From a1dd84cc5354dd71fc494da33cc21c90505936ff Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=99=88=E5=8D=8E?= <chenhua26@hisilicon.com>
Date: Mon, 19 Jul 2021 22:32:54 +0800
Subject: [PATCH] aicpu op

---
 ge/engine_manager/dnnengine_manager.cc        |   4 +
 ge/graph/load/model_manager/davinci_model.cc  |  52 ++++
 ge/graph/load/model_manager/davinci_model.h   |   6 +
 .../task_info/kernel_ex_task_info.cc          | 109 ++++++++-
 .../task_info/kernel_ex_task_info.h           |   8 +
 .../task_info/kernel_task_info.cc             | 106 +++++++-
 .../task_info/kernel_task_info.h              |   8 +
 .../node_executor/aicpu/aicpu_ext_info.cc     |  30 +++
 .../node_executor/aicpu/aicpu_ext_info.h      |   5 +
 .../aicpu/aicpu_node_executor.cc              | 121 ++++++++++
 .../node_executor/aicpu/aicpu_node_executor.h |  10 +-
 ge/single_op/task/op_task.cc                  | 117 +++++++++
 ge/single_op/task/op_task.h                   |   7 +
 tests/depends/runtime/src/runtime_stub.cc     |  93 ++++++-
 tests/depends/runtime/src/runtime_stub.h      |  70 ++++++
 tests/ut/ge/CMakeLists.txt                    |   1 +
 .../load/kernel_ex_task_info_unittest.cc      | 141 ++++++++++-
 .../graph/load/kernel_task_info_unittest.cc   | 140 ++++++++++-
 .../aicpu/aicpu_node_executor_unittest.cc     | 227 +++++++++++++++++-
 .../ge/single_op/single_op_task_unittest.cc   | 131 +++++++++-
 .../fwkacllib/inc/cce/fwk_adpt_struct.h       |  16 ++
 third_party/fwkacllib/inc/runtime/config.h    |   8 +
 third_party/fwkacllib/inc/runtime/dev.h       |  12 +
 23 files changed, 1396 insertions(+), 26 deletions(-)
 create mode 100644 tests/depends/runtime/src/runtime_stub.h

diff --git a/ge/engine_manager/dnnengine_manager.cc b/ge/engine_manager/dnnengine_manager.cc
index 0fadd993..36f11828 100644
--- a/ge/engine_manager/dnnengine_manager.cc
+++ b/ge/engine_manager/dnnengine_manager.cc
@@ -239,6 +239,10 @@ std::string DNNEngineManager::GetDNNEngineName(const ge::NodePtr &node_ptr) {
         op_desc->SetOpEngineName(it.engine);
         op_desc->SetOpKernelLibName(kernel_name);
         // set attrs for taking information when load txt to graph object
+        if (it.flagAsync) {
+          GELOGD("Set aicpu blocking op:%s attribute(is_blocking_op):true", op_desc->GetName().c_str());
+          (void)AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+        }
         (void) AttrUtils::SetStr(op_desc, ATTR_NAME_ENGINE_NAME_FOR_LX, it.engine);
         (void) AttrUtils::SetStr(op_desc, ATTR_NAME_KKERNEL_LIB_NAME_FOR_LX, kernel_name);
         GELOGD("DNNEngineManager:Set OpKernelLibName %s and engine name %s to op_desc %s", kernel_name.c_str(),
diff --git a/ge/graph/load/model_manager/davinci_model.cc b/ge/graph/load/model_manager/davinci_model.cc
index aba06173..495ec28e 100755
--- a/ge/graph/load/model_manager/davinci_model.cc
+++ b/ge/graph/load/model_manager/davinci_model.cc
@@ -238,6 +238,12 @@ DavinciModel::~DavinciModel() {
         GE_LOGW_IF(rtEventDestroy(event_list_[i]) != RT_ERROR_NONE, "Destroy event failed, index: %zu", i);
       }
 
+      for (const auto &it : stream_2_event_) {
+        if (rtEventDestroy(it.second) != RT_ERROR_NONE) {
+          GELOGW("Destroy event failed");
+        }
+      }
+
       FreeWeightsMem();
 
