AnfNodePtr> &anf_nodes) {
+  std::vector<std::pair<AkgAscendKernelBuilder, AnfNodePtr>> json_and_node;
+  for (const auto &anf_node : anf_nodes) {
+    MS_EXCEPTION_IF_NULL(anf_node);
+    AkgAscendKernelBuilder akg_cce_kernel_builder;
+    KernelPackPtr kernel_pack = nullptr;
+    auto cnode = anf_node->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    if (AnfAlgo::IsGraphKernel(cnode)) {
+      auto func_graph = AnfAlgo::GetCNodeFuncGraphPtr(cnode);
+      auto mng = func_graph->manager();
+      if (mng == nullptr) {
+        mng = Manage(func_graph, true);
+        func_graph->set_manager(mng);
+      }
+      MS_EXCEPTION_IF_NULL(func_graph);
+      std::vector<AnfNodePtr> node_list;
+      std::vector<AnfNodePtr> input_list;
+      std::vector<AnfNodePtr> output_list;
+      std::string op_name = AnfAlgo::GetCNodeName(anf_node);
+      MS_LOG(INFO) << "Akg start compile composite op[" << op_name << "]";
+      GetValidKernelNodes(func_graph, &node_list, &input_list, &output_list);
+      if (!akg_cce_kernel_builder.CollectFusedJson(node_list, input_list, output_list)) {
+        MS_EXCEPTION(UnknownError) << "Akg build failed composite op[" << op_name << "].";
+      }
+    } else {
+      if (!akg_cce_kernel_builder.CollectJson(anf_node)) {
+        MS_EXCEPTION(UnknownError) << "Akg build failed op[" << AnfAlgo::GetCNodeName(anf_node) << "].";
+      }
+    }
+    json_and_node.push_back({akg_cce_kernel_builder, anf_node});
+  }
+
+  if (json_and_node.empty()) {
+    MS_LOG(DEBUG) << "There is no kernel needed to be compiled.";
+    return true;
+  }
+
+  return AkgOpParallelBuild(json_and_node);
+}
+
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/kernel/akg/ascend/akg_ascend_kernel_build.h

@@ -0,0 +1,52 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_KERNEL_AKG_ASCEND_AKG_ASCEND_KERNEL_BUILD_H_
+#define MINDSPORE_CCSRC_KERNEL_AKG_ASCEND_AKG_ASCEND_KERNEL_BUILD_H_
+
+#include <string>
+#include <memory>
+#include <vector>
+#include "ir/anf.h"
+#include "kernel/kernel.h"
+#include "kernel/akg/akg_kernel_build.h"
+
+namespace mindspore {
+namespace kernel {
+class AkgAscendKernelBuilder : public AkgKernelBuild {
+ public:
+  AkgAscendKernelBuilder() = default;
+  ~AkgAscendKernelBuilder() = default;
+
+  bool CollectJson(const AnfNodePtr &anf_node);
+  bool CollectFusedJson(const std::vector<AnfNodePtr> &anf_nodes, const std::vector<AnfNodePtr> &input_list,
+                        const std::vector<AnfNodePtr> &output_list);
+  std::string json_name() const { return json_name_; }
+  std::string kernel_json() const { return kernel_json_; }
+  const std::vector<size_t> &input_size_list() const { return input_size_list_; }
+  const std::vector<size_t> &output_size_list() const { return output_size_list_; }
+
+ private:
+  std::string kernel_json_;
+  std::vector<size_t> input_size_list_;
+  std::vector<size_t> output_size_list_;
+};
+
+bool AkgAscendKernelParallelBuild(const std::vector<AnfNodePtr> &anf_nodes);
+}  // namespace kernel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_KERNEL_AKG_ASCEND_AKG_ASCEND_KERNEL_BUILD_H_

mindspore/ccsrc/kernel/akg/ascend/akg_ascend_kernel_mod.cc

@@ -0,0 +1,181 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "kernel/akg/ascend/akg_ascend_kernel_mod.h"
+#include <algorithm>
+#include <fstream>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+#include <vector>
+#include "nlohmann/json.hpp"
+#include "runtime/rt.h"
+#include "utils/log_adapter.h"
+#include "utils/convert_utils.h"
+
+namespace mindspore {
+namespace kernel {
+using std::fstream;
+using std::map;
+using std::mutex;
+using std::string;
+using TbeTaskInfoPtr = std::shared_ptr<ge::model_runner::TbeTaskInfo>;
+using tbe::KernelManager;
+constexpr uint32_t DEFAULT_BLOCK_DIM = 1;
+/**
+ * @brief infotable contain func_stub\blockdim\kernel file buffer
+ */
+AkgKernelMod::AkgKernelMod(const KernelPackPtr &kernel_pack) : kernel_pack_(kernel_pack) {}
+
+void AkgKernelMod::SetInputSizeList(const std::vector<size_t> &size_list) { input_size_list_ = size_list; }
+
+void AkgKernelMod::SetOutputSizeList(const std::vector<size_t> &size_list) { output_size_list_ = size_list; }
+
+void AkgKernelMod::SetWorkspaceSizeList(const std::vector<size_t> &size_list) { workspace_size_list_ = size_list; }
+
+const std::vector<size_t> &AkgKernelMod::GetInputSizeList() const { return input_size_list_; }
+
+const std::vector<size_t> &AkgKernelMod::GetOutputSizeList() const { return output_size_list_; }
+
+const std::vector<size_t> &AkgKernelMod::GetWorkspaceSizeList() const { return workspace_size_list_; }
+
+void DumpData(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &outputs) {
+  const char *dump_data = getenv("MS_KERNEL_DUMP_DATA");
+  if (dump_data) {
+    int idx = 0;
+    for (const auto &x : inputs) {
+      std::vector<char> buf(x->size);
+      if (RT_ERROR_NONE != rtMemcpy(buf.data(), buf.size(), reinterpret_cast<const void *>(x->addr), x->size,
+                                    RT_MEMCPY_DEVICE_TO_HOST)) {
+        MS_LOG(WARNING) << "Call runtime rtMemcpy error.";
+        return;
+      }
+
+      std::string file_name("input_");
+      file_name += std::to_string(idx);
+      std::ofstream file(file_name, std::ios::binary);
+      if (file.is_open()) {
+        (void)file.write(buf.data(), SizeToLong(buf.size()));
+        file.close();
+        idx++;
+      } else {
+        MS_LOG(ERROR) << "Open file failed.";
+        return;
+      }
+    }
+    idx = 0;
+    for (const auto &x : outputs) {
+      std::vector<char> buf(x->size);
+      if (RT_ERROR_NONE != rtMemcpy(buf.data(), buf.size(), reinterpret_cast<const void *>(x->addr), x->size,
+                                    RT_MEMCPY_DEVICE_TO_HOST)) {
+        MS_LOG(WARNING) << "Call runtime rtMemcpy error.";
+        return;
+      }
+
+      std::string file_name("output_");
+      file_name += std::to_string(idx);
+      std::ofstream file(file_name, std::ios::binary);
+      if (file.is_open()) {
+        (void)file.write(buf.data(), SizeToLong(buf.size()));
+        file.close();
+        idx++;
+      } else {
+        MS_LOG(ERROR) << "Open file failed.";
+        return;
+      }
+    }
+  }
+}
+
+bool AkgKernelMod::Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
+                          const std::vector<AddressPtr> &outputs, void *stream_ptr) {
+  if (stream_ptr == 0) {
+    MS_LOG(ERROR) << "stream_ptr should not be nullptr.";
+    return false;
+  }
+
+  if (kernel_pack_ == nullptr) {
+    MS_LOG(ERROR) << "kernel pack should not be nullptr.";
+    return false;
+  }
+
+  uint32_t block_dim = DEFAULT_BLOCK_DIM;  // default blockdim equal to 1.
+  auto func_stub = KernelManager::GenFuncStub(*kernel_pack_, false, &block_dim);
+  if (func_stub == 0) {
+    MS_LOG(ERROR) << "GenFuncStub failed.";
+    return false;
+  }
+
+  // pack all addresses into a vector.
+  std::vector<void *> runtime_args;
+  (void)std::transform(std::begin(inputs), std::end(inputs), std::back_inserter(runtime_args),
+                       [](const AddressPtr &input) -> void * { return input->addr; });
+  (void)std::transform(std::begin(outputs), std::end(outputs), std::back_inserter(runtime_args),
+                       [](const AddressPtr &output) -> void * { return output->addr; });
+
+  rtL2Ctrl_t *l2ctrl = nullptr;
+  auto stream = reinterpret_cast<rtStream_t *>(stream_ptr);
+  if (RT_ERROR_NONE != rtKernelLaunch(reinterpret_cast<void *>(func_stub), block_dim, runtime_args.data(),
+                                      SizeToUint(sizeof(void *) * runtime_args.size()), l2ctrl, stream)) {
+    MS_LOG(ERROR) << "Call runtime rtKernelLaunch error.";
+    return false;
+  }
+
+  DumpData(inputs, outputs);
+
+  return true;
+}
+
+std::vector<TaskInfoPtr> AkgKernelMod::GenTask(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
+                                               const std::vector<AddressPtr> &outputs, uint32_t stream_id) {
+  if (kernel_pack_ == nullptr) {
+    MS_LOG(EXCEPTION) << "kernel pack should not be nullptr.";
+  }
+
+  std::vector<uint8_t> args;
+  uint32_t args_size = 0;
+  std::vector<uint8_t> sm_desc;
+  void *binary = nullptr;
+  uint32_t binary_size = 0;
+  std::vector<uint8_t> meta_data;
+  std::vector<void *> input_data_addrs;
+  std::vector<void *> output_data_addrs;
+  std::vector<void *> workspace_addrs;
+
+  // pack all addresses into a vector.
+  (void)std::transform(std::begin(inputs), std::end(inputs), std::back_inserter(input_data_addrs),
+                       [](const AddressPtr &input) -> void * { return input->addr; });
+  (void)std::transform(std::begin(outputs), std::end(outputs), std::back_inserter(output_data_addrs),
+                       [](const AddressPtr &output) -> void * { return output->addr; });
+
+  uint32_t block_dim = DEFAULT_BLOCK_DIM;  // default blockdim equal to 1.
+  auto func_stub = KernelManager::GenFuncStub(*kernel_pack_, false, &block_dim);
+  if (func_stub == 0) {
+    MS_LOG(EXCEPTION) << "GenFuncStub failed.";
+  }
+
+  std::string stub_func = KernelManager::GetStubFuncName(kernel_pack_);
+
+  MS_LOG(DEBUG) << "The block_dim is:" << block_dim;
+
+  TbeTaskInfoPtr task_info_ptr = make_shared<ge::model_runner::TbeTaskInfo>(
+    stream_id, stub_func, block_dim, args, args_size, sm_desc, binary, binary_size, meta_data, input_data_addrs,
+    output_data_addrs, workspace_addrs);
+  return {task_info_ptr};
+}
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/kernel/akg/ascend/akg_ascend_kernel_mod.h

@@ -0,0 +1,54 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_KERNEL_AKG_ASCEND_AKG_ASCEND_KERNEL_MOD_H_
+#define MINDSPORE_CCSRC_KERNEL_AKG_ASCEND_AKG_ASCEND_KERNEL_MOD_H_
+#include <string>
+#include <vector>
+#include <memory>
+#include "kernel/ascend_kernel_mod.h"
+#include "kernel/tbe/tbe_utils.h"
+
+namespace mindspore {
+namespace kernel {
+class AkgKernelMod : public AscendKernelMod {
+ public:
+  explicit AkgKernelMod(const KernelPackPtr &kernel_pack);
+  ~AkgKernelMod() final {}
+
+  void SetInputSizeList(const std::vector<size_t> &size_list);
+  void SetOutputSizeList(const std::vector<size_t> &size_list);
+  void SetWorkspaceSizeList(const std::vector<size_t> &size_list);
+  const std::vector<size_t> &GetInputSizeList() const override;
+  const std::vector<size_t> &GetOutputSizeList() const override;
+  const std::vector<size_t> &GetWorkspaceSizeList() const override;
+  bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
+              const std::vector<AddressPtr> &outputs, void *stream_ptr) override;
+  std::vector<TaskInfoPtr> GenTask(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
+                                   const std::vector<AddressPtr> &outputs, uint32_t stream_id) override;
+
+ private:
+  KernelPackPtr kernel_pack_;
+  std::vector<size_t> input_size_list_;
+  std::vector<size_t> output_size_list_;
+  std::vector<size_t> workspace_size_list_;
+};
+
+using AkgKernelModPtr = std::shared_ptr<AkgKernelMod>;
+}  // namespace kernel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_KERNEL_AKG_ASCEND_AKG_ASCEND_KERNEL_MOD_H_

mindspore/ccsrc/kernel/akg/gpu/akg_gpu_kernel_build.cc

@@ -18,7 +18,7 @@
 #include <vector>
 #include <memory>
 #include "kernel/kernel.h"
-#include "kernel/akg/akgkernelbuild.h"
+#include "kernel/akg/akg_kernel_build.h"
 #include "kernel/akg/gpu/akg_gpu_kernel_mod.h"
 #include "common/utils.h"
 

mindspore/ccsrc/kernel/common_utils.cc

@@ -23,6 +23,11 @@
 #include "nlohmann/json.hpp"
 #include "session/anf_runtime_algorithm.h"
 #include "common/utils.h"
+#include "ir/manager.h"
+#include "ir/meta_tensor.h"
+#include "ir/func_graph.h"
+#include "operator/ops.h"
+#include "utils/graph_utils.h"
 
 namespace mindspore {
 namespace kernel {
@@ -48,12 +53,6 @@ const std::map<TypeId, std::string> type_id_str_map = {
   {TypeId::kNumberTypeBool, "bool"},
 };
 
-const std::map<std::string, std::string> DATATYPE_STRING_MAP{
-  {"Float32", "float32"}, {"Float16", "float16"}, {"Int8", "int8"},   {"Int16", "int16"},
-  {"UInt16", "uint16"},   {"UInt8", "uint8"},     {"Int32", "int32"}, {"UInt32", "uint32"},
-  {"Int64", "int64"},     {"UInt64", "uint64"},   {"Bool_", "bool"},  {"Float64", "double"},
-};
-
 const std::unordered_map<std::string, std::string> dtype_shortdtype_map_ = {
   {"float16", "f16"}, {"float32", "f32"}, {"float64", "f64"}, {"int8", "i8"},    {"int16", "i16"},  {"int32", "i32"},
   {"int64", "i64"},   {"uint8", "u8"},    {"uint16", "u16"},  {"uint32", "u32"}, {"uint64", "u64"}, {"bool", "bool"},
@@ -243,14 +242,6 @@ TypeId DtypeToTypeId(const std::string &dtypes) {
   }
 }
 
-std::string Dtype2String(const std::string &dtypes) {
-  auto iter = DATATYPE_STRING_MAP.find(dtypes);
-  if (iter == DATATYPE_STRING_MAP.end()) {
-    MS_EXCEPTION(ArgumentError) << "Illegal input dtype:" << dtypes;
-  }
-  return iter->second;
-}
-
 std::string TypeId2String(TypeId type_id) {
   auto iter = type_id_str_map.find(type_id);
   if (iter == type_id_str_map.end()) {
@@ -361,7 +352,7 @@ bool SetOutputKernelBuilderInfo(const std::vector<std::shared_ptr<OpIOInfo>> &ou
       output_num = 1;
     } else {
       if (output_idx < real_output_num) {
-        MS_LOG(INFO) << "Set output kernel builder info, output type is optional, output index is :" << output_idx;
+        MS_LOG(DEBUG) << "Set output kernel builder info, output type is optional, output index is :" << output_idx;
         output_num = 1;
       }
     }
@@ -403,7 +394,7 @@ void SetKernelBuildInfo(const std::shared_ptr<KernelBuildInfo::KernelBuildInfoBu
   }
 
   if (imply_type == kAKG) {
-    builder->SetKernelType(AUTO_DIFF_KERNEL);
+    builder->SetKernelType(AKG_KERNEL);
   } else if (imply_type == kAICPU) {
     builder->SetKernelType(AICPU_KERNEL);
   } else {
@@ -634,5 +625,256 @@ void ReduceSparseGradient(const SparseGradient &origin_sparse_grad, SparseGradie
   }
   unique_grad->indices_size_ = unique_indices_size + 1;
 }
+
+std::pair<AnfNodePtr, size_t> GetKernelInput(const AnfNodePtr &anf_node, size_t index) {
+  MS_EXCEPTION_IF_NULL(anf_node);
+
+  if (index >= AnfAlgo::GetInputTensorNum(anf_node)) {
+    MS_EXCEPTION(ArgumentError) << "Index is out of the size of anf_node inputs.";
+  }
+
+  auto cnode = anf_node->cast<CNodePtr>();
+  if (cnode == nullptr) {
+    return AnfAlgo::VisitKernel(anf_node, 0);
+  } else {
+    return AnfAlgo::VisitKernel(anf_node->cast<CNodePtr>()->input(index + 1), 0);
+  }
+}
+
+std::vector<std::pair<AnfNodePtr, std::pair<size_t, size_t>>> GetInputIndex(const std::vector<AnfNodePtr> &node_list,
+                                                                            const std::vector<AnfNodePtr> &input_list) {
+  std::vector<std::pair<AnfNodePtr, std::pair<size_t, size_t>>> input_index;
+  for (size_t i = 0; i < input_list.size(); ++i) {
+    auto const &input = input_list[i];
+    MS_EXCEPTION_IF_NULL(input);
+    bool found = false;
+    // using NodeUsersMap = std::unordered_map<AnfNodePtr, std::set<std::pair<AnfNodePtr, int>>>;
+    auto mng = input->func_graph()->manager();
+    MS_EXCEPTION_IF_NULL(mng);
+    const NodeUsersMap &users = mng->node_users();
+    auto input_users = users.find(input);
+    if (input_users == users.end() || input_users->second.empty()) {
+      MS_EXCEPTION(ArgumentError) << "Input [" << i << "][" << input->DebugString(2) << "] of ["
+                                  << input->func_graph()->ToString() << "] has no users.";
+    }
+
+    for (auto const &input_user : input_users->second) {
+      for (auto const &anf_node : node_list) {
+        if (anf_node != input_user.first) {
+          continue;
+        }
+
+        std::vector<int> dyn_input_sizes;
+        auto prim = AnfAlgo::GetCNodePrimitive(anf_node);
+        MS_EXCEPTION_IF_NULL(prim);
+        if (prim->GetAttr(kAttrDynInputSizes) != nullptr) {
+          dyn_input_sizes = GetValue<const std::vector<int>>(prim->GetAttr(kAttrDynInputSizes));
+        }
+
+        if (dyn_input_sizes.empty()) {
+          input_index.push_back(std::make_pair(anf_node, std::make_pair(IntToSize(input_user.second - 1), 0)));
+          found = true;
+          break;
+        } else {
+          int used_as_idx = input_user.second - 1;
+          int accum_idx = 0;
+          size_t dyn_i = 0;
+          for (; dyn_i < dyn_input_sizes.size(); ++dyn_i) {
+            accum_idx += dyn_input_sizes[dyn_i];
+            if (used_as_idx < accum_idx) {
+              input_index.push_back(std::make_pair(
+                anf_node, std::make_pair(dyn_i, IntToSize(used_as_idx - (accum_idx - dyn_input_sizes[dyn_i])))));
+              break;
+            }
+          }
+          if (dyn_i != dyn_input_sizes.size()) {
+            found = true;
+            break;
+          }
+        }
+      }
+      if (found) {
+        break;
+      }
+    }
+
+    if (!found) {
+      MS_EXCEPTION(ArgumentError) << "Input [" << i << "][" << input->DebugString(2) << "] of ["
+                                  << input->func_graph()->ToString() << "] found no related kernel info.";
+    }
+  }
+  return input_index;
+}
+
+std::vector<std::pair<AnfNodePtr, size_t>> GetOutputIndex(const std::vector<AnfNodePtr> &node_list,
+                                                          const std::vector<AnfNodePtr> &input_list,
+                                                          const std::vector<AnfNodePtr> &output_list) {
+  std::vector<std::pair<AnfNodePtr, size_t>> output_index;
+  for (size_t i = 0; i < output_list.size(); ++i) {
+    auto const &output = output_list[i];
+    MS_EXCEPTION_IF_NULL(output);
+    bool found = false;
+    auto pree_node = AnfAlgo::VisitKernel(output, 0);
+
+    auto pos = std::find(std::begin(node_list), std::end(node_list), pree_node.first);
+    if (pos != std::end(node_list)) {
+      output_index.push_back(pree_node);
+      continue;
+    }
+
+    auto ret = std::find(std::begin(input_list), std::end(input_list), pree_node.first);
+    if (ret != std::end(input_list)) {
+      output_index.push_back(std::make_pair(pree_node.first, 0));
+      found = true;
+    }
+
+    if (!found) {
+      MS_EXCEPTION(ArgumentError) << "Output [" << i << "][" << output->DebugString(2) << "] of ["
+                                  << output->func_graph()->ToString() << "] found no related kernel info.";
+    }
+  }
+  return output_index;
+}
+
+void GetValidKernelNodes(const FuncGraphPtr &func_graph, std::vector<AnfNodePtr> *node_list) {
+  MS_EXCEPTION_IF_NULL(node_list);
+
+  MS_EXCEPTION_IF_NULL(func_graph);
+
+  std::vector<AnfNodePtr> node_lists = TopoSort(func_graph->get_return());
+  for (auto const &node : node_lists) {
+    if (!AnfAlgo::IsRealKernel(node) || !node->isa<CNode>()) {
+      continue;
+    }
+
+    auto cnode = node->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+
+    if (IsValueNode<Primitive>(cnode->input(kAnfPrimitiveIndex))) {
+      node_list->push_back(node);
+    }
+  }
+}
+
+void GetValidKernelNodes(const FuncGraphPtr &func_graph, std::vector<AnfNodePtr> *node_list,
+                         std::vector<AnfNodePtr> *input_list, std::vector<AnfNodePtr> *output_list) {
+  MS_EXCEPTION_IF_NULL(node_list);
+  MS_EXCEPTION_IF_NULL(input_list);
+  MS_EXCEPTION_IF_NULL(output_list);
+  MS_EXCEPTION_IF_NULL(func_graph);
+
+  GetValidKernelNodes(func_graph, node_list);
+
+  auto parameters = func_graph->parameters();
+  input_list->insert(input_list->begin(), parameters.begin(), parameters.end());
+
+  auto func_output = func_graph->output();
+  MS_EXCEPTION_IF_NULL(func_output);
+  if (func_output->isa<CNode>()) {
+    // multi output.
+    auto cnode = func_output->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    auto input0 = cnode->input(kAnfPrimitiveIndex);
+    MS_EXCEPTION_IF_NULL(input0);
+    if (IsPrimitive(input0, prim::kPrimMakeTuple)) {
+      for (size_t input_idx = 1; input_idx < cnode->inputs().size(); ++input_idx) {
+        auto input_node = cnode->input(input_idx);
+        MS_EXCEPTION_IF_NULL(input_node);
+        output_list->push_back(AnfAlgo::VisitKernel(input_node, 0).first);
+      }
+    } else {
+      // single output.
+      output_list->push_back(AnfAlgo::VisitKernel(func_output, 0).first);
+    }
+  } else {
+    // single output.
+    output_list->push_back(AnfAlgo::VisitKernel(func_output, 0).first);
+  }
+}
+
+bool GetInputTensorValue(const AnfNodePtr &anf_node, size_t input_idx, nlohmann::json *const node_json) {
+  MS_EXCEPTION_IF_NULL(anf_node);
+  MS_EXCEPTION_IF_NULL(node_json);
+  auto cnode = anf_node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(cnode);
+  if (input_idx + 1 >= cnode->size()) {
+    MS_EXCEPTION(ArgumentError) << "input_idx [" << input_idx << "] is out of index of inputs of ["
+                                << cnode->inputs().size() << "][" << cnode->DebugString() << "]";
+  }
+
+  auto input_node = cnode->input(input_idx + 1);
+  if (!IsValueNode<tensor::Tensor>(input_node)) {
+    return false;
+  }
+
+  auto tensor = GetValueNode<tensor::TensorPtr>(input_node);
+  if (tensor == nullptr) {
+    return false;
+  }
+
+  auto type_id = tensor->data_type();
+  auto *data = tensor->data_c();
+  MS_EXCEPTION_IF_NULL(data);
+  if (tensor->DataDim() > 1 || tensor->DataSize() != 1) {
+    // not const tensor.
+    MS_LOG(WARNING) << "We take first value of tensor whose datasize != 1, [" << input_node->DebugString(2) << "]";
+  }
+
+  if (type_id == kFloat32->type_id()) {
+    float *val = static_cast<float *>(data);
+    MS_EXCEPTION_IF_NULL(val);
+    (*node_json)["value"] = val[0];
+    MS_LOG(DEBUG) << "Value of tensor[" << cnode->DebugString() << "] is [float32][" << *val << "].";
+    return true;
+  } else if (type_id == kFloat16->type_id()) {
+    float16 *val = static_cast<float16 *>(data);
+    MS_EXCEPTION_IF_NULL(val);
+    (*node_json)["value"] = static_cast<float>(val[0]);
+    MS_LOG(INFO) << "Value of tensor[" << cnode->DebugString() << "] is [float16][" << *val << "].";
+    return true;
+  } else if (type_id == kInt32->type_id()) {
+    int *val = static_cast<int *>(data);
+    MS_EXCEPTION_IF_NULL(val);
+    (*node_json)["value"] = val[0];
+    MS_LOG(INFO) << "Value of tensor[" << cnode->DebugString() << "] is [int32][" << *val << "].";
+    return true;
+  }
+  MS_LOG(ERROR) << "Unknown value type of tensor[" << cnode->DebugString() << "]";
+  return false;
+}
+
+void GetGraphRealOutput(const FuncGraphPtr &func_graph, std::vector<std::pair<AnfNodePtr, size_t>> *node_list) {
+  MS_EXCEPTION_IF_NULL(func_graph);
+  MS_EXCEPTION_IF_NULL(node_list);
+  auto output = func_graph->output();
+  MS_EXCEPTION_IF_NULL(output);
+  if (AnfAlgo::IsRealKernel(output)) {
+    // single output.
+    node_list->push_back(std::make_pair(output, 0));
+    return;
+  } else if (IsPrimitiveCNode(output, prim::kPrimMakeTuple)) {
+    auto output_cnode = output->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(output_cnode);
+    // multi output.
+    auto &inputs = output_cnode->inputs();
+    for (size_t i = 1; i < inputs.size(); ++i) {
+      auto in_with_idx = AnfAlgo::VisitKernel(inputs[i], 0);
+      node_list->push_back(in_with_idx);
+    }
+    return;
+  }
+  MS_EXCEPTION(ArgumentError) << "Unknown  output type: " << output->DebugString(2)
+                              << " of graph: " << func_graph->ToString();
+}
+
+bool IsWeightBoundary(const AnfNodePtr &node) {
+  if (node->isa<ValueNode>()) {
+    return true;
+  }
+  if (node->isa<Parameter>() && AnfAlgo::IsParameterWeight(node->cast<ParameterPtr>())) {
+    return true;
+  }
+  return false;
+}
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/common_utils.h