       FreeFeatureMapMem();
@@ -4648,4 +4654,50 @@ Status DavinciModel::GetTotalMemSizeExcludeZeroCopy(int64_t &total_useful_size)
   total_useful_size = runtime_param_.mem_size - runtime_param_.zero_copy_size;
   return SUCCESS;
 }
+
+Status DavinciModel::GetEventIdForBlockingAicpuOp(const OpDescPtr &op_desc, rtStream_t stream, uint32_t &event_id) {
+  GELOGI("Get event id for aicpu blocking op:%s", op_desc->GetName().c_str());
+  auto it = stream_2_event_.find(stream);
+  if (it != stream_2_event_.end()) {
+    auto rt_ret = rtGetEventID(it->second, &event_id);
+    if (rt_ret != RT_ERROR_NONE) {
+      REPORT_CALL_ERROR("E19999", "Call rtGetEventID failed for op:%s(%s), ret:0x%X",
+                        op_desc->GetName().c_str(), op_desc->GetType().c_str(), rt_ret);
+      GELOGE(RT_FAILED, "[Call][rtGetEventID] failed for op:%s(%s), ret:0x%X",
+             op_desc->GetName().c_str(), op_desc->GetType().c_str(), rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+  } else {
+    rtEvent_t rt_event = nullptr;
+    auto rt_ret = rtEventCreateWithFlag(&rt_event, RT_EVENT_WITH_FLAG);
+    if (rt_ret != RT_ERROR_NONE) {
+      REPORT_CALL_ERROR("E19999", "Call rtEventCreateWithFlag failed for op:%s(%s), ret:0x%X",
+                        op_desc->GetName().c_str(), op_desc->GetType().c_str(), rt_ret);
+      GELOGE(RT_FAILED, "[Call][rtEventCreateWithFlag] failed for op:%s(%s), ret:0x%X",
+             op_desc->GetName().c_str(), op_desc->GetType().c_str(), rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+    rt_ret = rtGetEventID(rt_event, &event_id);
+    if (rt_ret != RT_ERROR_NONE) {
+      REPORT_CALL_ERROR("E19999", "Call rtGetEventID failed for op:%s(%s), ret:0x%X",
+                        op_desc->GetName().c_str(), op_desc->GetType().c_str(), rt_ret);
+      GELOGE(RT_FAILED, "[Call][rtGetEventID] failed for op:%s(%s), ret:0x%X",
+             op_desc->GetName().c_str(), op_desc->GetType().c_str(), rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+    stream_2_event_.emplace(stream, rt_event);
+  }
+  return SUCCESS;
+}
+
+Status DavinciModel::GetEventByStream(const rtStream_t &stream, rtEvent_t &rt_event) {
+  auto it = stream_2_event_.find(stream);
+  if (it == stream_2_event_.end()) {
+    REPORT_INNER_ERROR("E19999", "Get event failed");
+    GELOGE(FAILED, "[Get][Event] Get event failed");
+    return FAILED;
+  }
+  rt_event = it->second;
+  return SUCCESS;
+}
 }  // namespace ge
diff --git a/ge/graph/load/model_manager/davinci_model.h b/ge/graph/load/model_manager/davinci_model.h
index fe89f66f..76b0beef 100755
--- a/ge/graph/load/model_manager/davinci_model.h
+++ b/ge/graph/load/model_manager/davinci_model.h
@@ -582,6 +582,10 @@ class DavinciModel {
   void SetRunningFlag(bool flag) { running_flg_ = flag; }
   Status SetRunAsyncListenerCallback(const RunAsyncCallback &callback);
 
+  // for blocking aicpu op
+  Status GetEventByStream(const rtStream_t &stream, rtEvent_t &rt_event);
+  Status GetEventIdForBlockingAicpuOp(const OpDescPtr &op_desc, rtStream_t stream, uint32_t &event_id);
+
  private:
   // memory address of weights
   uint8_t *weights_mem_base_;
@@ -1107,6 +1111,8 @@ class DavinciModel {
 
   // op name to attrs mapping
   std::map<std::string, std::map<std::string, std::vector<std::string>>> op_name_to_attrs_;
+
+  std::map<rtStream_t, rtEvent_t> stream_2_event_;
 };
 }  // namespace ge
 #endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_DAVINCI_MODEL_H_
diff --git a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc
index 1a6ab542..fe9cd0cc 100644
--- a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc
+++ b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.cc
@@ -26,8 +26,8 @@
 #include "external/graph/attr_value.h"
 #include "graph/load/model_manager/davinci_model.h"
 #include "graph/load/model_manager/model_manager.h"
-#include "hybrid/node_executor/aicpu/aicpu_ext_info.h"
 #include "framework/common/debug/log.h"
+#include "runtime/rt.h"
 