@@ -20,9 +20,12 @@
 #include <dirent.h>
 #include <memory>
 #include <unordered_map>
+#include <unordered_set>
 #include <map>
 #include <string>
 #include <vector>
+#include <utility>
+#include <nlohmann/json.hpp>
 #include "kernel/kernel.h"
 #include "kernel/oplib/opinfo.h"
 #include "kernel/kernel_build_info.h"
@@ -79,13 +82,11 @@ bool CheckCache(const std::string &kernel_name);
 KernelPackPtr SearchCache(const std::string &kernel_name, const std::string &processor);
 KernelPackPtr InsertCache(const std::string &kernel_name, const std::string &processor);
 TypeId DtypeToTypeId(const std::string &dtypes);
-std::string Dtype2String(const std::string &dtypes);
 std::string Dtype2ShortType(const std::string &dtypes);
 std::string TypeId2String(TypeId type_id);
 size_t GetDtypeNbyte(const std::string &dtypes);
 bool ParseMetadata(const CNodePtr &kernel_node, const std::shared_ptr<const OpInfo> &op_info_ptr, Processor processor,
                    std::vector<std::shared_ptr<KernelBuildInfo>> *const kernel_info_list);
-bool IsAtomicNode(const CNodePtr &kernel_node);
 void SaveJsonInfo(const std::string &json_name, const std::string &info);
 std::string GetProcessor(const AnfNodePtr &anf_node);
 bool IsSameShape(const std::vector<size_t> &shape_a, const std::vector<size_t> &shape_b);
@@ -94,6 +95,18 @@ void DeduplicateIndexedSlices(const SparseGradient &origin_sparse_grad, SparseGr
                               size_t outer_dim);
 void ReduceSparseGradient(const SparseGradient &origin_sparse_grad, SparseGradient *unique_grad, size_t first_dim,
                           size_t outer_dim);
+std::pair<AnfNodePtr, size_t> GetKernelInput(const AnfNodePtr &anf_node, size_t index);
+std::vector<std::pair<AnfNodePtr, std::pair<size_t, size_t>>> GetInputIndex(const std::vector<AnfNodePtr> &node_list,
+                                                                            const std::vector<AnfNodePtr> &input_list);
+std::vector<std::pair<AnfNodePtr, size_t>> GetOutputIndex(const std::vector<AnfNodePtr> &node_list,
+                                                          const std::vector<AnfNodePtr> &input_list,
+                                                          const std::vector<AnfNodePtr> &output_list);
+void GetValidKernelNodes(const FuncGraphPtr &func_graph, std::vector<AnfNodePtr> *node_list,
+                         std::vector<AnfNodePtr> *input_list, std::vector<AnfNodePtr> *output_list);
+void GetValidKernelNodes(const FuncGraphPtr &func_graph, std::vector<AnfNodePtr> *node_list);
+bool GetInputTensorValue(const AnfNodePtr &anf_node, size_t input_idx, nlohmann::json *const node_json);
+void GetGraphRealOutput(const FuncGraphPtr &func_graph, std::vector<std::pair<AnfNodePtr, size_t>> *node_list);
+bool IsWeightBoundary(const AnfNodePtr &node);
 }  // namespace kernel
 }  // namespace mindspore
 

mindspore/ccsrc/kernel/kash/kernel_pack.cc

@@ -17,7 +17,7 @@
 #include <fstream>
 #include "mindspore/ccsrc/kernel/kernel.h"
 #include "kernel/kernel.h"
-#include "kernel/akg/akgkernelbuild.h"
+#include "kernel/akg/akg_kernel_build.h"
 #include "nlohmann/json.hpp"
 #include "securec/include/securec.h"
 #include "pipeline/parse/python_adapter.h"

mindspore/ccsrc/kernel/kernel.h

@@ -27,7 +27,7 @@
 #include "utils/log_adapter.h"
 
 namespace mindspore {
-enum KernelType : int { UNKNOWN_KERNEL_TYPE = 0, AUTO_DIFF_KERNEL, AICPU_KERNEL, RT_KERNEL, HCCL_KERNEL, TBE_KERNEL };
+enum KernelType : int { UNKNOWN_KERNEL_TYPE = 0, AKG_KERNEL, AICPU_KERNEL, RT_KERNEL, HCCL_KERNEL, TBE_KERNEL };
 
 namespace kernel {
 

mindspore/ccsrc/kernel/kernel_query.cc

@@ -21,6 +21,7 @@
 #include "kernel/rts/rt_kernel_info.h"
 #include "kernel/hccl/hccl_kernel_metadata.h"
 #include "kernel/tbe/tbe_kernel_select/tbe_kernel_select.h"
+#include "kernel/akg/akg_kernel_metadata.h"
 #include "session/anf_runtime_algorithm.h"
 
 namespace mindspore {
@@ -59,10 +60,14 @@ void FilterInvalidKernelInfo(const CNodePtr &kernel_node,
   }
 }
 }  // namespace
-void KernelQuery(const CNodePtr &kernel_node, std::vector<std::shared_ptr<kernel::KernelBuildInfo>> *kernel_info_list) {
+
+void KernelQueryAll(const CNodePtr &kernel_node,
+                    std::vector<std::shared_ptr<kernel::KernelBuildInfo>> *kernel_info_list) {
   MS_EXCEPTION_IF_NULL(kernel_node);
   MS_EXCEPTION_IF_NULL(kernel_info_list);
+
   TbeMetadataInfo(kernel_node, kernel_info_list);
+
   if (kernel_info_list->empty()) {
     AicpuMetadataInfo(kernel_node, kernel_info_list);
     if (!kernel_info_list->empty()) {
@@ -82,6 +87,28 @@ void KernelQuery(const CNodePtr &kernel_node, std::vector<std::shared_ptr<kernel
   if (kernel_info_list->empty()) {
     MS_LOG(EXCEPTION) << "Op " << kernel_node->DebugString() << "kernel query fail!";
   }
+}
+
+void KernelQuery(const CNodePtr &kernel_node, std::vector<std::shared_ptr<kernel::KernelBuildInfo>> *kernel_info_list,
+                 KernelType kernel_type) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  MS_EXCEPTION_IF_NULL(kernel_info_list);
+
+  std::string op_name = AnfAlgo::GetCNodeName(kernel_node);
+
+  switch (kernel_type) {
+    case KernelType::AKG_KERNEL:
+      AkgMetadataInfo(kernel_node, kernel_info_list);
+      break;
+    default:
+      KernelQueryAll(kernel_node, kernel_info_list);
+      break;
+  }
+
+  if (kernel_info_list->empty()) {
+    MS_EXCEPTION(NotExistsError) << "Op[" << kernel_node->DebugString() << "] kernel query fail!";
+  }
+  // check output
   FilterInvalidKernelInfo(kernel_node, kernel_info_list);
 }
 

mindspore/ccsrc/kernel/kernel_query.h

@@ -25,7 +25,8 @@
 
 namespace mindspore {
 namespace kernel {
-void KernelQuery(const CNodePtr &kernel_node, std::vector<std::shared_ptr<kernel::KernelBuildInfo>> *kernel_info_list);
+void KernelQuery(const CNodePtr &kernel_node, std::vector<std::shared_ptr<kernel::KernelBuildInfo>> *kernel_info_list,
+                 KernelType kernel_type = KernelType::UNKNOWN_KERNEL_TYPE);
 void AICPUQuery(const CNodePtr &kernel_node, std::vector<std::shared_ptr<kernel::KernelBuildInfo>> *kernel_info_list);
 bool IsSupportedByAICPU(const AnfNodePtr &kernel_node, const KernelBuildInfoPtr &select_kernel_build_info);
 bool IsSupportedByAICore(const AnfNodePtr &kernel_node, const KernelBuildInfoPtr &select_kernel_build_info);

mindspore/ccsrc/kernel/oplib/oplib.cc

@@ -272,8 +272,7 @@ std::shared_ptr<OpInfo> OpLib::FindOp(const std::string &op_name, OpImplyType im
   auto context = MsContext::GetInstance();
   MS_EXCEPTION_IF_NULL(context);
   bool is_gpu = (context->device_target() == kGPUDevice);
-  if ((is_gpu && (imply_type == kTBE || imply_type == kAICPU)) ||
-      (!is_gpu && (imply_type != kTBE && imply_type != kAICPU))) {
+  if (is_gpu && (imply_type == kTBE || imply_type == kAICPU)) {
     MS_LOG(ERROR) << "FindOp failed: opname: " << op_name << ", imply_type: " << ImplTypeToStr(imply_type)
                   << ", current op num: " << op_info_.size();
     return nullptr;

mindspore/ccsrc/kernel/tbe/tbe_adapter.cc

@@ -347,7 +347,7 @@ static int TypeStrToDstType(const std::string &type_str) {
     ret = 4;
   } else if (type_str == "UInt64") {
     ret = 10;
-  } else if (type_str == "Bool_") {
+  } else if (type_str == "Bool") {
     ret = 12;
   } else {
     MS_LOG(INFO) << "Error type str is invailed: " << type_str;

mindspore/ccsrc/kernel/tbe/tbe_convert_utils.cc

@@ -51,7 +51,7 @@ const std::map<TypeId, std::string> type_id_str_maps = {
 const std::map<std::string, std::string> type_str_maps = {
   {"Float32", "float32"}, {"Float16", "float16"}, {"Int8", "int8"},   {"Int16", "int16"},
   {"UInt16", "uint16"},   {"UInt8", "uint8"},     {"Int32", "int32"}, {"UInt32", "uint32"},
-  {"Int64", "int64"},     {"UInt64", "uint64"},   {"Bool_", "int8"},  {"Float64", "float64"},
+  {"Int64", "int64"},     {"UInt64", "uint64"},   {"Bool", "int8"},   {"Float64", "float64"},
 };
 
 const std::unordered_map<std::string, size_t> type_nbyte_maps = {

mindspore/ccsrc/operator/composite/composite.cc

@@ -334,8 +334,8 @@ ArgsPairList HyperMap::Harmonize(const FuncGraphPtr &func_graph, const ArgsPairL
 
 FuncGraphPtr HyperMap::GenerateFromTypes(const TypePtrList &args_spec_list) {
   FuncGraphPtr ptrGraph = std::make_shared<FuncGraph>();
-  ptrGraph->set_flags(FUNC_GRAPH_FLAG_CORE, true);
-  ptrGraph->set_flags(FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER, true);
+  ptrGraph->set_flag(FUNC_GRAPH_FLAG_CORE, true);
+  ptrGraph->set_flag(FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER, true);
   ptrGraph->debug_info()->set_name("hyper_map");
 
   AnfNodePtr ptrFnArg = nullptr;
@@ -389,7 +389,7 @@ FuncGraphPtr Tail::GenerateTupleFuncGraph(const abstract::AbstractTuplePtr &a_tu
   MS_EXCEPTION_IF_NULL(a_tuple);
 
   FuncGraphPtr ret = std::make_shared<FuncGraph>();
-  ret->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   ret->debug_info()->set_name("tail");
   AnfNodePtr ptrTup = ret->add_parameter();
 
@@ -409,7 +409,7 @@ FuncGraphPtr Tail::GenerateListFuncGraph(const abstract::AbstractListPtr &a_list
   MS_EXCEPTION_IF_NULL(a_list);
 
   FuncGraphPtr ret = std::make_shared<FuncGraph>();
-  ret->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   ret->debug_info()->set_name("tail");
   AnfNodePtr ptrList = ret->add_parameter();
 
@@ -481,10 +481,10 @@ FuncGraphPtr MakeTupleGradient::GenerateFuncGraph(const AbstractBasePtrList &arg
     grads.push_back(b->NewCNode({NewValueNode(prim::kPrimTupleGetItem), dout, NewValueNode(i)}));
   }
 
-  b->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  b->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   b->set_output(b->NewCNode(grads));
 
-  fg->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  fg->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   fg->set_output(fg->NewCNode({NewValueNode(prim::kPrimMakeTuple), out, NewValueNode(b)}));
   (void)fg->transforms().emplace("primal", FuncGraphTransform(prim::kPrimMakeTuple));
   return fg;
@@ -504,7 +504,7 @@ FuncGraphPtr GradOperation::GetGrad(AnfNodePtr node, const AnfNodePtr &weights,
                                     const std::vector<AnfNodePtr> &params_list, const std::vector<AnfNodePtr> &args,
                                     bool applyJ) {
   FuncGraphPtr ret = std::make_shared<FuncGraph>();
-  ret->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret->set_flag(FUNC_GRAPH_FLAG_CORE, true);
 
   auto weights_node = weights;
   if (weights == nullptr && !args.empty()) {
@@ -625,7 +625,7 @@ FuncGraphPtr GradOperation::GenerateFuncGraph(const AbstractBasePtrList &args_sp
 
   std::ostringstream ss;
   ss << "grad{" << nparam << "}";
-  dfBuilder->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  dfBuilder->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   dfBuilder->debug_info()->set_name(ss.str());
   ParameterPtr param_graph = dfBuilder->add_parameter();
 
@@ -671,7 +671,7 @@ FuncGraphPtr ListMap::GenerateFuncGraph(const AbstractBasePtrList &args_spec_lis
   }
 
   FuncGraphPtr fg_ptr = std::make_shared<FuncGraph>();
-  fg_ptr->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  fg_ptr->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   fg_ptr->debug_info()->set_name("list_map");
   AnfNodePtr fn = fg_ptr->add_parameter();
 
@@ -741,7 +741,7 @@ void ListMap::MakeCond(const std::vector<AnfNodePtr> &lists, const FuncGraphPtr
   // cond = reduce(lambda a, b: g.apply(P.bool_and, a, b), hasnexts)
   FuncGraphPtr fgtrue_ptr = std::make_shared<FuncGraph>();
   fgtrue_ptr->debug_info()->set_name("ftrue");
-  fgtrue_ptr->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  fgtrue_ptr->set_flag(FUNC_GRAPH_FLAG_CORE, true);
 
   CNodePtr fgtrue_output_cnode = fgtrue_ptr->NewCNode({NewValueNode(fgnext_ptr), fn, resl});
   auto inputs = fgtrue_output_cnode->inputs();
@@ -751,7 +751,7 @@ void ListMap::MakeCond(const std::vector<AnfNodePtr> &lists, const FuncGraphPtr
 
   FuncGraphPtr fgfalse_ptr = std::make_shared<FuncGraph>();
   fgfalse_ptr->debug_info()->set_name("ffalse");
-  fgfalse_ptr->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  fgfalse_ptr->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   fgfalse_ptr->set_output(resl);
 
   AnfNodePtr output_cnode = fg_ptr->NewCNode({NewValueNode(prim::kPrimSwitch), NewValueNode(std::string("cond")),
@@ -808,7 +808,7 @@ FuncGraphPtr TupleAdd::GenerateFuncGraph(const AbstractBasePtrList &args_spec_li
   }
 
   FuncGraphPtr ret = std::make_shared<FuncGraph>();
-  ret->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   AnfNodePtr p_tup_a = ret->add_parameter();
   AnfNodePtr p_tup_b = ret->add_parameter();
 
@@ -912,7 +912,7 @@ FuncGraphPtr TupleSlice::GenerateFuncGraph(const AbstractBasePtrList &args_spec_
   GenerateTupleSliceParameter(tuple, slice, &start_index, &stop_index, &step_value);
 
   FuncGraphPtr ret = std::make_shared<FuncGraph>();
-  ret->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   AnfNodePtr p_tuple = ret->add_parameter();
   (void)ret->add_parameter();
 