 namespace {
 const char *const kAicpuAllshape = "_AllShape";
@@ -43,7 +43,7 @@ Status KernelExTaskInfo::InitTaskExtInfo(const std::string &ext_info, const OpDe
   UnknowShapeOpType unknown_type = static_cast<UnknowShapeOpType>(unknown_shape_type_val);
   uint32_t num_inputs = op_desc->GetInputsSize();
   uint32_t num_outputs = op_desc->GetOutputsSize();
-  std::unique_ptr<ge::hybrid::AicpuExtInfoHandler> ext_handle(
+  std::shared_ptr<ge::hybrid::AicpuExtInfoHandler> ext_handle(
           new(std::nothrow) ::ge::hybrid::AicpuExtInfoHandler(op_desc->GetName(),
                                                               num_inputs,
                                                               num_outputs,
@@ -76,6 +76,16 @@ Status KernelExTaskInfo::InitTaskExtInfo(const std::string &ext_info, const OpDe
       }
     }
   }
+
+  AttrUtils::GetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, is_blocking_aicpu_op_);
+  GELOGD("Get op:%s attribute(is_blocking_op), value:%d", op_desc->GetName().c_str(), is_blocking_aicpu_op_);
+
+  if (UpdateEventIdForAicpuBlockingOp(op_desc, ext_handle) != SUCCESS) {
+    GELOGE(FAILED, "[Call][UpdateEventIdForAicpuBlockingOp] failed for op:%s(%s)",
+           op_desc->GetName().c_str(), op_desc->GetType().c_str());
+    return FAILED;
+  }
+
   auto rt_ret = rtMalloc(&ext_info_addr_, ext_handle->GetExtInfoLen(), RT_MEMORY_HBM);
   GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE,
                   REPORT_CALL_ERROR("E19999", "Call rtMalloc failed, size:%zu, ret:0x%X", ext_info.size(), rt_ret);
@@ -448,6 +458,101 @@ Status KernelExTaskInfo::Distribute() {
   stream_id_ = stream_id;
 
   GELOGI("KernelExTaskInfo Distribute Success. task id: %u, stream id: %u", task_id_, stream_id_);
+  if (is_blocking_aicpu_op_) {
+    if (DistributeWaitTaskForAicpuBlockingOp() != SUCCESS) {
+      GELOGE(FAILED, "[Call][DistributeWaitTaskForAicpuBlockingOp] Call DistributeWaitTaskForAicpuBlockingOp failed");
+      return FAILED;
+    }
+  }
+  return SUCCESS;
+}
+
+Status KernelExTaskInfo::CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support) {
+  int32_t device_id = 0;
+  auto rt_ret = rtGetDevice(&device_id);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDevice failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDevice] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  int32_t value = 0;
+  rt_ret = rtGetDeviceCapability(device_id, FEATURE_TYPE_BLOCKING_OPERATOR, RT_MODULE_TYPE_AICPU, &value);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDeviceCapability failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDeviceCapability] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  if (value != RT_AICPU_BLOCKING_OP_NOT_SUPPORT && value != RT_AICPU_BLOCKING_OP_SUPPORT) {
+    REPORT_INNER_ERROR("E19999", "Value should be %d or %d but %d",
+                       RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    GELOGE(FAILED, "[Check][Value] Value should be %d or %d but %d",
+           RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    return FAILED;
+  }
+  is_support = (value == RT_AICPU_BLOCKING_OP_SUPPORT ? true : false);
+  return SUCCESS;
+}
+
+Status KernelExTaskInfo::UpdateEventIdForAicpuBlockingOp(const OpDescPtr &op_desc,
+    std::shared_ptr<ge::hybrid::AicpuExtInfoHandler> &ext_handle) {
+  if (is_blocking_aicpu_op_) {
+    bool is_support = false;
+    if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+      GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+      return FAILED;
+    }
+    if (!is_support) {
+      GELOGD("Device not support blocking aicpu op process");
+      return SUCCESS;
+    }
+    uint32_t event_id = 0;
+    if (davinci_model_->GetEventIdForBlockingAicpuOp(op_desc, stream_, event_id) != SUCCESS) {
+      REPORT_CALL_ERROR("E19999", "Get event id failed for op:%s(%s).", op_desc->GetName().c_str(),
+                        op_desc->GetType().c_str());
+      GELOGE(FAILED, "[Get][EventId] Get event id failed for op:%s(%s)", op_desc->GetName().c_str(),
+             op_desc->GetType().c_str());
+      return FAILED;
+    }
+    if (ext_handle->UpdateEventId(event_id) != SUCCESS) {
+      REPORT_CALL_ERROR("E19999", "Update event id failed for op:%s(%s).", op_desc->GetName().c_str(),
+                        op_desc->GetType().c_str());
+      GELOGE(FAILED, "[Update][EventId] Update event id failed for op:%s(%s)", op_desc->GetName().c_str(),
+             op_desc->GetType().c_str());
+      return FAILED;
+    }
+    GELOGI("Update event_id=%u success", event_id);
+  }
+  return SUCCESS;
+}
+
+Status KernelExTaskInfo::DistributeWaitTaskForAicpuBlockingOp() {
+  bool is_support = false;
+  if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+    GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+    return FAILED;
+  }
+  if (!is_support) {
+    GELOGD("Device not support blocking aicpu op process.");
+    return SUCCESS;
+  }
+  GELOGD("Distribute wait task begin");
+  rtEvent_t rt_event = nullptr;
+  if (davinci_model_->GetEventByStream(stream_, rt_event) != SUCCESS) {
+    GELOGE(FAILED, "[Call][GetEventByStream] Call GetEventByStream failed");
+    return FAILED;
+  }
+  auto rt_ret = rtStreamWaitEvent(stream_, rt_event);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtStreamWaitEvent failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  rt_ret = rtEventReset(rt_event, stream_);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtEventReset failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
   return SUCCESS;
 }
 