@@ -941,7 +941,7 @@ FuncGraphPtr TupleGetItemTensor::GenerateFuncGraph(const AbstractBasePtrList &ar
   AbstractBasePtrList branches = branches_abs->elements();
   if (branches.size() > 0 && branches[0] != nullptr && branches[0]->isa<AbstractFunction>()) {
     FuncGraphPtr ret_graph = std::make_shared<FuncGraph>();
-    ret_graph->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+    ret_graph->set_flag(FUNC_GRAPH_FLAG_CORE, true);
     AnfNodePtr functions = ret_graph->add_parameter();
     auto index = ret_graph->add_parameter();
 

mindspore/ccsrc/operator/composite/do_signature.cc

@@ -304,7 +304,7 @@ FuncGraphPtr DoSignatureMetaFuncGraph::GenerateFuncGraph(const AbstractBasePtrLi
   }
   auto new_cnode = BuildNewCNode(func_graph, name_, function_, args_spec_list, func_graph->parameters());
   func_graph->set_output(new_cnode);
-  func_graph->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  func_graph->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   return func_graph;
 }
 }  // namespace prim

mindspore/ccsrc/operator/composite/list_append_operation.cc

@@ -35,7 +35,7 @@ FuncGraphPtr ListAppend::GenerateFuncGraph(const abstract::AbstractBasePtrList &
   MS_EXCEPTION_IF_NULL(arg0_list);
 
   FuncGraphPtr ret = std::make_shared<FuncGraph>();
-  ret->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   ret->debug_info()->set_name("append");
   AnfNodePtr arg0_node = ret->add_parameter();
 

mindspore/ccsrc/operator/composite/map.cc

@@ -51,8 +51,8 @@ AnfNodePtr Map::FullMakeLeaf(const FuncGraphPtr &func_graph, const AnfNodePtr &f
 FuncGraphPtr Map::GenerateLeafFunc(const size_t &args_size) {
   // Generate func for leaf nodes
   FuncGraphPtr ptrGraph = std::make_shared<FuncGraph>();
-  ptrGraph->set_flags(FUNC_GRAPH_FLAG_CORE, true);
-  ptrGraph->set_flags(FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER, true);
+  ptrGraph->set_flag(FUNC_GRAPH_FLAG_CORE, true);
+  ptrGraph->set_flag(FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER, true);
   ptrGraph->debug_info()->set_name("map");
   AnfNodePtr ptrFnArg = nullptr;
   if (fn_leaf_ == nullptr) {
@@ -237,8 +237,8 @@ AnfNodePtr Map::Make(const FuncGraphPtr &func_graph, const AnfNodePtr &fn_arg, c
 
 FuncGraphPtr Map::GenerateFromTypes(const TypePtrList &args_spec_list) {
   FuncGraphPtr ptrGraph = std::make_shared<FuncGraph>();
-  ptrGraph->set_flags(FUNC_GRAPH_FLAG_CORE, true);
-  ptrGraph->set_flags(FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER, true);
+  ptrGraph->set_flag(FUNC_GRAPH_FLAG_CORE, true);
+  ptrGraph->set_flag(FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER, true);
   ptrGraph->debug_info()->set_name("map");
 
   AnfNodePtr ptrFnArg = nullptr;

mindspore/ccsrc/operator/composite/unpack_call.cc

@@ -51,7 +51,7 @@ FuncGraphPtr UnpackCall::GenerateFuncGraph(const AbstractBasePtrList &args_spec_
 
   (void)abstract::CheckArg<AbstractFunction>(op_name, args_spec_list, 0);
   auto ret_graph = std::make_shared<FuncGraph>();
-  ret_graph->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret_graph->set_flag(FUNC_GRAPH_FLAG_CORE, true);
 
   AnfNodePtr fnNode = ret_graph->add_parameter();
   std::vector<AnfNodePtr> elems;

mindspore/ccsrc/operator/composite/zip_operation.cc

@@ -57,7 +57,7 @@ FuncGraphPtr ZipOperation::GenerateFuncGraph(const AbstractBasePtrList &args_spe
                                     return (x->cast<AbstractTuplePtr>()->size() < y->cast<AbstractTuplePtr>()->size());
                                   });
   FuncGraphPtr ret_graph = std::make_shared<FuncGraph>();
-  ret_graph->set_flags(FUNC_GRAPH_FLAG_CORE, true);
+  ret_graph->set_flag(FUNC_GRAPH_FLAG_CORE, true);
   for (size_t idx = 0; idx < args_spec_list.size(); idx++) {
     (void)ret_graph->add_parameter();
   }

mindspore/ccsrc/operator/ops.cc

@@ -50,6 +50,12 @@ const PrimitivePtr kPrimBoolNot = std::make_shared<Primitive>("bool_not");
 const PrimitivePtr kPrimBoolAnd = std::make_shared<Primitive>("bool_and");
 const PrimitivePtr kPrimBoolOr = std::make_shared<Primitive>("bool_or");
 const PrimitivePtr kPrimBoolEq = std::make_shared<Primitive>("bool_eq");
+const PrimitivePtr kPrimGreater = std::make_shared<Primitive>("Greater");
+const PrimitivePtr kPrimGreaterEqual = std::make_shared<Primitive>("GreaterEqual");
+const PrimitivePtr kPrimLess = std::make_shared<Primitive>("Less");
+const PrimitivePtr kPrimLessEqual = std::make_shared<Primitive>("LessEqual");
+const PrimitivePtr kPrimEqual = std::make_shared<Primitive>("Equal");
+const PrimitivePtr kPrimNotEqual = std::make_shared<Primitive>("NotEqual");
 
 // Type introspection
 const PrimitivePtr kPrimTypeOf = std::make_shared<Primitive>("typeof");
@@ -166,17 +172,20 @@ const PrimitivePtr kPrimMul = std::make_shared<Primitive>("Mul");
 const PrimitivePtr kPrimMinimum = std::make_shared<Primitive>("Minimum");
 const PrimitivePtr kPrimMaximum = std::make_shared<Primitive>("Maximum");
 const PrimitivePtr kPrimSquare = std::make_shared<Primitive>("Square");
-const PrimitivePtr kPrimEqual = std::make_shared<Primitive>("Equal");
-const PrimitivePtr kPrimLess = std::make_shared<Primitive>("Less");
-const PrimitivePtr kPrimLessEqual = std::make_shared<Primitive>("LessEqual");
 const PrimitivePtr kPrimCumSum = std::make_shared<Primitive>("CumSum");
 const PrimitivePtr kPrimCumProd = std::make_shared<Primitive>("CumProd");
 const PrimitivePtr kPrimSubscalar = std::make_shared<Primitive>("Subscalar");
 const PrimitivePtr kPrimInplaceAdd = std::make_shared<Primitive>("InplaceAdd");
 const PrimitivePtr kPrimInplaceSub = std::make_shared<Primitive>("InplaceSub");
+const PrimitivePtr kPrimPow = std::make_shared<Primitive>("Pow");
+const PrimitivePtr kPrimRealDiv = std::make_shared<Primitive>("RealDiv");
+const PrimitivePtr kPrimSqrt = std::make_shared<Primitive>("Sqrt");
+const PrimitivePtr kPrimReciprocal = std::make_shared<Primitive>("Reciprocal");
+const PrimitivePtr kPrimExpandDims = std::make_shared<Primitive>("ExpandDims");
 
 // NN
 const PrimitivePtr kPrimFlatten = std::make_shared<Primitive>("Flatten");
+const PrimitivePtr kPrimSoftmax = std::make_shared<Primitive>("Softmax");
 const PrimitivePtr kPrimLogSoftmax = std::make_shared<Primitive>("LogSoftmax");
 const PrimitivePtr kPrimLogSoftmaxGrad = std::make_shared<Primitive>("LogSoftmaxGrad");
 const PrimitivePtr kPrimTanh = std::make_shared<Primitive>("Tanh");
@@ -253,6 +262,7 @@ const PrimitivePtr kPrimInDict = std::make_shared<Primitive>("in_dict");
 const PrimitivePtr kPrimNotInDict = std::make_shared<Primitive>("not_in_dict");
 const PrimitivePtr kPrimMixedPrecisionCast = std::make_shared<Primitive>("mixed_precision_cast");
 const PrimitivePtr kPrimIsConsant = std::make_shared<Primitive>("is_constant");
+const PrimitivePtr kPrimEquivFormat = std::make_shared<Primitive>("EquivFormat");
 
 // Comm ops
 const PrimitivePtr kPrimMirror = std::make_shared<Primitive>("_MirrorOperator");

mindspore/ccsrc/operator/ops.h

@@ -59,6 +59,12 @@ extern const PrimitivePtr kPrimBoolNot;
 extern const PrimitivePtr kPrimBoolAnd;
 extern const PrimitivePtr kPrimBoolOr;
 extern const PrimitivePtr kPrimBoolEq;
+extern const PrimitivePtr kPrimGreater;
+extern const PrimitivePtr kPrimGreaterEqual;
+extern const PrimitivePtr kPrimLess;
+extern const PrimitivePtr kPrimLessEqual;
+extern const PrimitivePtr kPrimEqual;
+extern const PrimitivePtr kPrimNotEqual;
 
 // Type introspection
 extern const PrimitivePtr kPrimTypeOf;
@@ -157,6 +163,10 @@ extern const PrimitivePtr KPrimTransData;
 extern const PrimitivePtr kPrimNMSWithMask;
 extern const PrimitivePtr kPrimPad;
 extern const PrimitivePtr kPrimArgMaxWithValue;
+extern const PrimitivePtr kPrimRealDiv;
+extern const PrimitivePtr kPrimSqrt;
+extern const PrimitivePtr kPrimReciprocal;
+extern const PrimitivePtr kPrimExpandDims;
 
 // Maths
 extern const PrimitivePtr kPrimTensorAdd;
@@ -183,9 +193,11 @@ extern const PrimitivePtr kPrimCumProd;
 extern const PrimitivePtr kPrimSubscalar;
 extern const PrimitivePtr kPrimInplaceAdd;
 extern const PrimitivePtr kPrimInplaceSub;
+extern const PrimitivePtr kPrimPow;
 
 // NN
 extern const PrimitivePtr kPrimFlatten;
+extern const PrimitivePtr kPrimSoftmax;
 extern const PrimitivePtr kPrimLogSoftmax;
 extern const PrimitivePtr kPrimLogSoftmaxGrad;
 extern const PrimitivePtr kPrimApplyCenteredRMSProp;
@@ -263,6 +275,7 @@ extern const PrimitivePtr kPrimInDict;
 extern const PrimitivePtr kPrimNotInDict;
 extern const PrimitivePtr kPrimMixedPrecisionCast;
 extern const PrimitivePtr kPrimIsConsant;
+extern const PrimitivePtr kPrimEquivFormat;
 
 // Comm ops
 extern const PrimitivePtr kPrimAllReduce;

mindspore/ccsrc/optimizer/ad/dfunctor.cc

@@ -45,10 +45,19 @@ DFunctor::DFunctor(const FuncGraphPtr &primal_graph, const pipeline::ResourceBas
     : primal_graph_(primal_graph), resources_(resources), need_cut_(false), is_top_(false) {
   TraceManager::DebugTrace(std::make_shared<TraceGradFprop>(primal_graph->debug_info()));
   k_graph_ = std::make_shared<FuncGraph>();
+  if (primal_graph->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+    std::string grad_op_name = GetValue<std::string>(primal_graph->get_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL));
+    k_graph_->set_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL, MakeValue(grad_op_name));
+  }
   TraceManager::EndTrace();
 
   TraceManager::DebugTrace(std::make_shared<TraceGradBprop>(primal_graph->debug_info()));
   tape_ = std::make_shared<FuncGraph>();
+  // Add "_Grad" postfix
+  if (primal_graph->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+    std::string grad_op_name = GetValue<std::string>(primal_graph->get_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) + "_Grad";
+    tape_->set_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL, MakeValue(grad_op_name));
+  }
   TraceManager::EndTrace();
 
   dout_ = tape_->add_parameter();
@@ -368,7 +377,7 @@ FuncGraphPtr DFunctor::KUserDefined(const FuncGraphPtr &primal) {
     (void)primal->transforms().insert(std::make_pair("grad", FuncGraphTransform(fg)));
     (void)fg->transforms().insert(std::make_pair("primal", FuncGraphTransform(primal)));
     // Reset defer_inline to enable successive inlining
-    primal->set_flags(FUNC_GRAPH_FLAG_DEFER_INLINE, false);
+    primal->set_flag(FUNC_GRAPH_FLAG_DEFER_INLINE, false);
 
     auto functor = std::make_shared<DFunctor>(primal, resources_);
     functor->Init();

mindspore/ccsrc/optimizer/ad/grad.cc

@@ -37,7 +37,7 @@ FuncGraphPtr Grad(const FuncGraphPtr &func_graph, const pipeline::ResourceBasePt
   auto multi_graph_sink = [&func_graph](const FuncGraphPtr &f) {
     if (MsContext::GetInstance()->is_multi_graph_sink()) {
       if (func_graph->has_flag(FUNC_GRAPH_FLAG_IGNORE_VALUES)) {
-        f->set_flags(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
+        f->set_flag(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
       }
     }
   };

mindspore/ccsrc/optimizer/clean.cc

@@ -78,7 +78,10 @@ AnfNodePtr ConvertGetAttrToTupleGetItem(const CNodePtr &node) {
   MS_EXCEPTION_IF_NULL(cons);
 
   auto dt = data->abstract();
-  MS_EXCEPTION_IF_NULL(dt);
+  if (dt == nullptr) {
+    return nullptr;
+  }
+
   if (!dt->isa<AbstractClass>()) {
     MS_LOG(EXCEPTION) << "First parameter of getattr is not AbstractClass, but " << dt->type_name() << ".";
   }

mindspore/ccsrc/optimizer/graph_kernel_reuse.cc

@@ -0,0 +1,157 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "optimizer/graph_kernel_reuse.h"
+#include <vector>
+#include <algorithm>
+#include <string>
+#include "./common.h"
+#include "utils/graph_utils.h"
+
+namespace mindspore {
+/* namespace to support opt */
+namespace opt {
+
+bool GraphKernelReuse::CompareNode(const AnfNodePtr a, const AnfNodePtr b) {
+  if (a->abstract() && b->abstract()) {
+    auto a_type = a->abstract()->GetTypeTrack();
+    auto b_type = b->abstract()->GetTypeTrack();
+
+    if (a_type != b_type) {
+      return false;
+    }
+
+    auto a_shape = a->abstract()->GetShapeTrack();
+    auto b_shape = b->abstract()->GetShapeTrack();
+    if (a_shape != nullptr && a_shape == b_shape) {
+      return true;
+    }
+
+    if (a_shape != nullptr && b_shape != nullptr && a_shape->isa<abstract::Shape>() &&
+        b_shape->isa<abstract::Shape>()) {
+      return a_shape->cast<abstract::ShapePtr>()->shape() == b_shape->cast<abstract::ShapePtr>()->shape();
+    }
+  }
+  return false;
+}
+
+bool GraphKernelReuse::DoReplace(const FuncGraphManagerPtr manager) {
+  bool changed = false;
+  auto fgs = manager->func_graphs();
+  for (FuncGraphPtr &fg : fgs) {
+    if (!fg->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+      continue;
+    }
+    std::string key = GetValue<std::string>(fg->get_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL));
+    if (graph_kernel_ops.find(key) != graph_kernel_ops.end()) {
+      if (find(graph_kernel_ops[key].begin(), graph_kernel_ops[key].end(), fg) == graph_kernel_ops[key].end()) {
+        FuncGraphPtr new_fg = nullptr;
+        for (auto &cfg : graph_kernel_ops[key]) {
+          // If two graphs have different size then continue
+          auto fg_topos = TopoSort(fg->get_return());
+          auto cfg_topos = TopoSort(cfg->get_return());
+          if (fg_topos.size() != cfg_topos.size()) {
+            continue;
+          }
+
+          // Compare const tensor
+          bool has_same = true;
+          for (size_t i = 0; i < fg_topos.size(); ++i) {
+            if (IsValueNode<tensor::Tensor>(fg_topos[i])) {
+              if (!IsValueNode<tensor::Tensor>(cfg_topos[i])) {
+                has_same = false;
+                break;
+              }
+
+              auto tensor1 = GetValueNode<tensor::TensorPtr>(fg_topos[i]);
+              auto tensor2 = GetValueNode<tensor::TensorPtr>(cfg_topos[i]);
+              if (!tensor1->ValueEqual(*tensor2)) {
+                has_same = false;
+                break;
+              }
+            }
+          }
+
+          if (!has_same) {
+            continue;
+          }
+
+          auto fg_input = fg->parameters();
+          auto cfg_input = cfg->parameters();
+          if (fg_input.size() != cfg_input.size()) {
+            continue;
+          }
+          // Compare input
+          for (size_t i = 0; i < fg_input.size(); ++i) {
+            if (!CompareNode(fg_input[i], cfg_input[i])) {
+              has_same = false;
+              break;
+            }
+          }
+          if (!has_same) {
+            continue;
+          }
+
+          // Compare output
+          if (!CompareNode(fg->output(), cfg->output())) {
+            continue;
+          }
+
+          // Find reusable fg
+          new_fg = cfg;
+          break;
+        }
+
+        if (new_fg != nullptr) {
+          // Replace current fg with existing fg
+          auto users = fg->func_graph_cnodes_index();
+          for (auto &iter : users) {
+            auto cnode = iter.first->first->cast<CNodePtr>();
+            auto new_input = cnode->inputs();
+            auto main_graph = cnode->func_graph();
+            MS_EXCEPTION_IF_NULL(main_graph);
+            if (IsPrimitiveCNode(cnode, prim::kPrimPartial)) {
+              new_input[1] = NewValueNode(new_fg);
+            } else {
+              new_input[0] = NewValueNode(new_fg);
+            }
+            auto new_cnode = main_graph->NewCNode(new_input);
+            manager->Replace(iter.first->first, new_cnode);
+            changed = true;
+          }
+
+        } else {
+          // Add current fg to map
+          graph_kernel_ops[key].push_back(fg);
+        }
+      }
+    } else {
+      graph_kernel_ops[key] = {fg};
+    }
+  }
+
+  return changed;
+}
+
+bool GraphKernelReuse::ReuseGraphKernel(const FuncGraphPtr root, const FuncGraphManagerPtr manager) {
+  MS_EXCEPTION_IF_NULL(manager);
+  manager->AddFuncGraph(root);
+
+  return DoReplace(manager);
+}
+
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/optimizer/graph_kernel_reuse.h

@@ -0,0 +1,53 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_OPTIMIZER_GRAPH_KERNEL_OP_REUSE_H
+#define MINDSPORE_CCSRC_OPTIMIZER_GRAPH_KERNEL_OP_REUSE_H
+
+#include <mindspore/ccsrc/session/anf_runtime_algorithm.h>
+#include <unordered_map>
+#include <string>
+#include <vector>
+
+#include "optimizer/optimizer.h"
+
+namespace mindspore {
+namespace opt {
+
+// Common subexpression elimination.
+class GraphKernelReuse {
+ public:
+  GraphKernelReuse() : count(0) {}
+  virtual ~GraphKernelReuse() = default;
+
+  bool operator()(const FuncGraphPtr &root, const OptimizerPtr &optimizer) {
+    bool chg = ReuseGraphKernel(root, optimizer->resource()->manager());
+    return chg;
+  }
+
+  bool CompareNode(const AnfNodePtr a, const AnfNodePtr other);
+  bool DoReplace(const FuncGraphManagerPtr manager);
+
+  bool ReuseGraphKernel(const FuncGraphPtr root, const FuncGraphManagerPtr manager);
+
+ private:
+  std::unordered_map<std::string, std::vector<FuncGraphPtr>> graph_kernel_ops;
+  int count;
+};
+
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_OPTIMIZER_GRAPH_KERNEL_OP_REUSE_H

mindspore/ccsrc/optimizer/irpass.cc

@@ -41,6 +41,8 @@
 #include "optimizer/irpass/incorporate_call.h"
 #include "optimizer/irpass/grad_var_prepare.h"
 #include "optimizer/irpass/param_replace.h"
+#include "optimizer/irpass/mark_interface_fusion.h"
+#include "optimizer/opt.h"
 
 namespace mindspore {
 namespace opt {
@@ -48,7 +50,7 @@ namespace irpass {
 OptimizeIRPassLib::OptimizeIRPassLib() {
   arithmetic_simplify_ = MakeSubstitution(ArithmeticSimplify(), "arithmetic_simplify",
                                           {prim::kPrimScalarAdd, prim::kPrimScalarMul, prim::kPrimTensorAdd,
-                                           prim::kPrimIdentity, prim::kPrimMomentum, prim::kPrimMul});
+                                           prim::kPrimIdentity, prim::kPrimMomentum, prim::kPrimMul, prim::kPrimPow});
   special_op_eliminate_ =
     MakeSubstitution(SpecialOpEliminater(), "special_op_eliminate",
                      {prim::kPrimInsertGradientOf, prim::kPrimStopGradient, prim::kPrimHookBackward,
@@ -90,7 +92,6 @@ OptimizeIRPassLib::OptimizeIRPassLib() {
   replace_refkey_by_param_ =
     MakeSubstitution(ReplaceRefkeyByParam(), "replace_refkey_by_param", IsValueNode<RefKey>, opt::FORCE_RENORM);
   replace_old_param_ = MakeSubstitution(ReplaceOldParam(), "replace_old_param", IsParam);
-
   // Gradient transforms
   expand_jprim_ = MakeSubstitution(ExpandJPrim(), "expand_jprim", prim::kPrimJ);
   minmaximum_grad_ = MakeSubstitution(MinMaximumGrad(), "minmaximum_grad", prim::kPrimTupleGetItem);
@@ -115,6 +116,8 @@ OptimizeIRPassLib::OptimizeIRPassLib() {
   // Incorporation
   incorporate_getitem_set_ =
     MakeSubstitution(IncorporateGetitemSet(), "incorporate_getitem_set", prim::kPrimTupleGetItem);
+  incorporate_getitem_from_param_ =
+    MakeSubstitution(IncorporateGetitemFromParam(), "incorporate_getitem_from_param", IsCNodeGraphKernel);
   incorporate_call_ = MakeSubstitution(IncorporateCall(), "incorporate_call", IsCNodeDup);
   incorporate_call_switch_ = MakeSubstitution(IncorporateCallSwitch(), "incorporate_call_switch", IsCNodeDup);
 
@@ -124,6 +127,17 @@ OptimizeIRPassLib::OptimizeIRPassLib() {
 
   // Convert
   print_tuple_wrapper_ = MakeSubstitution(PrintTupleWrapper(), "print_tuple_wrapper", prim::kPrimPrint);
+
+  // Unused parameter eliminate
+  unused_parameter_eliminate_ =
+    MakeSubstitution(UnusedParasEliminater(), "unused_parameter_eliminate", IsCNodeGraphKernel);
+  unused_output_eliminate_ = MakeSubstitution(UnusedOutputEliminater(), "unused_output_eliminate", IsCNodeGraphKernel);
+
+  // AddN eliminate
+  addn_eliminate_ = MakeSubstitution(AddNEliminater(), "addn_eliminate", IsCNodeGraphKernel);
+
+  // Mark interface fusion
+  mark_interface_fusion_ = MakeSubstitution(MarkInterfaceFusion(), "mark_interface_fusion", prim::kPrimSelect);
 }
 
 ResolveIRPassLib::ResolveIRPassLib() {

mindspore/ccsrc/optimizer/irpass.h

@@ -84,6 +84,7 @@ class OptimizeIRPassLib {
 
   // Incorporation
   SubstitutionPtr incorporate_getitem_set_;
+  SubstitutionPtr incorporate_getitem_from_param_;
   SubstitutionPtr incorporate_call_;
   SubstitutionPtr incorporate_call_switch_;
 
@@ -92,6 +93,16 @@ class OptimizeIRPassLib {
 
   // Convert
   SubstitutionPtr print_tuple_wrapper_;
+
+  // Unused parameter eliminate
+  SubstitutionPtr unused_parameter_eliminate_;
+  SubstitutionPtr unused_output_eliminate_;
+
+  // AddN eliminate
+  SubstitutionPtr addn_eliminate_;
+
+  // Fusion
+  SubstitutionPtr mark_interface_fusion_;
 };
 
 // the collection of irpass for resolve action
@@ -145,6 +156,23 @@ inline bool IsCNodeGraph(const AnfNodePtr &node) {
   return IsValueNode<FuncGraph>(inp0);
 }
 
+// Check if CNode Input 0 is Func Graph of graph kernel.
+inline bool IsCNodeGraphKernel(const AnfNodePtr &node) {
+  if (node == nullptr || !node->isa<CNode>()) {
+    return false;
+  }
+
+  auto inp0 = node->cast<CNodePtr>()->input(0);
+  if (IsValueNode<FuncGraph>(inp0)) {
+    auto fg = GetValueNode<FuncGraphPtr>(inp0);
+    if (fg == nullptr) {
+      return false;
+    }
+    return fg->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL);
+  }
+  return false;
+}
+
 // Check if CNode Input 0 is CNode
 inline bool IsCNodeDup(const AnfNodePtr &node) {
   if (node == nullptr || !node->isa<CNode>()) {

mindspore/ccsrc/optimizer/irpass/arithmetic_simplify.h

@@ -83,6 +83,216 @@ class MultiplyByZeroOrOne : public AnfVisitor {
   AnfNodePtr x_{nullptr};
 };
 
+// Support class used for checking if all values of a Tensor are equal `check_value_`
+// Supported data types: double, float/float32, int/int32
+class CheckTensorConstant {
+ public:
+  explicit CheckTensorConstant(int _check_value = 0) : check_value_(_check_value) {}
+  ~CheckTensorConstant() = default;
+  bool IsTensorConstant(const ValuePtr &value) {
+    if (!value->isa<tensor::Tensor>()) {
+      return false;
+    }
+    auto tensor_ptr = dyn_cast<tensor::Tensor>(value);
+    TypeId tensor_type = tensor_ptr->Dtype()->type_id();
+    if ((tensor_type == TypeId::kNumberTypeFloat32) || (tensor_type == TypeId::kNumberTypeFloat)) {
+      float *data2 = reinterpret_cast<float *>(tensor_ptr->data_c());
+      for (int i = 0; i < tensor_ptr->DataSize(); i++) {
+        if (fabs(data2[i] - check_value_) > FLT_EPSILON) {
+          return false;
+        }
+      }
+      return true;
+    } else if (tensor_type == TypeId::kNumberTypeFloat64) {
+      double *data2 = reinterpret_cast<double *>(tensor_ptr->data_c());
+      for (int i = 0; i < tensor_ptr->DataSize(); i++) {
+        if (fabs(data2[i] - check_value_) > DBL_EPSILON) {
+          return false;
+        }
+      }
+      return true;
+    } else if ((tensor_type == TypeId::kNumberTypeInt32) || (tensor_type == TypeId::kNumberTypeInt)) {
+      int *data2 = reinterpret_cast<int *>(tensor_ptr->data_c());
+      for (int i = 0; i < tensor_ptr->DataSize(); i++) {
+        if (data2[i] != check_value_) {
+          return false;
+        }
+      }
+      return true;
+    }
+    // Un-support Data Types
+    return false;
+  }
+
+  bool IsTensorScalarConstant(const ValuePtr &value) {
+    if (!value->isa<tensor::Tensor>()) {
+      return false;
+    }
+    auto tensor_ptr = dyn_cast<tensor::Tensor>(value);
+    if ((tensor_ptr->DataSize() > 1) || (tensor_ptr->DataDim() > 0)) {
+      return false;
+    }
+    return IsTensorConstant(value);
+  }
+
+ private:
+  int check_value_;
+};
+
+// {prim::kPrimMul, 0, X}, {prim::kPrimMul, X, 0}
+// {prim::kPrimMul, 1, X}, {prim::kPrimMul, X, 1}
+class TensorMultiplyByZeroOrOne : public AnfVisitor {
+ public:
+  TensorMultiplyByZeroOrOne() : zero_(MakeValue(0)) {}
+  ~TensorMultiplyByZeroOrOne() override = default;
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    Reset();
+    AnfVisitor::Match(prim::kPrimMul)(node);
+
+    if (is_zero_) {
+      if (x_->func_graph() != node->func_graph()) {
+        return nullptr;
+      }
+      return NewTensorFilledWithData(node);
+    }
+    if (is_one_) {
+      return NewTensorFilledWithData(node, x_);
+    }
+    return nullptr;
+  }
+
+  void Visit(const AnfNodePtr &node) override {
+    if (is_zero_ || is_one_) {
+      x_ = node;
+      return;
+    }
+
+    if (IsParam(node)) {
+      x_ = node;
+      return;
+    }
+
+    if (IsCNode(node)) {
+      CNodePtr cnode = node->cast<CNodePtr>();
+      if (IsPrimitive(cnode->input(0), prim::kPrimZerosLike)) {
+        is_zero_ = true;
+        return;
+      }
+      x_ = node;
+      return;
+    }
+    auto value = node->cast<ValueNodePtr>()->value();
+    if (CheckTensorConstant(0).IsTensorConstant(value)) {
+      is_zero_ = true;
+      return;
+    } else if (CheckTensorConstant(1).IsTensorConstant(value)) {
+      is_one_ = true;
+      return;
+    }
+    x_ = node;
+  }
+
+  void Visit(const ValueNodePtr &vnode) override {
+    auto value = vnode->value();
+    if (CheckTensorConstant(0).IsTensorConstant(value)) {
+      is_zero_ = true;
+      return;
+    } else if (CheckTensorConstant(1).IsTensorConstant(value)) {
+      is_one_ = true;
+      return;
+    }
+    x_ = vnode;
+  }
+  void Reset() {
+    x_ = nullptr;
+    is_one_ = false;
+    is_zero_ = false;
+  }
+
+  void *GetPointerToTensorData(const AnfNodePtr &node, bool writable = false) {
+    if (!node->isa<ValueNode>()) {
+      return nullptr;
+    }
+
+    auto value = node->cast<ValueNodePtr>()->value();
+
+    if (!value->isa<tensor::Tensor>()) {
+      return nullptr;
+    }
+
+    tensor::TensorPtr tensor_ptr = dyn_cast<tensor::Tensor>(value);
+    return tensor_ptr->data_c(writable);
+  }
+
+  // Make a new tensor (when possible) with the same shape as of `node`
+  // If x is nullptr then fill new tensor will "0"
+  // If x is a tensor with empty shape then fill new tensor with the single value of x
+  // If x is a tensor with same shape as `node` then return x as result
+  AnfNodePtr NewTensorFilledWithData(const AnfNodePtr &node, const AnfNodePtr &x = nullptr) {
+    if ((node->abstract() == nullptr) || !node->abstract()->isa<abstract::AbstractTensor>()) {
+      return nullptr;
+    }
+
+    auto tensor_abstract = node->abstract()->cast<abstract::AbstractTensorPtr>();
+    TypePtr tensor_type_ptr = tensor_abstract->element()->BuildType();
+    std::vector<int> tensor_shape = tensor_abstract->shape()->shape();
+
+    auto new_tensor_ptr = std::make_shared<tensor::Tensor>(tensor_type_ptr->type_id(), tensor_shape);
+    size_t mem_size = GetTypeByte(tensor_type_ptr) * IntToSize(new_tensor_ptr->ElementsNum());
+    char *data = reinterpret_cast<char *>(new_tensor_ptr->data_c(true));
+
+    if (x == nullptr) {
+      std::memset(data, 0, mem_size);
+      auto new_vnode = NewValueNode(new_tensor_ptr);
+      new_vnode->set_abstract(new_tensor_ptr->ToAbstract());
+      return new_vnode;
+    }
+    // x is not nullptr
+    if (x->isa<CNode>()) {
+      if ((x->abstract() == nullptr) || !x->abstract()->isa<abstract::AbstractTensor>()) {
+        return nullptr;
+      }
+      auto x_abstract = x->abstract()->cast<abstract::AbstractTensorPtr>();
+      std::vector<int> x_shape = x_abstract->shape()->shape();
+
+      if (x_shape != tensor_shape) {
+        return nullptr;
+      }
+      return x;
+    }
+
+    if (!x->isa<ValueNode>()) {
+      return nullptr;
+    }
+    auto x_value = x->cast<ValueNodePtr>()->value();
+    if (!x_value->isa<tensor::Tensor>()) {
+      return nullptr;
+    }
+
+    auto x_tensor_ptr = dyn_cast<tensor::Tensor>(x_value);
+
+    if ((x_tensor_ptr->DataSize() > 1) && (x_tensor_ptr->DataSize() != new_tensor_ptr->DataSize())) {
+      return nullptr;
+    }
+    char *source_data = reinterpret_cast<char *>(GetPointerToTensorData(x));
+    if (x_tensor_ptr->DataSize() == 1) {
+      for (int i = 0; i < new_tensor_ptr->ElementsNum(); i++) {
+        memcpy(source_data, data + i * GetTypeByte(tensor_type_ptr), GetTypeByte(tensor_type_ptr));
+      }
+    } else {
+      memcpy(source_data, data, mem_size);
+    }
+    auto new_vnode = NewValueNode(new_tensor_ptr);
+    new_vnode->set_abstract(new_tensor_ptr->ToAbstract());
+    return new_vnode;
+  }
+
+ private:
+  bool is_zero_{false}, is_one_{false};
+  ValuePtr zero_;
+  AnfNodePtr x_{nullptr};
+};
+
 // {prim::kPrimScalarAdd, X, 0}
 // {prim::kPrimScalarAdd, 0, X}
 class AddByZero : public AnfVisitor {
@@ -101,7 +311,8 @@ class AddByZero : public AnfVisitor {
   }
 
   void Visit(const AnfNodePtr &node) override {
-    if (node->isa<ValueNode>() && *GetValueNode(node) == *zero_) {
+    if (node->isa<ValueNode>() &&
+        ((*GetValueNode(node) == *zero_) || CheckTensorConstant(0).IsTensorScalarConstant(GetValueNode(node)))) {
       is_zero_ = true;
       return;
     }
@@ -139,10 +350,22 @@ class TensorAddByZero : public AnfVisitor {
       is_zero_ = true;
       return;
     }
+    if (node->isa<ValueNode>() && CheckTensorConstant(0).IsTensorScalarConstant(GetValueNode(node))) {
+      is_zero_ = true;
+      return;
+    }
 
     x_ = node;
   }
 
+  void Visit(const ValueNodePtr &vnode) override {
+    auto value = vnode->value();
+    if (CheckTensorConstant(0).IsTensorConstant(value)) {
+      is_zero_ = true;
+      return;
+    }
+  }
+
   void Reset() {
     x_ = nullptr;
     is_zero_ = false;
@@ -183,29 +406,143 @@ class OptUpdateZeroTensor : public AnfVisitor {
 // {prim::kPrimMul, {...}, {prim::kPrimMul, Tensor1, Tensor2}}
 class ConstantDuplicateMul : public AnfVisitor {
  public:
+  // Support function to multiply two constant tensors: partially support broadcasting shapes
+  template <typename T>
+  void Multiply(void *in_data_1, int in_data_1_size, void *in_data_2, int in_data_2_size, void **out_data,
+                int out_data_size) {
+    T *data_1 = reinterpret_cast<T *>(in_data_1);
+    T *data_2 = reinterpret_cast<T *>(in_data_2);
+    T *data_out = new T[out_data_size];
+
+    if (in_data_1_size == 1) {
+      for (int i = 0; i < out_data_size; i++) {
+        data_out[i] = data_1[0];
+      }
+    } else {
+      for (int i = 0; i < out_data_size; i++) {
+        data_out[i] = data_1[i];
+      }
+    }
+    if (in_data_2_size == 1) {
+      for (int i = 0; i < out_data_size; i++) {
+        data_out[i] *= data_2[0];
+      }
+    } else {
+      for (int i = 0; i < out_data_size; i++) {
+        data_out[i] *= data_2[i];
+      }
+    }
+    *out_data = reinterpret_cast<void *>(data_out);
+    return;
+  }
+
+  AnfNodePtr MulConstantTensors(const AnfNodePtr &vnode_1, const AnfNodePtr &vnode_2, const AnfNodePtr &node_3) {
+    if (!vnode_1->isa<ValueNode>() || !vnode_2->isa<ValueNode>() || (vnode_1->abstract() == nullptr) ||
+        (vnode_2->abstract() == nullptr) || (node_3->abstract() == nullptr)) {
+      return nullptr;
+    }
+
+    auto value_1 = GetValueNode(vnode_1);
+    auto value_2 = GetValueNode(vnode_2);
+
+    if (!value_1->isa<tensor::Tensor>() || !value_2->isa<tensor::Tensor>()) {
+      return nullptr;
+    }
+
+    auto tensor_ptr_1 = dyn_cast<tensor::Tensor>(value_1);
+    auto tensor_ptr_2 = dyn_cast<tensor::Tensor>(value_2);
+
+    auto tensor_1_abstract = vnode_1->abstract()->cast<abstract::AbstractTensorPtr>();
+    auto tensor_2_abstract = vnode_1->abstract()->cast<abstract::AbstractTensorPtr>();
+    auto tensor_3_abstract = node_3->abstract()->cast<abstract::AbstractTensorPtr>();
+
+    TypePtr tensor_1_type_ptr = tensor_1_abstract->element()->BuildType();
+    TypePtr tensor_2_type_ptr = tensor_2_abstract->element()->BuildType();
+    TypePtr tensor_3_type_ptr = tensor_3_abstract->element()->BuildType();
+
+    if ((tensor_1_type_ptr->type_id() != tensor_3_type_ptr->type_id()) ||
+        (tensor_2_type_ptr->type_id() != tensor_3_type_ptr->type_id())) {
+      return nullptr;
+    }
+
+    std::vector<int> tensor_out_shape = tensor_3_abstract->shape()->shape();
+
+    int data_out_size = 1;
+    for (auto it : tensor_out_shape) {
+      data_out_size *= it;
+    }
+    if ((tensor_ptr_1->DataSize() > 1) && (tensor_ptr_1->DataSize() != data_out_size)) {
+      return nullptr;
+    }
+    if ((tensor_ptr_2->DataSize() > 1) && (tensor_ptr_2->DataSize() != data_out_size)) {
+      return nullptr;
+    }
+
+    void *data_out;
+
+    if ((tensor_3_type_ptr->type_id() == TypeId::kNumberTypeFloat32) ||
+        (tensor_3_type_ptr->type_id() == TypeId::kNumberTypeFloat)) {
+      Multiply<float>(tensor_ptr_1->data_c(), tensor_ptr_1->DataSize(), tensor_ptr_2->data_c(),
+                      tensor_ptr_2->DataSize(), &data_out, data_out_size);
+    } else {
+      if (tensor_3_type_ptr->type_id() == TypeId::kNumberTypeFloat64) {
+        Multiply<double>(tensor_ptr_1->data_c(), tensor_ptr_1->DataSize(), tensor_ptr_2->data_c(),
+                         tensor_ptr_2->DataSize(), &data_out, data_out_size);
+      } else {
+        if ((tensor_3_type_ptr->type_id() == TypeId::kNumberTypeInt32) ||
+            (tensor_3_type_ptr->type_id() == TypeId::kNumberTypeInt)) {
+          Multiply<int>(tensor_ptr_1->data_c(), tensor_ptr_1->DataSize(), tensor_ptr_2->data_c(),
+                        tensor_ptr_2->DataSize(), &data_out, data_out_size);
+        } else {
+          // Un-support data types
+          return nullptr;
+        }
+      }
+    }
+
+    auto new_tensor_ptr = std::make_shared<tensor::Tensor>(tensor_3_type_ptr->type_id(), tensor_out_shape);
+    size_t mem_size = GetTypeByte(tensor_3_type_ptr) * IntToSize(new_tensor_ptr->ElementsNum());
+    char *data = reinterpret_cast<char *>(new_tensor_ptr->data_c(true));
+    memcpy(data, data_out, mem_size);
+
+    auto new_vnode = NewValueNode(new_tensor_ptr);
+    new_vnode->set_abstract(new_tensor_ptr->ToAbstract());
+    return new_vnode;
+  }
+
   AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
     Reset();
     // {prim::kPrimMul, Tensor1, {...}}
     AnfVisitor::Match(prim::kPrimMul, {IsNode, IsNode})(node);
-    if (vnode_ == nullptr || cnode_ == nullptr) {
+    if (vnode_ == nullptr || c_p_node_ == nullptr) {
       return nullptr;
     }
+
+    if (!IsCNode(c_p_node_)) {
+      return nullptr;
+    }
+
     auto tensor1 = vnode_;
-    auto mul = cnode_;
+    auto mul = c_p_node_->cast<CNodePtr>();
 
     Reset();
     // {prim::kPrimMul, Tensor2, {...}}
     AnfVisitor::Match(prim::kPrimMul, {IsNode, IsNode})(mul);
-    if (vnode_ == nullptr || cnode_ == nullptr) {
+    if (vnode_ == nullptr || c_p_node_ == nullptr) {
       return nullptr;
     }
     auto tensor2 = vnode_;
-    auto cnode = cnode_;
+    auto c_p_node = c_p_node_;
 
     auto PrimMul = GetValueNode<PrimitivePtr>(mul->input(0));
     auto fg = node->func_graph();
-    auto ttmul = NewCNode({NewValueNode(PrimMul), tensor1, tensor2}, fg);
-    return NewCNode({NewValueNode(PrimMul), cnode, ttmul}, fg);
+
+    auto new_mul_tensor = MulConstantTensors(tensor1, tensor2, c_p_node);
+    if (new_mul_tensor == nullptr) {
+      auto ttmul = NewCNode({NewValueNode(PrimMul), tensor1, tensor2}, fg);
+      return NewCNode({NewValueNode(PrimMul), c_p_node, ttmul}, fg);
+    }
+    return NewCNode({NewValueNode(PrimMul), c_p_node, new_mul_tensor}, fg);
   }
 
   void Visit(const AnfNodePtr &node) override {
@@ -213,19 +550,40 @@ class ConstantDuplicateMul : public AnfVisitor {
       vnode_ = node;
     }
 
-    if (IsCNode(node)) {
-      cnode_ = node->cast<CNodePtr>();
+    if (IsCNode(node) || IsParam(node)) {
+      c_p_node_ = node;
     }
   }
 
   void Reset() {
     vnode_ = nullptr;
-    cnode_ = nullptr;
+    c_p_node_ = nullptr;
   }
 
  private:
   AnfNodePtr vnode_;
-  CNodePtr cnode_;
+  AnfNodePtr c_p_node_;
+};
+
+class PowerOneEliminate : public AnfVisitor {
+ public:
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    if (!IsPrimitiveCNode(node, prim::kPrimPow) || node->func_graph() == nullptr) {
+      return nullptr;
+    }
+
+    auto &inputs = node->cast<CNodePtr>()->inputs();
+    if (!IsValueNode<Scalar>(inputs[2])) {
+      return nullptr;
+    }
+    auto scalar = GetValueNode<ScalarPtr>(inputs[2]);
+    if (scalar->isa<FloatImm>() && GetValue<float>(scalar) == 1.0) {
+      return inputs[1];
+    } else if (scalar->isa<IntergerImm>() && GetValue<int>(scalar) == 1) {
+      return inputs[1];
+    }
+    return nullptr;
+  }
 };
 