diff --git a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h
index 7d07eb7f..eb411576 100644
--- a/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h
+++ b/ge/graph/load/model_manager/task_info/kernel_ex_task_info.h
@@ -19,6 +19,7 @@
 
 #include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/op_desc.h"
+#include "hybrid/node_executor/aicpu/aicpu_ext_info.h"
 
 namespace ge {
 class KernelExTaskInfo : public TaskInfo {
@@ -65,6 +66,12 @@ class KernelExTaskInfo : public TaskInfo {
   void InitDumpArgs(void *addr, const OpDescPtr &op_desc);
   Status InitTaskExtInfo(const std::string &ext_info, const OpDescPtr &op_desc);
 
+  // for blocking aicpu op
+  Status DistributeWaitTaskForAicpuBlockingOp();
+  Status CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support);
+  Status UpdateEventIdForAicpuBlockingOp(const OpDescPtr &op_desc,
+                                         std::shared_ptr<ge::hybrid::AicpuExtInfoHandler> &ext_handle);
+
   uint32_t task_id_;
   uint32_t stream_id_;
   uint32_t dump_flag_;
@@ -79,6 +86,7 @@ class KernelExTaskInfo : public TaskInfo {
   uint32_t args_offset_ = 0;
   int64_t fixed_addr_offset_ = 0;
   int32_t topic_type_flag_ = -1;
+  bool is_blocking_aicpu_op_ = false;
 };
 }  // namespace ge
 #endif  // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_KERNEL_EX_TASK_INFO_H_
diff --git a/ge/graph/load/model_manager/task_info/kernel_task_info.cc b/ge/graph/load/model_manager/task_info/kernel_task_info.cc
index 019a0a8b..6bbfe58e 100755
--- a/ge/graph/load/model_manager/task_info/kernel_task_info.cc
+++ b/ge/graph/load/model_manager/task_info/kernel_task_info.cc
@@ -28,11 +28,10 @@
 #include "graph/load/model_manager/davinci_model.h"
 #include "graph/load/model_manager/model_manager.h"
 #include "graph/load/model_manager/model_utils.h"
-#include "runtime/kernel.h"
+#include "runtime/rt.h"
 #include "graph/load/model_manager/task_info/super_kernel/super_kernel.h"
 #include "graph/load/model_manager/task_info/super_kernel/super_kernel_factory.h"
 #include "cce/aicpu_engine_struct.h"
-#include "hybrid/node_executor/aicpu/aicpu_ext_info.h"
 #include "framework/common/debug/log.h"
 
 namespace {
@@ -474,6 +473,12 @@ Status KernelTaskInfo::Distribute() {
   }
   // set for task_id_
   UpdateTaskId();
+  if (is_blocking_aicpu_op_) {
+    if (DistributeWaitTaskForAicpuBlockingOp() != SUCCESS) {
+      GELOGE(FAILED, "[Call][DistributeWaitTaskForAicpuBlockingOp] Call DistributeWaitTaskForAicpuBlockingOp failed");
+      return FAILED;
+    }
+  }
   GELOGD(
       "KernelTaskInfo Distribute Success. sktenable:%d taskid:%d sktid:%d stubfunc_name:%s stubfunc:%p "
       "blockdim:%d stream:%p",
@@ -482,6 +487,91 @@ Status KernelTaskInfo::Distribute() {
   return SUCCESS;
 }
 