 // grad = AllReduce(grad) / worker_number
@@ -341,17 +699,21 @@ class ArithmeticSimplify {
  public:
   ArithmeticSimplify()
       : multiply_by_zero_or_one_(),
+        tensor_multiply_by_zero_or_one_(),
         add_by_zero_(),
         tensor_add_by_zero_(),
         identity_(prim::kPrimIdentity),
         opt_update_zero_tensor_(),
-        constant_duplicate_mul_() {
+        constant_duplicate_mul_(),
+        power_one_() {
     eliminaters_.emplace_back(multiply_by_zero_or_one_);
+    eliminaters_.emplace_back(tensor_multiply_by_zero_or_one_);
     eliminaters_.emplace_back(add_by_zero_);
     eliminaters_.emplace_back(tensor_add_by_zero_);
     eliminaters_.emplace_back(identity_);
     eliminaters_.emplace_back(opt_update_zero_tensor_);
     eliminaters_.emplace_back(constant_duplicate_mul_);
+    eliminaters_.emplace_back(power_one_);
   }
   ~ArithmeticSimplify() = default;
 
@@ -368,11 +730,13 @@ class ArithmeticSimplify {
 
  private:
   MultiplyByZeroOrOne multiply_by_zero_or_one_;
+  TensorMultiplyByZeroOrOne tensor_multiply_by_zero_or_one_;
   AddByZero add_by_zero_;
   TensorAddByZero tensor_add_by_zero_;
   PrimEliminater identity_;
   OptUpdateZeroTensor opt_update_zero_tensor_;
   ConstantDuplicateMul constant_duplicate_mul_;
+  PowerOneEliminate power_one_;
   std::vector<TransformFuncType> eliminaters_{};
 };
 }  // namespace irpass

mindspore/ccsrc/optimizer/irpass/incorporate_getitem.h

@@ -21,6 +21,7 @@
 #include <algorithm>
 #include <unordered_map>
 #include <memory>
+#include <unordered_set>
 
 #include "optimizer/irpass.h"
 #include "optimizer/optimizer.h"
@@ -28,7 +29,6 @@
 #include "ir/func_graph.h"
 #include "ir/func_graph_cloner.h"
 #include "operator/ops.h"
-
 namespace mindspore {
 namespace opt {
 namespace irpass {
@@ -81,13 +81,32 @@ class IncorporateGetitem : public AnfVisitor {
   AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
     Reset();
     AnfVisitor::Match(prim::kPrimTupleGetItem, {IsCNode, IsValueNode<Int32Imm>})(node);
-
-    if (node->func_graph() != nullptr && idx_ >= 0 && fg_ != nullptr) {
-      auto new_fg = getitem_transform_(fg_, idx_);
-      (void)args_.insert(args_.begin(), NewValueNode(new_fg));
-      return node->func_graph()->NewCNode(args_);
+    if (node->func_graph() == nullptr || idx_ == -1 || fg_ == nullptr) {
+      return nullptr;
     }
-    return nullptr;
+
+    if (fg_->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+      // If graph kernel has muti output, do not split.
+      // some graph kernel output has EnvInstance node or DeadCode node should split.
+      auto output = fg_->output();
+      if (IsPrimitiveCNode(output, prim::kPrimMakeTuple)) {
+        auto output_cnode = output->cast<CNodePtr>();
+        auto outputs = output_cnode->inputs();
+        int real_output_cnt = 0;
+        for (size_t i = 1; i < outputs.size(); ++i) {
+          if (IsCNode(outputs[i]) || IsValueNode<tensor::Tensor>(outputs[i]) || IsParam(outputs[i])) {
+            real_output_cnt++;
+            if (real_output_cnt > 1) {
+              return nullptr;
+            }
+          }
+        }
+      }
+    }
+
+    auto new_fg = getitem_transform_(fg_, idx_);
+    (void)args_.insert(args_.begin(), NewValueNode(new_fg));
+    return node->func_graph()->NewCNode(args_);
   }
 
   void Visit(const CNodePtr &cnode) override {
@@ -115,6 +134,172 @@ class IncorporateGetitem : public AnfVisitor {
   internal::GetitemTransform getitem_transform_;
 };
 
+class IncorporateGetitemFromParam : public AnfVisitor {
+ public:
+  void Process(const FuncGraphPtr &func_graph, const CNodePtr &cnode, const AnfNodePtr &param, size_t input_idx) {
+    auto mng = func_graph->manager();
+    MS_EXCEPTION_IF_NULL(mng);
+    auto &node_users = mng->node_users();
+    if (node_users.find(param) == node_users.end() || node_users[param].empty()) {
+      args_.push_back(cnode->input(input_idx + 1));
+      return;
+    }
+
+    for (auto &user : node_users[param]) {
+      if (!IsPrimitiveCNode(user.first, prim::kPrimTupleGetItem)) {
+        // we do not process this case.
+        args_.push_back(cnode->input(input_idx + 1));
+        return;
+      }
+    }
+
+    // update new args.
+    if (IsPrimitiveCNode(cnode->input(input_idx + 1), prim::kPrimMakeTuple)) {
+      // case 1
+      replace_parameters_[input_idx] = true;
+      need_update_ = true;
+      auto make_tuple_cnode = cnode->input(input_idx + 1)->cast<CNodePtr>();
+      auto &make_tuple_cnode_inputs = make_tuple_cnode->inputs();
+      inputs_num_[input_idx] = make_tuple_cnode_inputs.size() - 1;
+      args_.insert(args_.end(), make_tuple_cnode_inputs.begin() + 1, make_tuple_cnode_inputs.end());
+    } else {
+      // case 2
+      auto prev_cnode = cnode->input(input_idx + 1)->cast<CNodePtr>();
+      auto prev_fg = GetValueNode<FuncGraphPtr>(prev_cnode->input(0));
+      auto fg_output = prev_fg->output();
+      if (!IsPrimitiveCNode(fg_output, prim::kPrimMakeTuple)) {
+        MS_LOG(ERROR) << "The return of: " << prev_fg->ToString()
+                      << " should be a make tuple, but got: " << fg_output->DebugString();
+        return;
+      }
+      replace_parameters_[input_idx] = true;
+      need_update_ = true;
+      auto make_tuple_cnode = fg_output->cast<CNodePtr>();
+      inputs_num_[input_idx] = make_tuple_cnode->inputs().size() - 1;
+      for (size_t output_i = 0; output_i < inputs_num_[input_idx]; ++output_i) {
+        auto new_getitem =
+          func_graph->NewCNode({NewValueNode(prim::kPrimTupleGetItem), prev_cnode, NewValueNode(SizeToInt(output_i))});
+        auto aptr = std::make_shared<abstract::AbstractScalar>(std::make_shared<Int32Imm>(SizeToInt(output_i)));
+        new_getitem->input(2)->set_abstract(aptr);
+        new_getitem->set_abstract(make_tuple_cnode->input(output_i + 1)->abstract());
+        args_.push_back(new_getitem);
+      }
+    }
+  }
+
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    if (node->func_graph() == nullptr) {
+      return nullptr;
+    }
+
+    Reset();
+
+    auto cnode = node->cast<CNodePtr>();
+    if (cnode == nullptr) {
+      return nullptr;
+    }
+    auto &inputs = cnode->inputs();
+    auto fg = GetValueNode<FuncGraphPtr>(inputs[0]);
+    if (fg == nullptr) {
+      return nullptr;
+    }
+    auto mng = fg->manager();
+    MS_EXCEPTION_IF_NULL(mng);
+    auto parameters = fg->parameters();
+    if (parameters.size() != inputs.size() - 1) {
+      return nullptr;
+    }
+    replace_parameters_ = std::vector<bool>(parameters.size(), false);
+    inputs_num_ = std::vector<size_t>(parameters.size(), 1);
+    auto node_fg = node->func_graph();
+
+    for (size_t i = 1; i < inputs.size(); ++i) {
+      if (IsPrimitiveCNode(inputs[i], prim::kPrimMakeTuple) || IsCNodeGraphKernel(inputs[i])) {
+        Process(node_fg, cnode, parameters[i - 1], i - 1);
+      } else {
+        args_.push_back(inputs[i]);
+      }
+    }
+
+    if (!need_update_) {
+      return nullptr;
+    }
+
+    FuncGraphPtr new_fg = TransformableClone(fg, std::make_shared<TraceTransform>("sp"));
+    mng->AddFuncGraph(new_fg);
+
+    auto node_users = mng->node_users();
+    std::vector<AnfNodePtr> new_fg_parameters = new_fg->parameters();
+    std::vector<AnfNodePtr> new_parameters;
+    size_t curr_input_idx{0};
+    for (size_t param_i = 0; param_i < new_fg_parameters.size(); ++param_i) {
+      if (!replace_parameters_[param_i]) {
+        if (parameters[param_i]->abstract() != nullptr) {
+          new_fg_parameters[param_i]->set_abstract(parameters[param_i]->abstract());
+        }
+        new_parameters.push_back(new_fg_parameters[param_i]);
+        curr_input_idx++;
+        continue;
+      }
+
+      // make a new parameter.
+      for (size_t input_i = 0; input_i < inputs_num_[param_i]; ++input_i) {
+        auto new_param = std::make_shared<Parameter>(new_fg);
+        new_param->set_abstract(args_.at(curr_input_idx)->abstract());
+
+        // update users of new parameter.
+        for (auto &user : node_users[new_fg_parameters[param_i]]) {
+          idx_ = -1;
+          AnfVisitor::Match(prim::kPrimTupleGetItem, {IsParam, IsValueNode<Int32Imm>})(user.first);
+          if (idx_ == -1) {
+            MS_LOG(ERROR) << "User of: " << new_fg_parameters[param_i]->DebugString()
+                          << " must be tuple getitem here, but got: " << user.first->DebugString();
+            return nullptr;
+          }
+
+          if (input_i == IntToSize(idx_)) {
+            for (auto &sub_user : node_users[user.first]) {
+              auto sub_user_cnode = sub_user.first->cast<CNodePtr>();
+              MS_EXCEPTION_IF_NULL(sub_user_cnode);
+              sub_user_cnode->set_input(sub_user.second, new_param);
+              (void)mng->Replace(sub_user.first, sub_user_cnode);
+            }
+          }
+        }
+
+        // (void)mng->Replace(new_fg_parameters[param_i], new_param);
+        new_parameters.push_back(new_param);
+        curr_input_idx++;
+      }
+    }
+
+    mng->SetParameters(new_fg, new_parameters);
+    (void)args_.insert(args_.begin(), NewValueNode(new_fg));
+    auto new_call = node_fg->NewCNode(args_);
+    new_call->set_abstract(node->abstract());
+    return new_call;
+  }
+
+  void Visit(const ValueNodePtr &vnode) override { idx_ = GetValue<int>(vnode->value()); }
+
+  void Visit(const CNodePtr &cnode) override {}
+
+  void Reset() {
+    replace_parameters_.clear();
+    args_.clear();
+    inputs_num_.clear();
+    need_update_ = false;
+    idx_ = -1;
+  }
+
+ private:
+  std::vector<bool> replace_parameters_{};
+  std::vector<AnfNodePtr> args_{};
+  std::vector<size_t> inputs_num_{};
+  bool need_update_{false};
+  int idx_{-1};
+};
+
 // {prim::kPrimTupleGetItem, {{prim::kPrimSwitch, X, G1, G2}, Xs}, C}
 class IncorporateGetitemSwitch : public AnfVisitor {
  public:

mindspore/ccsrc/optimizer/irpass/inline.h

@@ -86,20 +86,10 @@ bool IsUniqueUse(const FuncGraphPtr &fg, AnfNodePtr) {
 
 bool IsInside(FuncGraphPtr, const AnfNodePtr &node) {
   MS_EXCEPTION_IF_NULL(node->func_graph());
-  auto &flags = node->func_graph()->flags();
-  if (flags.find("inline_inside") != flags.end()) {
-    return flags["inline_inside"];
-  }
-  return false;
+  return node->func_graph()->has_flag("inline_inside");
 }
 
-bool IsCore(const FuncGraphPtr &fg, AnfNodePtr) {
-  auto &flags = fg->flags();
-  if (flags.find("core") != flags.end()) {
-    return flags["core"];
-  }
-  return false;
-}
+bool IsCore(const FuncGraphPtr &fg, AnfNodePtr) { return fg->has_flag("core"); }
 
 bool NoCriterion(FuncGraphPtr, AnfNodePtr) { return true; }
 
@@ -123,6 +113,13 @@ class InlinerBase : public AnfVisitor {
     if (fg->has_flag(FUNC_GRAPH_FLAG_DEFER_INLINE)) {
       return nullptr;
     }
+    // Do not inline GraphKernel to Cell.
+    if (fg->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL) && !node->func_graph()->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+      // If the GraphKernel only contains a return node, we make it inlined.
+      if (fg->nodes().size() - fg->parameters().size() > 1) {
+        return nullptr;
+      }
+    }
 
     Reset();
     bool is_match = false;

mindspore/ccsrc/optimizer/irpass/mark_interface_fusion.h

@@ -0,0 +1,86 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_OPTIMIZER_IRPASS_MARK_INTERFACE_FUSION_H
+#define MINDSPORE_CCSRC_OPTIMIZER_IRPASS_MARK_INTERFACE_FUSION_H
+
+#include <string>
+#include <sstream>
+#include <unordered_map>
+
+#include "session/anf_runtime_algorithm.h"
+#include "optimizer/optimizer.h"
+#include "optimizer/irpass.h"
+#include "ir/visitor.h"
+#include "operator/ops.h"
+#include "utils/graph_utils.h"
+#include "operator/composite/composite.h"
+
+namespace mindspore {
+namespace opt {
+namespace irpass {
+
+static int count = 0;
+
+std::string GetFusionNumber() {
+  std::stringstream ss;
+  ss << std::setw(4) << std::setfill('0') << count;
+  std::string num = ss.str();
+  ++count;
+
+  return "_" + num;
+}
+
+// Mark CNodes which can be merged in kernel build
+class MarkInterfaceFusion : public AnfVisitor {
+ public:
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    if (node->func_graph()->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL) && IsPrimitiveCNode(node, prim::kPrimSelect)) {
+      auto cnode = node->cast<CNodePtr>();
+      auto condition = cnode->input(1);
+      std::string cmp;
+      std::unordered_map<std::string, std::string> cmp_list = {{"GreaterEqual", "GE"}, {"Greater", "GT"},
+                                                               {"LessEqual", "LE"},    {"Less", "LT"},
+                                                               {"Equal", "EQ"},        {"NotEqual", "NE"}};
+      if (IsPrimitiveCNode(condition)) {
+        auto prim_name = GetCNodeFuncName(condition->cast<CNodePtr>());
+        if (cmp_list.count(prim_name) != 0) {
+          // Mark Select and compare node
+          cmp = cmp_list[prim_name];
+          auto cnt = GetFusionNumber();
+          AnfAlgo::SetNodeAttr("fusion", MakeValue("Select" + cmp + cnt), condition);
+          AnfAlgo::SetNodeAttr("fusion", MakeValue("Select" + cmp + cnt + "_end"), node);
+          for (size_t i = 1; i < cnode->inputs().size(); ++i) {
+            if (IsPrimitiveCNode(cnode->input(i), prim::kPrimZerosLike)) {
+              AnfAlgo::SetNodeAttr("fusion", MakeValue("Select" + cmp + cnt), cnode->input(i));
+            }
+          }
+        }
+      }
+    }
+    return nullptr;
+  }
+
+  void Visit(const AnfNodePtr &) override {}
+
+ private:
+  AnfNodePtr y_{nullptr};
+};
+
+}  // namespace irpass
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_OPTIMIZER_IRPASS_MARK_INTERFACE_FUSION_H

mindspore/ccsrc/optimizer/irpass/merge_addn.h

@@ -19,6 +19,7 @@
 
 #include <vector>
 #include <algorithm>
+#include <memory>
 
 #include "optimizer/irpass.h"
 #include "optimizer/optimizer.h"
@@ -196,6 +197,131 @@ class AddNZeroFilter : public AnfVisitor {
   std::vector<AnfNodePtr> filtered_Xs_{}, Xs_{};
   bool has_zero_like_{false};
 };
+
+// {PrimAddN, {kPrimMakeTuple, Xs}}
+// Akg don't support AddN(ValueNode, Tensor, ...), converted to TensorAdd.
+// case0: AddN(inputs)(inputs size < 2) -> error
+// case1: AddN(inputs)(all inputs is ValueNode) -> error
+// case2: AddN(inputs)(inputs size = 2) -> TensorAdd(Tensor, Tensor)
+// case3: AddN(ValueNode, Tensor, Tensor, ...)(has one ValueNode input)
+//   -> TensorAdd(ValueNode, AddN(Tensor, Tensor, ...))
+class AddNEliminater : public AnfVisitor {
+ public:
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    if (!node->isa<CNode>() || node->func_graph() == nullptr) {
+      return nullptr;
+    }
+
+    auto &inputs = node->cast<CNodePtr>()->inputs();
+    auto fg = GetValueNode<FuncGraphPtr>(inputs[0]);
+    MS_EXCEPTION_IF_NULL(fg);
+    auto mng = fg->manager();
+    MS_EXCEPTION_IF_NULL(mng);
+    if (fg->recursive()) {
+      return nullptr;
+    }
+
+    auto new_fg = TransformableClone(fg, std::make_shared<TraceTransform>("fg"));
+    mng->AddFuncGraph(new_fg);
+    need_update_ = false;
+    bool changed = false;
+    do {
+      changed = false;
+      changed |= Process(new_fg);
+    } while (changed);
+
+    if (!need_update_) {
+      return nullptr;
+    } else {
+      auto new_sx = inputs;
+      new_sx[0] = NewValueNode(new_fg);
+      return node->func_graph()->NewCNode(new_sx);
+    }
+  }
+
+  bool Process(const FuncGraphPtr &func_graph) {
+    auto mng = func_graph->manager();
+    MS_EXCEPTION_IF_NULL(mng);
+    auto nodes = TopoSort(func_graph->output());
+    bool changed = false;
+
+    for (size_t i = 0; i < nodes.size(); ++i) {
+      auto node = nodes[i];
+      if (!IsPrimitiveCNode(node, prim::kPrimAddN)) {
+        continue;
+      }
+
+      auto cnode = node->cast<CNodePtr>();
+      MS_EXCEPTION_IF_NULL(cnode);
+      auto &tuple_input = cnode->input(1);
+      MS_EXCEPTION_IF_NULL(tuple_input);
+      auto tuple_input_cnode = tuple_input->cast<CNodePtr>();
+      MS_EXCEPTION_IF_NULL(tuple_input_cnode);
+      auto &tuple_inputs = tuple_input_cnode->inputs();
+      if (tuple_inputs.size() < 3) {
+        // case0: inputs size < 2, error
+        MS_EXCEPTION(ArgumentError) << "Inputs size of AddN less than 2. " << cnode->DebugString(2);
+      }
+
+      int valuenode_num =
+        std::accumulate(tuple_inputs.begin() + 1, tuple_inputs.end(), 0, [](int accumulator, const AnfNodePtr &node) {
+          if (IsValueNode<tensor::Tensor>(node)) {
+            return accumulator + 1;
+          } else {
+            return accumulator;
+          }
+        });
+      if (IntToSize(valuenode_num) == tuple_inputs.size()) {
+        // case1: all inputs is ValueNode, error
+        MS_EXCEPTION(ArgumentError) << "All inputs of AddN is ValueNode. " << cnode->DebugString(2);
+      }
+
+      if (tuple_inputs.size() == 3) {
+        // case2: inputs size = 2, -> TensorAdd(Tensor, Tensor)
+        MS_LOG(DEBUG) << "Replace AddN with two inputs with TensorAdd. " << cnode->DebugString(2);
+        ValuePtr prim_tensoradd = prim::GetPythonOps("TensorAdd", "mindspore.ops.operations");
+        std::vector<AnfNodePtr> new_xs{func_graph->NewCNode({NewValueNode(prim_tensoradd)}), tuple_inputs[1],
+                                       tuple_inputs[2]};
+        mng->Replace(node, func_graph->NewCNode(new_xs));
+        changed = true;
+        continue;
+      }
+
+      auto first_valuenode = std::find_if(tuple_inputs.begin() + 1, tuple_inputs.end(),
+                                          [](const AnfNodePtr &node) { return IsValueNode<tensor::Tensor>(node); });
+      if (first_valuenode == tuple_inputs.end()) {
+        // no ValueNode input found.
+        continue;
+      } else {
+        // case3: has one ValueNode input -> TensorAdd(ValueNode, AddN(Tensor, Tensor, ...))
+        std::vector<AnfNodePtr> make_tuple_new_xs{
+          NewValueNode(prim::kPrimMakeTuple),
+        };
+        std::for_each(tuple_inputs.begin() + 1, tuple_inputs.end(),
+                      [&make_tuple_new_xs, &first_valuenode](const AnfNodePtr &node) {
+                        if (node != *first_valuenode) {
+                          make_tuple_new_xs.push_back(node);
+                        }
+                      });
+        ValuePtr prim_addn = prim::GetPythonOps("AddN", "mindspore.ops.operations");
+        auto new_addn = func_graph->NewCNode(
+          {func_graph->NewCNode({NewValueNode(prim_addn)}), func_graph->NewCNode(make_tuple_new_xs)});
+        ValuePtr prim_tensoradd = prim::GetPythonOps("TensorAdd", "mindspore.ops.operations");
+        auto new_add =
+          func_graph->NewCNode({func_graph->NewCNode({NewValueNode(prim_tensoradd)}), *first_valuenode, new_addn});
+        (void)mng->Replace(node, new_add);
+        changed = true;
+        continue;
+      }
+    }
+
+    need_update_ |= changed;
+    return changed;
+  }
+
+ private:
+  bool need_update_{false};
+};
 }  // namespace irpass
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/optimizer/irpass/reduce_eliminate.h

@@ -79,7 +79,7 @@ class ReduceOneEliminater : public AnfVisitor {
   }
 
   void Visit(const AnfNodePtr &node) override {
-    if (x_ == nullptr) {
+    if (!IsVNode(node) && x_ == nullptr) {
       if (IsValueNode<tensor::Tensor>(node)) {
         is_tensor_ = true;
       }

mindspore/ccsrc/optimizer/irpass/ref_eliminate.h

@@ -23,6 +23,8 @@
 #include "optimizer/irpass.h"
 #include "ir/visitor.h"
 #include "operator/ops.h"
+#include "utils/graph_utils.h"
+#include "operator/composite/composite.h"
 
 namespace mindspore {
 namespace opt {
@@ -36,6 +38,7 @@ class MakeRefEliminater : public AnfVisitor {
       this->y_ = node;
       return true;
     };
+
     AnfVisitor::Match(prim::kPrimMakeRef, {IsNode, gety, IsNode})(node);
     return y_;
   }

mindspore/ccsrc/optimizer/irpass/special_op_eliminate.h

@@ -142,7 +142,7 @@ class ResetDeferInline : public AnfVisitor {
   AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
     if (IsValueNode<FuncGraph>(node)) {
       auto fg = GetValueNode<FuncGraphPtr>(node);
-      fg->set_flags(FUNC_GRAPH_FLAG_DEFER_INLINE, false);
+      fg->set_flag(FUNC_GRAPH_FLAG_DEFER_INLINE, false);
     }
     return nullptr;
   }

mindspore/ccsrc/optimizer/irpass/specialize_transform.h

@@ -22,6 +22,7 @@
 #include <memory>
 #include <utility>
 #include <unordered_map>
+#include <unordered_set>
 
 #include "optimizer/irpass.h"
 #include "optimizer/optimizer.h"
@@ -41,7 +42,7 @@ class SpecializeTransform {
   ~SpecializeTransform() = default;
 
   FuncGraphPtr operator()(const FuncGraphPtr &func_graph, std::vector<FuncGraphPtr> graph_args,
-                          std::vector<PrimitivePtr> prim_args) {
+                          std::vector<PrimitivePtr> prim_args, std::vector<tensor::TensorPtr> value_args) {
     if (cache_.count(func_graph) == 0) {
       cache_[func_graph] = {};
     }
@@ -69,6 +70,13 @@ class SpecializeTransform {
           (void)mng->Replace(params[i], arg);
           continue;
         }
+        if (value_args[i] != nullptr) {
+          auto const_tensor = *value_args[i];
+          auto const_tensor_ptr = std::make_shared<tensor::Tensor>(const_tensor);
+          AnfNodePtr arg = NewValueNode(const_tensor_ptr);
+          (void)mng->Replace(params[i], arg);
+          continue;
+        }
         new_params.push_back(params[i]);
       }
 
@@ -108,6 +116,7 @@ class SpecializeOnGraphArguments : public AnfVisitor {
 
     std::vector<FuncGraphPtr> graph_args;
     std::vector<PrimitivePtr> prim_args;
+    std::vector<tensor::TensorPtr> value_node_args;
     std::vector<AnfNodePtr> new_xs;
     bool hasVNode = false;
     for (size_t i = 1; i < inputs.size(); i++) {
@@ -115,15 +124,24 @@ class SpecializeOnGraphArguments : public AnfVisitor {
         auto fg_vnode = GetValueNode<FuncGraphPtr>(inputs[i]);
         graph_args.push_back(fg_vnode);
         prim_args.emplace_back(nullptr);
+        value_node_args.emplace_back(nullptr);
         hasVNode = true;
       } else if (IsValueNode<Primitive>(inputs[i])) {
         auto p_vnode = GetValueNode<PrimitivePtr>(inputs[i]);
         graph_args.emplace_back(nullptr);
         prim_args.push_back(p_vnode);
+        value_node_args.emplace_back(nullptr);
+        hasVNode = true;
+      } else if (IsValueNode<tensor::Tensor>(inputs[i])) {
+        tensor::TensorPtr t_vnode = GetValueNode<tensor::TensorPtr>(inputs[i]);
+        graph_args.emplace_back(nullptr);
+        prim_args.emplace_back(nullptr);
+        value_node_args.emplace_back(t_vnode);
         hasVNode = true;
       } else {
         graph_args.emplace_back(nullptr);
         prim_args.emplace_back(nullptr);
+        value_node_args.emplace_back(nullptr);
         new_xs.push_back(inputs[i]);
       }
     }
@@ -132,7 +150,7 @@ class SpecializeOnGraphArguments : public AnfVisitor {
       return nullptr;
     }
 
-    auto new_fg = specialize_transform_(inp0_fg, graph_args, prim_args);
+    auto new_fg = specialize_transform_(inp0_fg, graph_args, prim_args, value_node_args);
     (void)new_xs.insert(new_xs.begin(), NewValueNode(new_fg));
 
     return node->func_graph()->NewCNode(new_xs);
@@ -141,6 +159,146 @@ class SpecializeOnGraphArguments : public AnfVisitor {
  private:
   internal::SpecializeTransform specialize_transform_;
 };
+
+// Eliminate unused parameters.
+// {G, Xs}
+class UnusedParasEliminater : public AnfVisitor {
+ public:
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    if (!node->isa<CNode>() || node->func_graph() == nullptr) {
+      return nullptr;
+    }
+
+    auto cnode = node->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    auto &inputs = cnode->inputs();
+    auto fg = GetValueNode<FuncGraphPtr>(inputs[0]);
+    MS_EXCEPTION_IF_NULL(fg);
+
+    std::vector<AnfNodePtr> parameters = fg->parameters();
+    size_t size = parameters.size();
+    if (size != inputs.size() - 1) {
+      return nullptr;
+    }
+
+    std::vector<AnfNodePtr> new_xs;
+    std::vector<bool> keep_parameters;
+    auto mng = fg->manager();
+    MS_EXCEPTION_IF_NULL(mng);
+    auto &node_users = mng->node_users();
+    bool has_unused_para = false;
+    for (size_t i = 0; i < size; ++i) {
+      auto iter = node_users.find(parameters[i]);
+      if (iter != node_users.end() && !iter->second.empty()) {
+        keep_parameters.push_back(true);
+        new_xs.push_back(inputs[i + 1]);
+        continue;
+      }
+      keep_parameters.push_back(false);
+      has_unused_para = true;
+    }
+
+    if (!has_unused_para) {
+      return nullptr;
+    }
+    FuncGraphPtr new_fg = TransformableClone(fg, std::make_shared<TraceTransform>("sp"));
+    mng->AddFuncGraph(new_fg);
+
+    std::vector<AnfNodePtr> new_fg_parameters = new_fg->parameters();
+    std::vector<AnfNodePtr> new_parameters;
+    for (size_t i = 0; i < size; i++) {
+      if (keep_parameters[i]) {
+        if (parameters[i]->abstract() != nullptr) {
+          new_fg_parameters[i]->set_abstract(parameters[i]->abstract());
+        }
+        new_parameters.push_back(new_fg_parameters[i]);
+      }
+    }
+    mng->SetParameters(new_fg, new_parameters);
+
+    (void)new_xs.insert(new_xs.begin(), NewValueNode(new_fg));
+    return node->func_graph()->NewCNode(new_xs);
+  }
+};
+
+// Eliminate unused outputs.
+// {G, Xs}
+class UnusedOutputEliminater : public AnfVisitor {
+ public:
+  AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
+    if (!node->isa<CNode>() || node->func_graph() == nullptr) {
+      return nullptr;
+    }
+
+    auto &inputs = node->cast<CNodePtr>()->inputs();
+    auto fg = GetValueNode<FuncGraphPtr>(inputs[0]);
+    MS_EXCEPTION_IF_NULL(fg);
+    auto mng = fg->manager();
+    MS_EXCEPTION_IF_NULL(mng);
+    if (fg->recursive()) {
+      return nullptr;
+    }
+
+    auto new_fg = TransformableClone(fg, std::make_shared<TraceTransform>("fg"));
+    mng->AddFuncGraph(new_fg);
+    auto new_fg_output = new_fg->output();
+    if (!IsPrimitiveCNode(new_fg_output, prim::kPrimMakeTuple)) {
+      return nullptr;
+    }
+
+    auto output_cnode = new_fg_output->cast<CNodePtr>();
+    auto &node_users = mng->node_users();
+    if (node_users.count(node) == 0 || node_users[node].empty()) {
+      return nullptr;
+    }
+    std::unordered_set<int> used_output_idx;
+    std::vector<std::pair<AnfNodePtr, int>> all_users;
+    for (auto &node_user : node_users[node]) {
+      if (!IsPrimitiveCNode(node_user.first, prim::kPrimTupleGetItem)) {
+        return nullptr;
+      }
+      auto user_cnode = node_user.first->cast<CNodePtr>();
+      size_t used_idx = GetValue<int>(user_cnode->input(2)->cast<ValueNodePtr>()->value());
+      used_output_idx.insert(used_idx);
+      all_users.push_back(std::make_pair(node_user.first, used_idx));
+    }
+
+    if (used_output_idx.size() >= output_cnode->inputs().size() - 1) {
+      // all output has users.
+      return nullptr;
+    }
+
+    if (used_output_idx.empty()) {
+      // we do not process this case.
+      return nullptr;
+    } else if (used_output_idx.size() == 1) {
+      // after eliminate, only one output left.
+      new_fg->set_output(output_cnode->input(*used_output_idx.begin() + 1));
+      // update users.
+      for (auto &ret_user : all_users) {
+        (void)mng->Replace(ret_user.first, node);
+      }
+    } else {
+      // after eliminate, create new multi output.
+      std::vector<AnfNodePtr> new_output_inputs{output_cnode->input(0)};
+      std::unordered_map<int, int> new_idx_map;
+      for (auto idx : used_output_idx) {
+        new_idx_map[idx] = SizeToInt(new_output_inputs.size() - 1);
+        new_output_inputs.push_back(output_cnode->input(idx + 1));
+      }
+      new_fg->set_output(new_fg->NewCNode(new_output_inputs));
+      // update users.
+      for (auto &ret_user : all_users) {
+        auto ret_user_cnode = ret_user.first->cast<CNodePtr>();
+        ret_user_cnode->set_input(2, NewValueNode(new_idx_map[ret_user.second]));
+      }
+    }
+
+    auto new_sx = inputs;
+    new_sx[0] = NewValueNode(new_fg);
+    return node->func_graph()->NewCNode(new_sx);
+  }
+};
 }  // namespace irpass
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/optimizer/optimizer.h

@@ -89,7 +89,7 @@ using OptPassGroupMap = std::vector<std::pair<std::string, OptPassConfig>>;
 class Optimizer : public std::enable_shared_from_this<Optimizer> {
  public:
   Optimizer(const std::string &name, const pipeline::ResourceBasePtr &resource_ptr)
-      : name_(name), resource_(resource_ptr), run_only_once_(false), is_watch_renormalize_(false) {}
+      : name_(name), resource_(resource_ptr), run_only_once_(false), is_watch_renormalize_(false), is_enable_(true) {}
   virtual ~Optimizer() = default;
 
   void Init(const OptPassGroupMap &passes, bool run_only_once) {
@@ -132,6 +132,9 @@ class Optimizer : public std::enable_shared_from_this<Optimizer> {
   }
 
   FuncGraphPtr step(FuncGraphPtr func_graph, bool use_profile = true) {
+    if (!is_enable_) {
+      return func_graph;
+    }
     // Optimizer step counter;
     int counter = -1;
     bool changes = true;
@@ -171,7 +174,7 @@ class Optimizer : public std::enable_shared_from_this<Optimizer> {
           };
           use_profile ? (WITH(MsProfile::GetProfile()->Step(pass_names_[i])) opt_func) : opt_func();
           if (IS_OUTPUT_ON(mindspore::DEBUG) && MsContext::GetInstance()->save_graphs_flag()) {
-            MS_LOG(DEBUG) << name_ << " round " << counter << " OptPass " << pass_names_[i] << " end.";
+            MS_LOG(DEBUG) << "The opt " << name_ << " round " << counter << " OptPass " << pass_names_[i] << " end.";
             auto fg_name =
               "opt_substep_" + name_ + "_r" + std::to_string(counter) + "_" + std::to_string(i) + "_" + pass_names_[i];
             func_graph->DumpFuncGraph(fg_name);
@@ -211,6 +214,7 @@ class Optimizer : public std::enable_shared_from_this<Optimizer> {
   void enable_watch_renormalize() { is_watch_renormalize_ = true; }
   void disable_watch_renormalize() { is_watch_renormalize_ = false; }
   bool is_watch_renormalize() { return is_watch_renormalize_; }
+  void set_enable(bool enable) { is_enable_ = enable; }
 
  private:
   const std::string name_;
@@ -220,6 +224,7 @@ class Optimizer : public std::enable_shared_from_this<Optimizer> {
   bool run_only_once_;
   std::vector<AnfNodePtr> untyped_nodes_;
   bool is_watch_renormalize_;
+  bool is_enable_;
 };
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/parallel/allreduce_fusion/step_allreduce_fusion.cc

@@ -64,7 +64,7 @@ bool StepAllreduceFusion(const FuncGraphPtr &root, const opt::OptimizerPtr &opti
   DumpGraph(root, std::string(ALLREDUCE_FUSION_END));
 
   // allreduce fusion only run once
-  root->flags()[ALLREDUCE_FUSION_RUN_ONCE_ONLY] = true;
+  root->set_flag(ALLREDUCE_FUSION_RUN_ONCE_ONLY, true);
   res->results()[pipeline::kStepParallelGraph] = root;
 #if defined(_WIN32) || defined(_WIN64)
   auto end_time = std::chrono::steady_clock::now();

mindspore/ccsrc/parallel/context.cc

@@ -158,8 +158,8 @@ void ParallelParameterContextRestoreInNoTraining(const FuncGraphPtr &func_graph,
   MS_EXCEPTION_IF_NULL(func_graph);
   MS_EXCEPTION_IF_NULL(param_node);
   MS_EXCEPTION_IF_NULL(ptr);
-  if (!func_graph->has_flag(AUTO_PARALLEL) || (func_graph->flags().count(TRAINING) == 0) ||
-      func_graph->flags()[TRAINING]) {
+  if (!func_graph->has_flag(AUTO_PARALLEL) || (func_graph->attrs().count(TRAINING) == 0) ||
+      func_graph->has_flag(TRAINING)) {
     return;
   }
 

mindspore/ccsrc/parallel/step_auto_parallel.cc

@@ -107,7 +107,7 @@ bool StepAutoParallel(const FuncGraphPtr &root, const opt::OptimizerPtr &) {
   time += static_cast<uint64_t>(end_time.tv_usec - start_time.tv_usec);
   MS_LOG(INFO) << "Now leaving step auto parallel, used time: " << time << " us";
 
-  root->flags()[AUTO_PARALLEL_RUN_ONCE_ONLY] = true;
+  root->set_flag(AUTO_PARALLEL_RUN_ONCE_ONLY, true);
   return changes;
 }
 

mindspore/ccsrc/parallel/step_parallel.cc

@@ -2270,10 +2270,10 @@ bool StepParallel(const FuncGraphPtr &root, const opt::OptimizerPtr &optimizer)
       (root->has_flag(SEMI_AUTO_PARALLEL_RUN_ONCE_ONLY))) {
     if (!root->has_flag(CHECK_SET_STRATEGY_VALID_ONCE_ONLY)) {
       if (HasStrategy(root)) {
-        MS_LOG(INFO) << "strategies ignored in " << parallel_mode
+        MS_LOG(INFO) << "Strategies ignored in " << parallel_mode
                      << ", set_strategy() only valid in [semi_]auto_parallel.";
       }
-      root->flags()[CHECK_SET_STRATEGY_VALID_ONCE_ONLY] = true;
+      root->set_flag(CHECK_SET_STRATEGY_VALID_ONCE_ONLY, true);
     }
 
     return changes;
@@ -2330,11 +2330,11 @@ bool StepParallel(const FuncGraphPtr &root, const opt::OptimizerPtr &optimizer)
   DumpGraph(root, std::string(STEP_PARALLEL_END));
 
   // step parallel only run once
-  root->flags()[SEMI_AUTO_PARALLEL_RUN_ONCE_ONLY] = true;
+  root->set_flag(SEMI_AUTO_PARALLEL_RUN_ONCE_ONLY, true);
   res->results()[pipeline::kStepParallelGraph] = root;
 
   // in auto parallel mode, no need to check if stategies set
-  root->flags()[CHECK_SET_STRATEGY_VALID_ONCE_ONLY] = true;
+  root->set_flag(CHECK_SET_STRATEGY_VALID_ONCE_ONLY, true);
 
   (void)gettimeofday(&end_time, nullptr);
   uint64_t time = kUSecondInSecond * static_cast<uint64_t>(end_time.tv_sec - start_time.tv_sec);

mindspore/ccsrc/pipeline/init.cc

@@ -151,7 +151,10 @@ PYBIND11_MODULE(_c_expression, m) {
     .def("set_check_bprop_flag", &mindspore::MsContext::set_check_bprop_flag, "Set whether to check bprop.")
     .def("get_max_device_memory", &mindspore::MsContext::max_device_memory, "Get deivce memory max size.")
     .def("set_max_device_memory", &mindspore::MsContext::set_max_device_memory, "Set deivce memory max size.")
-    .def("set_print_file_path", &mindspore::MsContext::set_print_file_path, "Set path to print.");
+    .def("set_print_file_path", &mindspore::MsContext::set_print_file_path, "Set path to print.")
+    .def("set_enable_graph_kernel", &mindspore::MsContext::set_enable_graph_kernel,
+         "Set the GraphKernel switch to on or off.")
+    .def("get_enable_graph_kernel", &mindspore::MsContext::enable_graph_kernel, "Get the value of GraphKernel switch.");
 
   (void)py::class_<mindspore::MpiConfig, std::shared_ptr<mindspore::MpiConfig>>(m, "MpiConfig")
     .def_static("get_instance", &mindspore::MpiConfig::GetInstance, "Get mpi config instance.")

mindspore/ccsrc/pipeline/parse/data_converter.cc

@@ -278,7 +278,7 @@ bool ConvertCellObjToFuncGraph(py::object obj, ValuePtr *const data) {
     if (bprop_graph != nullptr) {
       (void)func_graph->transforms().insert(std::make_pair(CUSTOM_BPROP_NAME, FuncGraphTransform(bprop_graph)));
       (void)bprop_graph->transforms().insert(std::make_pair("primal", FuncGraphTransform(func_graph)));
-      func_graph->set_flags(FUNC_GRAPH_FLAG_DEFER_INLINE, true);
+      func_graph->set_flag(FUNC_GRAPH_FLAG_DEFER_INLINE, true);
     }
   }
   *data = func_graph;

mindspore/ccsrc/pipeline/parse/parse.cc

@@ -1448,15 +1448,23 @@ bool ParseAst::UpdateFuncGraphFlags(const FuncGraphPtr &func_graph) {
   }
   py::dict flags = python_adapter::GetPyObjAttr(obj_, PYTHON_EXTERN_MINDSPORE_FLAG);
   for (auto &item : flags) {
-    if (!py::isinstance<py::str>(item.first) || !py::isinstance<py::bool_>(item.second)) {
+    if (!py::isinstance<py::str>(item.first)) {
       MS_LOG(ERROR) << "Type error in flags dict convert";
       return false;
     }
     auto name = py::cast<std::string>(item.first);
-    auto value = py::cast<bool>(item.second);
-    MS_LOG(DEBUG) << "Flag name: " << name << ". Value: " << value;
-
-    func_graph->set_flags(name, value);
+    if (py::isinstance<py::bool_>(item.second)) {
+      auto value = py::cast<bool>(item.second);
+      MS_LOG(DEBUG) << "Flag name: " << name << ". Value: " << value;
+      func_graph->set_flag(name, value);
+    } else if (py::isinstance<py::str>(item.second)) {
+      auto value = py::cast<std::string>(item.second);
+      MS_LOG(DEBUG) << "Flag name: " << name << ". Value: " << value;
+      func_graph->set_attr(name, MakeValue(value));
+    } else {
+      MS_LOG(ERROR) << "Type error in flags/attrs dict convert";
+      return false;
+    }
   }
 
   return true;

mindspore/ccsrc/pipeline/parse/parse.h

@@ -223,8 +223,8 @@ class Parser {
     FunctionBlockPtr block = std::make_shared<FunctionBlock>(parse);
     // In order to keep effect order in the sub-graphs which generated by control flow.
     // We copy the flags from the top graph to the sub-graphs.
-    if (func_graph_ && !func_graph_->flags().empty()) {
-      block->func_graph()->set_flags(func_graph_->flags());
+    if (func_graph_ && !func_graph_->attrs().empty()) {
+      block->func_graph()->set_attrs(func_graph_->attrs());
     }
     func_block_list_.push_back(block);
     return block;

mindspore/ccsrc/pipeline/pass.cc

@@ -25,12 +25,14 @@
 #include <functional>
 
 #include "ir/func_graph_cloner.h"
+#include "debug/anf_ir_utils.h"
 #include "pipeline/parse/parse_base.h"
 #include "pipeline/parse/data_converter.h"
 #include "pipeline/resource.h"
 #include "pipeline/validator.h"
 #include "optimizer/optimizer.h"
 #include "optimizer/cse.h"
+#include "optimizer/graph_kernel_reuse.h"
 #include "optimizer/clean.h"
 #include "optimizer/irpass.h"
 #include "optimizer/control_depend.h"
@@ -38,6 +40,7 @@
 #include "parallel/step_auto_parallel.h"
 #include "parallel/allreduce_fusion/step_allreduce_fusion.h"
 #include "utils/any.h"
+#include "utils/log_adapter.h"
 
 namespace mindspore {
 namespace pipeline {
@@ -162,6 +165,40 @@ OptPassGroupMap GetOptPassesB(const opt::irpass::OptimizeIRPassLib &irpass) {
   return map;
 }
 
+OptPassGroupMap GetOptPassesGraphKernelA(const opt::irpass::OptimizeIRPassLib &irpass) {
+  opt::OptPassConfig interface_fusion = opt::OptPassConfig({
+    irpass.mark_interface_fusion_,
+  });
+  OptPassGroupMap map({
+    {"graph_kernel_reuse", opt::OptPassConfig(opt::GraphKernelReuse())},
+    {"interface_fusion", interface_fusion},
+    {"renormalize", opt::OptPassConfig::Renormalize()},
+    {"cse", opt::OptPassConfig(opt::CSE(false))},
+  });
+  return map;
+}
+
+OptPassGroupMap GetOptPassesGraphKernelB(const opt::irpass::OptimizeIRPassLib &irpass) {
+  opt::OptPassConfig elim_1 = opt::OptPassConfig({
+    irpass.addn_eliminate_,
+    irpass.incorporate_getitem_from_param_,
+  });
+  opt::OptPassConfig elim_2 = opt::OptPassConfig({
+    irpass.unused_parameter_eliminate_,
+    irpass.unused_output_eliminate_,
+  });
+  OptPassGroupMap map({
+    {"elim_1", elim_1},
+    {"renormalize", opt::OptPassConfig::Renormalize()},
+    {"elim_2", elim_2},
+  });
+  return map;
+}
+
+OptPassGroupMap GetOptPassesC(const opt::irpass::OptimizeIRPassLib &irpass) {
+  return OptPassGroupMap({{"renormalize", opt::OptPassConfig::Renormalize()}});
+}
+
 OptPassGroupMap GetControlPhases(const opt::irpass::OptimizeIRPassLib &irpass) {
   opt::OptPassConfig control_group = opt::OptPassConfig({irpass.convert_switch_replacement_}, true);
   OptPassGroupMap map({
@@ -191,8 +228,19 @@ void InitOpt(const ResourcePtr &res) {
     opt::irpass::OptimizeIRPassLib irpass;
     g_pass_opts["opt_a"] = Optimizer::MakeOptimizer("opt_a", res, GetOptPassesA(irpass));
     g_pass_opts["opt_b"] = Optimizer::MakeOptimizer("opt_b", res, GetOptPassesB(irpass), false, true);
+    g_pass_opts["opt_graph_kernel_a"] =
+      Optimizer::MakeOptimizer("opt_graph_kernel_a", res, GetOptPassesGraphKernelA(irpass), true);
+    g_pass_opts["opt_graph_kernel_b"] =
+      Optimizer::MakeOptimizer("opt_graph_kernel_b", res, GetOptPassesGraphKernelB(irpass), false);
+    g_pass_opts["renormal"] = Optimizer::MakeOptimizer("renormal", res, GetOptPassesC(irpass));
     g_pass_opts["opt_control"] = Optimizer::MakeOptimizer("opt_control", res, GetControlPhases(irpass), false, true);
     g_pass_opts["opt_prepare"] = Optimizer::MakeOptimizer("opt_prepare", res, GetPreparePhases(irpass));
+    auto context_ptr = MsContext::GetInstance();
+    MS_EXCEPTION_IF_NULL(context_ptr);
+    if (!(context_ptr->enable_graph_kernel())) {
+      g_pass_opts["opt_graph_kernel_a"]->set_enable(false);
+      g_pass_opts["opt_graph_kernel_b"]->set_enable(false);
+    }
   }
 }
 }  // namespace
@@ -224,9 +272,13 @@ bool OptPassGroup(const ResourcePtr &res, const std::string &name) {
 
 bool OptPassAGroup(const ResourcePtr &res) { return OptPassGroup(res, "opt_a"); }
 bool OptPassBGroup(const ResourcePtr &res) { return OptPassGroup(res, "opt_b"); }
+bool OptPassGraphKernelGroupA(const ResourcePtr &res) { return OptPassGroup(res, "opt_graph_kernel_a"); }
+bool OptPassGraphKernelGroupB(const ResourcePtr &res) { return OptPassGroup(res, "opt_graph_kernel_b"); }
 bool ControlGroup(const ResourcePtr &res) { return OptPassGroup(res, "opt_control"); }
 bool PrepareGroup(const ResourcePtr &res) { return OptPassGroup(res, "opt_prepare"); }
 
+bool OptPassRNGroup(const ResourcePtr &res) { return OptPassGroup(res, "renormal"); }
+
 bool AddControlDependPass(const ResourcePtr &res) {
   FuncGraphPtr func_graph = res->func_graph();
   MS_EXCEPTION_IF_NULL(func_graph);
@@ -270,8 +322,10 @@ bool InferenceOptPreparePass(const ResourcePtr &res) {
 std::vector<PassItem> kVmPasses = {{"simplify_data_structures", SimplifyDataStructuresPass},
                                    {"opt_a", OptPassAGroup},
                                    {"opt_b", OptPassBGroup},
-                                   {"add_control_depend", AddControlDependPass},
-                                   {"cconv", CconvPass}};
+                                   {"cconv", CconvPass},
+                                   {"opt_graph_kernel_a", OptPassGraphKernelGroupA},
+                                   {"opt_graph_kernel_b", OptPassGraphKernelGroupB},
+                                   {"add_control_depend", AddControlDependPass}};
 
 std::vector<PassItem> kGePasses = {{"simplify_data_structures", SimplifyDataStructuresPass},
                                    {"opt_a", OptPassAGroup},

mindspore/ccsrc/pipeline/pipeline_ge.cc

@@ -488,7 +488,7 @@ py::object ExecDFGraph(const std::map<std::string, ExecutorInfoPtr> &info, const
 #ifdef ENABLE_INFER
   // Now don't use the graph because the exec ge function don't take effect
   MS_EXCEPTION_IF_NULL(info.at(phase)->func_graph);
-  if (ENABLE_TRAIN != info.at(phase)->func_graph->flags()["training"]) {
+  if (ENABLE_TRAIN != info.at(phase)->func_graph->has_flag("training")) {
     MS_LOG(ERROR) << "Graph training mode mismatch mode of libraries";
     ConfigManager::GetInstance().ResetConfig();
     return py::none();

mindspore/ccsrc/pipeline/static_analysis/evaluator.cc

@@ -165,7 +165,7 @@ AbstractBasePtrList FuncGraphEvaluator::BroadenUndeterminedArgs(const AbstractBa
         MS_LOG(DEBUG) << "Joined args: " << ::mindspore::ToString(joined_args_spec_list);
         // If there is loop variant, all arguments need to be broaden to avoid wrong constant propagation.
         if (!(joined_args_spec_list == args_spec_list)) {
-          func_graph_->set_flags(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
+          func_graph_->set_flag(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
         }
         return joined_args_spec_list;
       }
@@ -178,7 +178,7 @@ AbstractBasePtrList FuncGraphEvaluator::BroadenUndeterminedArgs(const AbstractBa
       // If there is loop variant, all arguments need to be broaden to avoid wrong constant propagation.
       if (!(joined_args_spec_list == args_spec_list)) {
         trace_.push_back(joined_args_spec_list);
-        func_graph_->set_flags(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
+        func_graph_->set_flag(FUNC_GRAPH_FLAG_IGNORE_VALUES, true);
       }
       MS_LOG(DEBUG) << "Joined eval args: " << ::mindspore::ToString(joined_args_spec_list);
       return joined_args_spec_list;

mindspore/ccsrc/pipeline/static_analysis/static_analysis.cc

@@ -479,7 +479,7 @@ void AnalysisEngine::SetUndeterminedFlag(const EvaluatorPtr &evaluator) {
   if (undetermined_fgs) {
     auto fg_parent = fg->parent();
     MS_EXCEPTION_IF_NULL(fg_parent);
-    fg_parent->set_flags(kFuncGraphFlagUndetermined, true);
+    fg_parent->set_flag(kFuncGraphFlagUndetermined, true);
     MS_LOG(DEBUG) << "Set graph undetermined: " << fg_parent->ToString();
   }
 }

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc

@@ -16,6 +16,7 @@
 #include "pre_activate/ascend/ascend_backend_optimization.h"
 #include <memory>
 #include <string>
+#include <set>
 #include "pre_activate/common/optimizer.h"
 #include "pre_activate/ascend/ir_fission/bn_split.h"
 #include "pre_activate/ascend/ir_fission/bn_grad_split.h"
@@ -63,6 +64,9 @@
 #include "pre_activate/ascend/format_type/convert_unsupported_transnode_to_aicpu.h"
 #include "pre_activate/pass/eliminate_redundant_op.h"
 #include "pre_activate/pass/common_subexpression_elimination.h"
+#include "pre_activate/pass/fuse_graph_kernel.h"
+#include "pre_activate/pass/fuse_basic.h"
+#include "pre_activate/pass/add_atomic_clean.h"
 #include "pre_activate/ascend/format_type/merge_cast_to_op.h"
 #include "pre_activate/ascend/format_type/check_consistency.h"
 #include "pre_activate/ascend/buffer_fusion/ub_pattern_fusion.h"
@@ -88,6 +92,8 @@
 #include "pre_activate/ascend/enhancer/insert_memcpy_async_for_getnext.h"
 #include "pre_activate/ascend/ir_fission/batch_norm_grad_infer_fission.h"
 #include "pre_activate/ascend/ir_fission/split_fission.h"
+#include "pre_activate/ascend/format_type/modify_ops_attrs.h"
+#include "pre_activate/ascend/format_type/remove_no_use_reshape_op.h"
 #include "utils/context/ms_context.h"
 #include "utils/config_manager.h"
 #include "debug/anf_ir_dump.h"
@@ -164,6 +170,19 @@ void RunOpAscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_g
   kernel_graph->SetExecOrderByDefault();
 }
 
+void AscendGraphKernelCommonProcess(const std::shared_ptr<session::KernelGraph> &kernel_graph) {
+  MS_EXCEPTION_IF_NULL(kernel_graph);
+  auto optimizer = std::make_shared<GraphOptimizer>();
+  MS_EXCEPTION_IF_NULL(optimizer);
+  auto common_process = std::make_shared<PassManager>("graph_kernel_common_process");
+  MS_EXCEPTION_IF_NULL(common_process);
+  common_process->AddPass(std::make_shared<ModifyOpAttrs>());
+  common_process->AddPass(std::make_shared<RemoveNoUseReshapeOp>());
+  optimizer->AddPassManager(common_process);
+  (void)optimizer->Optimize(kernel_graph);
+  kernel_graph->SetExecOrderByDefault();
+}
+
 void AscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_graph) {
   MS_EXCEPTION_IF_NULL(kernel_graph);
   auto optimizer = std::make_shared<GraphOptimizer>();
@@ -332,7 +351,94 @@ void AscendBackendOptimization(const std::shared_ptr<session::KernelGraph> &kern
     std::string file_path =
       save_graphs_path + "/" + "hwopt_d_end" + "_graph_" + std::to_string(kernel_graph->graph_id()) + ".ir";
     DumpIR(file_path, kernel_graph, true);
-    DumpIRProto(kernel_graph, "after_hwopt_" + std::to_string(kernel_graph->graph_id()));
+    DumpIRProto(kernel_graph, "after_hwopt");
+    kernel_graph->DumpFuncGraph("hwopt_d_end");
+  }
+}
+
+void AscendBackendGraphKernelOpt(const std::shared_ptr<session::KernelGraph> &kernel_graph,
+                                 bool is_before_kernel_select) {
+  auto context_ptr = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(context_ptr);
+  if (!(context_ptr->enable_graph_kernel())) {
+    return;
+  }
+  bool save_graphs = context_ptr->save_graphs_flag();
+  auto save_graphs_path = context_ptr->save_graphs_path();
+  if (save_graphs_path.empty()) {
+    save_graphs_path = ".";
+  }
+  if (save_graphs) {
+    std::string file_path = save_graphs_path + "/" + "hwopt_d_graph_kernel_opt_before_graph_" +
+                            std::to_string(!is_before_kernel_select) + "_" + std::to_string(kernel_graph->graph_id()) +
+                            ".ir";
+    DumpIR(file_path, kernel_graph);
+  }
+
+  // Fuse graph kernels with basic ops
+  FuseGraphKernel(kernel_graph, is_before_kernel_select);
+
+  if (save_graphs) {
+    std::string file_path = save_graphs_path + "/" + "hwopt_d_graph_kernel_opt_end_graph_" +
+                            std::to_string(!is_before_kernel_select) + "_" + std::to_string(kernel_graph->graph_id()) +
+                            ".ir";
+    DumpIR(file_path, kernel_graph, true);
+  }
+}
+
+void AscendBackendFuseBasicOpt(const std::shared_ptr<session::KernelGraph> &kernel_graph,
+                               bool is_before_kernel_select) {
+  auto context_ptr = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(context_ptr);
+  if (!(context_ptr->enable_graph_kernel())) {
+    return;
+  }
+  bool save_graphs = context_ptr->save_graphs_flag();
+  auto save_graphs_path = context_ptr->save_graphs_path();
+  if (save_graphs_path.empty()) {
+    save_graphs_path = ".";
+  }
+  if (save_graphs) {
+    std::string file_path = save_graphs_path + "/" + "hwopt_d_fuse_basic_opt_before_graph_" +
+                            std::to_string(!is_before_kernel_select) + "_" + std::to_string(kernel_graph->graph_id()) +
+                            ".ir";
+    DumpIR(file_path, kernel_graph, true);
+  }
+
+  // Fuse basic ops with basic ops
+  FuseBasic(kernel_graph, is_before_kernel_select);
+
+  if (save_graphs) {
+    std::string file_path = save_graphs_path + "/" + "hwopt_d_fuse_basic_opt_end_graph_" +
+                            std::to_string(!is_before_kernel_select) + "_" + std::to_string(kernel_graph->graph_id()) +
+                            ".ir";
+    DumpIR(file_path, kernel_graph, true);
+  }
+}
+
+void AscendBackendAddAtomicClean(const std::shared_ptr<session::KernelGraph> &kernel_graph) {
+  auto context_ptr = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(context_ptr);
+  if (!(context_ptr->enable_graph_kernel())) {
+    return;
+  }
+  bool save_graphs = context_ptr->save_graphs_flag();
+  auto save_graphs_path = context_ptr->save_graphs_path();
+  if (save_graphs_path.empty()) {
+    save_graphs_path = ".";
+  }
+  if (save_graphs) {
+    std::string file_path = save_graphs_path + "/" + "hwopt_d_add_atomic_clean_before" + "_graph_" +
+                            std::to_string(kernel_graph->graph_id()) + ".ir";
+    DumpIR(file_path, kernel_graph);
+  }
+
+  AddAtomicClean(kernel_graph);
+
+  if (save_graphs) {
+    std::string file_path =
+      save_graphs_path + "/" + "hwopt_d_end" + "_graph_" + std::to_string(kernel_graph->graph_id()) + ".ir";
+    DumpIR(file_path, kernel_graph, true);
   }
 }
 

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.h

@@ -24,6 +24,12 @@ void RunOpAscendBackendIRFusionOptimization(const std::shared_ptr<session::Kerne
 void AscendDataLayout(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 void AscendMixPrecision(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 void AscendBackendOptimization(const std::shared_ptr<session::KernelGraph> &kernel_graph);
+void AscendGraphKernelCommonProcess(const std::shared_ptr<session::KernelGraph> &kernel_graph);
+void AscendBackendGraphKernelOpt(const std::shared_ptr<session::KernelGraph> &kernel_graph,
+                                 bool is_before_kernel_select = false);
+void AscendBackendFuseBasicOpt(const std::shared_ptr<session::KernelGraph> &kernel_graph,
+                               bool is_before_kernel_select = false);
+void AscendBackendAddAtomicClean(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 void AscendBackendIRFusionOptimization(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 void AscendBackendUBFusionOptimization(const std::shared_ptr<session::KernelGraph> &kernel_graph);
 }  // namespace opt

mindspore/ccsrc/pre_activate/ascend/ascend_helper.cc

@@ -22,6 +22,7 @@
 #include "utils/utils.h"
 #include "device/kernel_info.h"
 #include "kernel/oplib/oplib.h"
+#include "kernel/common_utils.h"
 #include "operator/ops.h"
 #include "session/anf_runtime_algorithm.h"
 #include "session/kernel_graph.h"
@@ -229,7 +230,7 @@ AnfNodePtr AddCastOpNodeToGraph(const FuncGraphPtr &func_graph, const AnfNodePtr
   if (kernel::OpLib::FindOp(prim::kPrimCast->name(), kernel::kTBE) != nullptr) {
     builder.SetKernelType(KernelType::TBE_KERNEL);
   } else {
-    builder.SetKernelType(KernelType::AUTO_DIFF_KERNEL);
+    builder.SetKernelType(KernelType::AKG_KERNEL);
   }
   // if kernel info is null , it remarks this function is running ut
   if (cast->kernel_info() == nullptr) {
@@ -284,22 +285,17 @@ CNodePtr InsertCastForInput(const FuncGraphPtr &func_graph, const CNodePtr &cnod
   MS_EXCEPTION_IF_NULL(cnode);
   std::vector<AnfNodePtr> new_inputs = {AnfAlgo::GetCNodePrimitiveNode(cnode)};
   for (size_t input_index = 0; input_index < AnfAlgo::GetInputTensorNum(cnode); ++input_index) {
-    TypeId origin_type;
+    const auto infer_type = AnfAlgo::GetPrevNodeOutputInferDataType(cnode, input_index);
+    TypeId origin_type(kTypeUnknown);
     auto cur_input = AnfAlgo::GetInputNode(cnode, input_index);
     auto kernel_with_index = AnfAlgo::VisitKernel(cur_input, 0);
-    auto is_weight_boundary = [](const AnfNodePtr &node) -> bool {
-      if (node->isa<ValueNode>()) {
-        return true;
-      }
-      if (node->isa<Parameter>() && AnfAlgo::IsParameterWeight(node->cast<ParameterPtr>())) {
-        return true;
-      }
-      return false;
-    };
     auto real_input_node = kernel_with_index.first;
-    if (is_weight_boundary(real_input_node)) {
+    if (kernel::IsWeightBoundary(real_input_node) || func_graph->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
       // weight
-      origin_type = AnfAlgo::GetPrevNodeOutputDeviceDataType(cnode, input_index);
+      origin_type = AnfAlgo::GetPrevNodeOutputPrecision(cnode, input_index);
+      if (origin_type == kTypeUnknown) {
+        origin_type = AnfAlgo::GetPrevNodeOutputDeviceDataType(cnode, input_index);
+      }
     } else {
       // feature map
       origin_type = AnfAlgo::GetPrevNodeOutputInferDataType(cnode, input_index);
@@ -307,9 +303,13 @@ CNodePtr InsertCastForInput(const FuncGraphPtr &func_graph, const CNodePtr &cnod
     const std::string dev_fmt = AnfAlgo::GetInputFormat(cnode, input_index);
     const std::vector<size_t> origin_shape = AnfAlgo::GetPrevNodeOutputInferShape(cnode, input_index);
     const TypeId device_type = AnfAlgo::GetInputDeviceDataType(cnode, input_index);
-    if (origin_type != device_type) {
+    // In graph kernel, we check parameter,
+    // the eliminate pass will not eliminate this case, so we just do not insert the noused cast.
+    if (func_graph->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL) && IsValueNode<tensor::Tensor>(cur_input)) {
+      new_inputs.push_back(cur_input);
+    } else if (origin_type != device_type) {
       auto cast =
-        AddCastOpNodeToGraph(func_graph, cur_input, dev_fmt, origin_type, device_type, origin_shape, origin_type);
+        AddCastOpNodeToGraph(func_graph, cur_input, dev_fmt, origin_type, device_type, origin_shape, infer_type);
       MS_EXCEPTION_IF_NULL(cast);
       cast->set_scope(cnode->scope());
       AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), cast);

mindspore/ccsrc/pre_activate/ascend/format_type/check_consistency.cc

@@ -17,9 +17,12 @@
 
 #include <string>
 #include <memory>
+#include <vector>
 
 #include "utils/utils.h"
 #include "session/anf_runtime_algorithm.h"
+#include "common/utils.h"
+#include "kernel/common_utils.h"
 
 namespace mindspore {
 namespace opt {
@@ -74,11 +77,21 @@ const AnfNodePtr CheckConsistency::Process(const FuncGraphPtr &, const AnfNodePt
   if (node == nullptr || !node->isa<CNode>() || !AnfAlgo::IsRealKernel(node)) {
     return nullptr;
   }
-  CNodePtr cnode = node->cast<CNodePtr>();
-  for (size_t i = 0; i < AnfAlgo::GetInputTensorNum(cnode); i++) {
-    if (!CheckFormatForConsistency(cnode, i) || !CheckDataTypeForConsistency(cnode, i)) {
-      MS_LOG(EXCEPTION) << "Found inconsistent format or data type! Op: " << AnfAlgo::GetCNodeName(node) << "["
-                        << node->DebugString() << "]";
+
+  std::vector<AnfNodePtr> todos = {node};
+  if (AnfAlgo::IsGraphKernel(node)) {
+    auto sub_graph = AnfAlgo::GetCNodeFuncGraphPtr(node);
+    MS_EXCEPTION_IF_NULL(sub_graph);
+    kernel::GetValidKernelNodes(sub_graph, &todos);
+  }
+
+  for (auto &t : todos) {
+    CNodePtr cnode = t->cast<CNodePtr>();
+    for (size_t i = 0; i < AnfAlgo::GetInputTensorNum(cnode); i++) {
+      if (!CheckFormatForConsistency(cnode, i) || !CheckDataTypeForConsistency(cnode, i)) {
+        MS_LOG(EXCEPTION) << "Found inconsistent format or data type! Op: " << AnfAlgo::GetCNodeName(cnode) << "["
+                          << cnode->DebugString() << "]";
+      }
     }
   }
   return nullptr;

mindspore/ccsrc/pre_activate/ascend/format_type/insert_cast.cc

@@ -18,6 +18,7 @@
 #include <memory>
 #include <string>
 #include <vector>
+#include <utility>
 
 #include "device/kernel_info.h"
 #include "pre_activate/ascend/ascend_helper.h"
@@ -27,34 +28,45 @@
 #include "session/anf_runtime_algorithm.h"
 #include "session/kernel_graph.h"
 #include "utils/utils.h"
+#include "kernel/common_utils.h"
 
 namespace mindspore {
 namespace opt {
 namespace {
-AnfNodePtr InsertCastForMultipleOutput(const FuncGraphPtr &func_graph, const CNodePtr &cnode) {
+AnfNodePtr InsertCastForMultipleOutput(const FuncGraphPtr &func_graph, const CNodePtr &cnode,
+                                       const std::vector<bool> &need_insert_cast) {
   MS_EXCEPTION_IF_NULL(func_graph);
   MS_EXCEPTION_IF_NULL(cnode);
   std::vector<AnfNodePtr> make_tuple_inputs;
   AbstractBasePtrList abstract_list;
   make_tuple_inputs.push_back(NewValueNode(prim::kPrimMakeTuple));
   for (size_t output_idx = 0; output_idx < AnfAlgo::GetOutputTensorNum(cnode); ++output_idx) {
-    const std::string dev_fmt = AnfAlgo::GetOutputFormat(cnode, output_idx);
-    const std::vector<size_t> origin_shape = AnfAlgo::GetOutputInferShape(cnode, output_idx);
-    const TypeId origin_type = AnfAlgo::GetOutputInferDataType(cnode, output_idx);
-    const TypeId device_type = AnfAlgo::GetOutputDeviceDataType(cnode, output_idx);
+    AnfNodePtr replace_node = nullptr;
+    const auto origin_shape = AnfAlgo::GetOutputInferShape(cnode, output_idx);
+    const auto infer_type = AnfAlgo::GetOutputInferDataType(cnode, output_idx);
     auto idx = NewValueNode(SizeToInt(output_idx));
     MS_EXCEPTION_IF_NULL(idx);
     auto imm = std::make_shared<Int32Imm>(output_idx);
     idx->set_abstract(std::make_shared<abstract::AbstractScalar>(imm));
     auto getitem = func_graph->NewCNode({NewValueNode(prim::kPrimTupleGetItem), cnode, idx});
-    AnfAlgo::SetOutputInferTypeAndShape({origin_type}, {origin_shape}, getitem.get());
-    AnfNodePtr replace_node = nullptr;
-    if (origin_type != device_type) {
-      replace_node =
-        AddCastOpNodeToGraph(func_graph, getitem, dev_fmt, device_type, origin_type, origin_shape, origin_type);
-      MS_EXCEPTION_IF_NULL(replace_node);
-      replace_node->set_scope(cnode->scope());
-      AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), replace_node);
+    AnfAlgo::SetOutputInferTypeAndShape({infer_type}, {origin_shape}, getitem.get());
+    if (need_insert_cast[output_idx]) {
+      const auto dev_fmt = AnfAlgo::GetOutputFormat(cnode, output_idx);
+      TypeId origin_type(kTypeUnknown);
+      if (func_graph->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+        origin_type = AnfAlgo::GetCNodeOutputPrecision(cnode);
+      }
+      origin_type = origin_type == kTypeUnknown ? infer_type : origin_type;
+      const auto device_type = AnfAlgo::GetOutputDeviceDataType(cnode, output_idx);
+      if (origin_type != device_type) {
+        replace_node =
+          AddCastOpNodeToGraph(func_graph, getitem, dev_fmt, device_type, origin_type, origin_shape, infer_type);
+        MS_EXCEPTION_IF_NULL(replace_node);
+        replace_node->set_scope(cnode->scope());
+        AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), replace_node);
+      } else {
+        replace_node = getitem;
+      }
     } else {
       replace_node = getitem;
     }
@@ -65,9 +77,10 @@ AnfNodePtr InsertCastForMultipleOutput(const FuncGraphPtr &func_graph, const CNo
   MS_EXCEPTION_IF_NULL(make_tuple);
   make_tuple->set_abstract(std::make_shared<abstract::AbstractTuple>(abstract_list));
   return make_tuple;
-}
+}  // namespace
 
-AnfNodePtr InsertCastForOutput(const FuncGraphPtr &func_graph, const CNodePtr &cnode) {
+AnfNodePtr InsertCastForOutput(const FuncGraphPtr &func_graph, const CNodePtr &cnode,
+                               const std::vector<bool> &need_insert_cast) {
   MS_EXCEPTION_IF_NULL(func_graph);
   MS_EXCEPTION_IF_NULL(cnode);
   if (AnfAlgo::GetOutputTensorNum(cnode) == 0) {
@@ -76,14 +89,23 @@ AnfNodePtr InsertCastForOutput(const FuncGraphPtr &func_graph, const CNodePtr &c
   MS_EXCEPTION_IF_NULL(cnode->Type());
   // Single output
   if (!cnode->Type()->isa<Tuple>()) {
+    if (!need_insert_cast[0]) {
+      return cnode;
+    }
+
     const std::string dev_fmt = AnfAlgo::GetOutputFormat(cnode, 0);
     std::vector<size_t> origin_shape = AnfAlgo::GetOutputInferShape(cnode, 0);
-    const TypeId origin_type = AnfAlgo::GetOutputInferDataType(cnode, 0);
+    const auto infer_type = AnfAlgo::GetOutputInferDataType(cnode, 0);
+    TypeId origin_type(kTypeUnknown);
+    if (func_graph->has_attr(FUNC_GRAPH_ATTR_GRAPH_KERNEL)) {
+      origin_type = AnfAlgo::GetCNodeOutputPrecision(cnode);
+    }
+    origin_type = origin_type == kTypeUnknown ? infer_type : origin_type;
     const TypeId device_type = AnfAlgo::GetOutputDeviceDataType(cnode, 0);
     AnfNodePtr replace_node = cnode;
     if (origin_type != device_type) {
       replace_node =
-        AddCastOpNodeToGraph(func_graph, cnode, dev_fmt, device_type, origin_type, origin_shape, origin_type);
+        AddCastOpNodeToGraph(func_graph, cnode, dev_fmt, device_type, origin_type, origin_shape, infer_type);
       MS_EXCEPTION_IF_NULL(replace_node);
       replace_node->set_scope(cnode->scope());
       AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), replace_node);
@@ -91,7 +113,57 @@ AnfNodePtr InsertCastForOutput(const FuncGraphPtr &func_graph, const CNodePtr &c
     return replace_node;
   }
   // Multiple output
-  return InsertCastForMultipleOutput(func_graph, cnode);
+  return InsertCastForMultipleOutput(func_graph, cnode, need_insert_cast);
+}
+
+AnfNodePtr ProcessGraphKernelOp(const FuncGraphPtr &func_graph, const AnfNodePtr &node) {
+  // insert cast for ops in graph kernel.
+  auto sub_graph = AnfAlgo::GetCNodeFuncGraphPtr(node);
+  MS_EXCEPTION_IF_NULL(sub_graph);
+  auto mng = sub_graph->manager();
+  MS_EXCEPTION_IF_NULL(mng);
+  std::vector<AnfNodePtr> todo;
+  std::vector<std::pair<AnfNodePtr, size_t>> graph_rets;
+  kernel::GetValidKernelNodes(sub_graph, &todo);
+  kernel::GetGraphRealOutput(sub_graph, &graph_rets);
+  for (auto &t : todo) {
+    AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), t);
+    // process input
+    CNodePtr t_cnode = t->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(t_cnode);
+    auto t_new_node = InsertCastForInput(sub_graph, t_cnode);
+    AnfNodePtr t_new_node_1 = nullptr;
+    std::vector<bool> need_insert_cast(AnfAlgo::GetOutputTensorNum(t), true);
+    // process output
+    auto iter = std::find_if(graph_rets.begin(), graph_rets.end(),
+                             [&t](const std::pair<AnfNodePtr, size_t> &ret) { return ret.first == t; });
+    if (iter != graph_rets.end()) {
+      auto t_fix_output_type = AnfAlgo::GetCNodeOutputPrecision(t);
+      auto t_output_type = AnfAlgo::GetOutputDeviceDataType(t, iter->second);
+      auto graph_output_type = AnfAlgo::GetOutputDeviceDataType(node, iter - graph_rets.begin());
+      if (t_fix_output_type == kTypeUnknown && t_output_type == graph_output_type) {
+        need_insert_cast[iter->second] = false;
+      } else if (t_fix_output_type == t_output_type && t_output_type == graph_output_type) {
+        need_insert_cast[iter->second] = false;
+      }
+      t_new_node_1 = InsertCastForOutput(sub_graph, t_new_node, need_insert_cast);
+    } else {
+      t_new_node_1 = InsertCastForOutput(sub_graph, t_new_node, need_insert_cast);
+    }
+
+    if (t_new_node_1 != nullptr && t_new_node_1 != t) {
+      (void)mng->Replace(t, t_new_node_1);
+    }
+  }
+
+  // insert cast for graph kernel.
+  AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), node);
+  // process input
+  CNodePtr cnode = node->cast<CNodePtr>();
+  MS_EXCEPTION_IF_NULL(cnode);
+  auto new_node = InsertCastForInput(func_graph, cnode);
+  // process output
+  return InsertCastForOutput(func_graph, new_node, std::vector<bool>(AnfAlgo::GetOutputTensorNum(new_node), true));
 }
 }  // namespace
 
@@ -106,13 +178,27 @@ const AnfNodePtr InsertCast::Process(const FuncGraphPtr &func_graph, const AnfNo
   if (!AnfAlgo::IsRealCNodeKernel(node) || func_graph == nullptr) {
     return nullptr;
   }
+
+  if (AnfAlgo::IsGraphKernel(node)) {
+    return ProcessGraphKernelOp(func_graph, node);
+  } else {
+    // insert cast for single op.
+    AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), node);
+    // process input
+    CNodePtr cnode = node->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    auto new_node = InsertCastForInput(func_graph, cnode);
+    // process output
+    return InsertCastForOutput(func_graph, new_node, std::vector<bool>(AnfAlgo::GetOutputTensorNum(new_node), true));
+  }
+  // insert cast for single op.
   AnfAlgo::SetNodeAttr(kAttrVisited, MakeValue(true), node);
   // process input
   CNodePtr cnode = node->cast<CNodePtr>();
   MS_EXCEPTION_IF_NULL(cnode);
   auto new_node = InsertCastForInput(func_graph, cnode);
   // process output
-  return InsertCastForOutput(func_graph, new_node);
+  return InsertCastForOutput(func_graph, new_node, std::vector<bool>(AnfAlgo::GetOutputTensorNum(new_node), true));
 }
 }  // namespace opt
 }  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/format_type/merge_cast_to_op.cc

@@ -133,6 +133,9 @@ AnfNodePtr MergeCastToNextOp(const FuncGraphPtr &graph, const CNodePtr &node, co
     return nullptr;
   }
   auto next_cnode = next_node->cast<CNodePtr>();
+  if (AnfAlgo::IsGraphKernel(next_node)) {
+    return nullptr;
+  }
   auto next_op_name = AnfAlgo::GetCNodeName(next_node);
   std::vector<std::shared_ptr<kernel::KernelBuildInfo>> kernel_info_list;
   kernel_query->Query(next_cnode, &kernel_info_list);
@@ -206,6 +209,9 @@ AnfNodePtr MergeCastToPriorOp(const FuncGraphPtr &graph, const CNodePtr &cur_nod
     return nullptr;
   }
   MS_EXCEPTION_IF_NULL(prior_op);
+  if (AnfAlgo::IsGraphKernel(prior_op)) {
+    return nullptr;
+  }
 
   std::vector<std::shared_ptr<kernel::KernelBuildInfo>> kernel_info_list;
   kernel_query->Query(prior_op, &kernel_info_list);

mindspore/ccsrc/pre_activate/ascend/format_type/modify_ops_attrs.cc

@@ -0,0 +1,99 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "pre_activate/ascend/format_type/modify_ops_attrs.h"
+#include <vector>
+#include <memory>
+#include "utils/utils.h"
+#include "pre_activate/common/helper.h"
+#include "kernel/common_utils.h"
+#include "session/anf_runtime_algorithm.h"
+#include "operator/ops.h"
+
+namespace mindspore {
+namespace opt {
+namespace {
+AnfNodePtr ModifyReduceOpsAttrs(const CNodePtr &cnode) {
+  auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(cnode, 0);
+  auto input_format = AnfAlgo::GetInputFormat(cnode, 0);
+  if (input_shape.size() == 5 || input_format != kOpFormat_NC1HWC0) {
+    return nullptr;
+  }
+  if (!AnfAlgo::HasNodeAttr(kAttrKeepDims, cnode)) {
+    return nullptr;
+  }
+
+  AnfAlgo::SetNodeAttr(kAttrKeepDims, MakeValue(true), cnode);
+  return cnode;
+}
+
+AnfNodePtr ModifyTileOpAttrs(const CNodePtr &cnode) {
+  auto input_shape = AnfAlgo::GetInputDeviceShape(cnode, 0);
+  if (input_shape.size() != 5) {
+    return nullptr;
+  }
+  if (!AnfAlgo::HasNodeAttr(kAttrMultiples, cnode)) {
+    return nullptr;
+  }
+
+  auto multiples = AnfAlgo::GetNodeAttr<std::vector<int>>(cnode, kAttrMultiples);
+  if (multiples.size() == 4 && multiples[1] == 1) {
+    multiples.push_back(1);
+    AnfAlgo::SetNodeAttr(kAttrMultiples, MakeValue(multiples), cnode);
+  }
+
+  return cnode;
+}
+
+AnfNodePtr ModifyAttrs(const CNodePtr &cnode) {
+  MS_EXCEPTION_IF_NULL(cnode);
+  auto op_name = AnfAlgo::GetCNodeName(cnode);
+  if (op_name == prim::kPrimTile->name()) {
+    return ModifyTileOpAttrs(cnode);
+  } else if (op_name == prim::kPrimReduceSum->name()) {
+    // kPrimReduceMean
+    // kPrimReduceSum
+    // kPrimReduceAll
+    // kPrimReduceMax
+    // kPrimReduceMin
+    return ModifyReduceOpsAttrs(cnode);
+  }
+  return nullptr;
+}
+}  // namespace
+
+const AnfNodePtr ModifyOpAttrs::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
+                                        const EquivPtr &) const {
+  if (node == nullptr || !node->isa<CNode>() || !AnfAlgo::IsGraphKernel(node)) {
+    return nullptr;
+  }
+  MS_LOG(DEBUG) << "====Process op: " << AnfAlgo::GetCNodeName(node);
+  auto fg = AnfAlgo::GetCNodeFuncGraphPtr(node);
+  MS_EXCEPTION_IF_NULL(fg);
+  auto manager = fg->manager();
+  MS_EXCEPTION_IF_NULL(manager);
+  std::vector<AnfNodePtr> todos;
+  kernel::GetValidKernelNodes(fg, &todos);
+  for (auto &t : todos) {
+    auto new_node = ModifyAttrs(t->cast<CNodePtr>());
+    if (new_node != nullptr && new_node != t) {
+      (void)manager->Replace(t, new_node);
+    }
+  }
+  return node;
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/format_type/modify_ops_attrs.h

@@ -0,0 +1,33 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_MODIFY_OPS_ATTRS_H
+#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_MODIFY_OPS_ATTRS_H
+
+#include "pre_activate/common/optimizer.h"
+
+namespace mindspore {
+namespace opt {
+class ModifyOpAttrs : public PatternProcessPass {
+ public:
+  explicit ModifyOpAttrs(bool multigraph = true) : PatternProcessPass("modify_ops_attrs", multigraph) {}
+  ~ModifyOpAttrs() override = default;
+  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
+};
+}  // namespace opt
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_MODIFY_OPS_ATTRS_H

mindspore/ccsrc/pre_activate/ascend/format_type/remove_no_use_reshape_op.cc

@@ -0,0 +1,66 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "pre_activate/ascend/format_type/remove_no_use_reshape_op.h"
+#include <vector>
+#include <memory>
+#include "pre_activate/common/helper.h"
+#include "kernel/common_utils.h"
+#include "session/anf_runtime_algorithm.h"
+#include "operator/ops.h"
+
+namespace mindspore {
+namespace opt {
+namespace {
+AnfNodePtr RemoveReshapeOp(const CNodePtr &cnode) {
+  MS_EXCEPTION_IF_NULL(cnode);
+  auto op_name = AnfAlgo::GetCNodeName(cnode);
+  if (op_name != prim::kPrimReshape->name()) {
+    return nullptr;
+  }
+
+  auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(cnode, 0);
+  auto input_format = AnfAlgo::GetPrevNodeOutputFormat(cnode, 0);
+  if (input_shape.size() != 1 || input_format != kOpFormat_NC1HWC0) {
+    return nullptr;
+  }
+
+  return cnode->input(1);
+}
+}  // namespace
+
+const AnfNodePtr RemoveNoUseReshapeOp::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
+                                               const EquivPtr &) const {
+  if (node == nullptr || !node->isa<CNode>() || !AnfAlgo::IsGraphKernel(node)) {
+    return nullptr;
+  }
+  MS_LOG(DEBUG) << "====process op: " << AnfAlgo::GetCNodeName(node);
+  auto fg = AnfAlgo::GetCNodeFuncGraphPtr(node);
+  MS_EXCEPTION_IF_NULL(fg);
+  auto manager = fg->manager();
+  MS_EXCEPTION_IF_NULL(manager);
+  std::vector<AnfNodePtr> todos;
+  kernel::GetValidKernelNodes(fg, &todos);
+  for (auto &t : todos) {
+    auto new_node = RemoveReshapeOp(t->cast<CNodePtr>());
+    if (new_node != nullptr && new_node != t) {
+      (void)manager->Replace(t, new_node);
+    }
+  }
+  return node;
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/pre_activate/ascend/format_type/remove_no_use_reshape_op.h

@@ -0,0 +1,33 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_REMOVE_NO_USE_RESHAPE_OP_H
+#define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_REMOVE_NO_USE_RESHAPE_OP_H
+
+#include "pre_activate/common/optimizer.h"
+
+namespace mindspore {
+namespace opt {
+class RemoveNoUseReshapeOp : public PatternProcessPass {
+ public:
+  explicit RemoveNoUseReshapeOp(bool multigraph = true) : PatternProcessPass("remove_no_use_reshape_op", multigraph) {}
+  ~RemoveNoUseReshapeOp() override = default;
+  const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
+};
+}  // namespace opt
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_FORMAT_TYPE_REMOVE_NO_USE_RESHAPE_OP_H

mindspore/ccsrc/pre_activate/ascend/ir_fusion/layer_norm_beta_gamma_backprop_fusion.cc

@@ -121,6 +121,9 @@ const AnfNodePtr LayerNormBetaGammaBackpropFusion::Process(const FuncGraphPtr &f
   if (node == nullptr || !node->isa<CNode>()) {
     return nullptr;
   }
+  if (AnfAlgo::IsGraphKernel(node)) {
+    return nullptr;
+  }
   auto cnode = node->cast<CNodePtr>();
   MS_EXCEPTION_IF_NULL(cnode);
   std::vector<CNodePtr> cast_nodes;

mindspore/ccsrc/pre_activate/common/helper.cc

@@ -102,9 +102,12 @@ bool UnVisited(const BaseRef &n) {
       auto prim_py = value->cast<PrimitivePtr>();
       MS_EXCEPTION_IF_NULL(prim_py);
       return !prim_py->HasAttr(kAttrVisited);
-    } else {
-      return false;
+    } else if (IsValueNode<FuncGraph>(in)) {
+      auto func_graph = GetValueNode<FuncGraphPtr>(in);
+      MS_EXCEPTION_IF_NULL(func_graph);
+      return !func_graph->has_flag(kAttrVisited);
     }
+    return false;
   }
   return false;
 }
@@ -188,9 +191,12 @@ bool Visited(const BaseRef &n) {
       auto prim_py = value->cast<PrimitivePtr>();
       MS_EXCEPTION_IF_NULL(prim_py);
       return prim_py->HasAttr(kAttrVisited);
-    } else {
-      return false;
+    } else if (IsValueNode<FuncGraph>(in)) {
+      auto func_graph = GetValueNode<FuncGraphPtr>(in);
+      MS_EXCEPTION_IF_NULL(func_graph);
+      return func_graph->has_flag(kAttrVisited);
     }
+    return false;
   }
   return false;
 }