+Status KernelTaskInfo::CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support) {
+  int32_t device_id = 0;
+  auto rt_ret = rtGetDevice(&device_id);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDevice failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDevice] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  int32_t value = 0;
+  rt_ret = rtGetDeviceCapability(device_id, FEATURE_TYPE_BLOCKING_OPERATOR, RT_MODULE_TYPE_AICPU, &value);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDeviceCapability failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDeviceCapability] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  if (value != RT_AICPU_BLOCKING_OP_NOT_SUPPORT && value != RT_AICPU_BLOCKING_OP_SUPPORT) {
+    REPORT_INNER_ERROR("E19999", "Value should be %d or %d but %d",
+                       RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    GELOGE(FAILED, "[Check][Value] Value should be %d or %d but %d",
+           RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    return FAILED;
+  }
+  is_support = (value == RT_AICPU_BLOCKING_OP_SUPPORT ? true : false);
+  return SUCCESS;
+}
+
+Status KernelTaskInfo::UpdateEventIdForAicpuBlockingOp(std::shared_ptr<ge::hybrid::AicpuExtInfoHandler> &ext_handle) {
+  if (is_blocking_aicpu_op_) {
+    bool is_support = false;
+    if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+      GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+      return FAILED;
+    }
+    if (!is_support) {
+      GELOGD("Device not support blocking aicpu op process");
+      return SUCCESS;
+    }
+    uint32_t event_id = 0;
+    if (davinci_model_->GetEventIdForBlockingAicpuOp(op_desc_, stream_, event_id) != SUCCESS) {
+      GELOGE(FAILED, "[Get][EventId] Get event id failed for op:%s(%s)", op_desc_->GetName().c_str(),
+             op_desc_->GetType().c_str());
+      return FAILED;
+    }
+    if (ext_handle->UpdateEventId(event_id) != SUCCESS) {
+      GELOGE(FAILED, "[Update][EventId] Update event id failed for op:%s(%s)", op_desc_->GetName().c_str(),
+             op_desc_->GetType().c_str());
+      return FAILED;
+    }
+    GELOGI("Update event_id=%u success", event_id);
+  }
+  return SUCCESS;
+}
+
+Status KernelTaskInfo::DistributeWaitTaskForAicpuBlockingOp() {
+  bool is_support = false;
+  if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+    GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+    return FAILED;
+  }
+  if (!is_support) {
+    GELOGD("device not support blocking aicpu op process.");
+    return SUCCESS;
+  }
+  GELOGD("Distribute wait task begin");
+  rtEvent_t rt_event = nullptr;
+  if (davinci_model_->GetEventByStream(stream_, rt_event) != SUCCESS) {
+    REPORT_CALL_ERROR("E19999", "Call GetEventByStream failed");
+    GELOGE(FAILED, "[Call][GetEventByStream] Call GetEventByStream failed");
+    return FAILED;
+  }
+  auto rt_ret = rtStreamWaitEvent(stream_, rt_event);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtStreamWaitEvent failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  rt_ret = rtEventReset(rt_event, stream_);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtEventReset failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  return SUCCESS;
+}
+
 void KernelTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) {
   const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam();
   vector<void *> input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc);
@@ -1109,7 +1199,7 @@ Status KernelTaskInfo::InitAicpuTaskExtInfo(const std::string &ext_info) {
   UnknowShapeOpType unknown_type = static_cast<UnknowShapeOpType>(unknown_shape_type_val);
   uint32_t num_inputs = op_desc_->GetInputsSize();
   uint32_t num_outputs = op_desc_->GetOutputsSize();
-  std::unique_ptr<ge::hybrid::AicpuExtInfoHandler> ext_handle(
+  std::shared_ptr<ge::hybrid::AicpuExtInfoHandler> ext_handle(
           new(std::nothrow) ::ge::hybrid::AicpuExtInfoHandler(op_desc_->GetName(),
                                                               num_inputs,
                                                               num_outputs,
@@ -1145,6 +1235,16 @@ Status KernelTaskInfo::InitAicpuTaskExtInfo(const std::string &ext_info) {
                         j, op_desc_->GetName().c_str());
     }
   }
+
+  AttrUtils::GetBool(op_desc_, ATTR_NAME_IS_BLOCKING_OP, is_blocking_aicpu_op_);
+  GELOGD("Get op:%s attribute(is_blocking_op), value:%d", op_desc_->GetName().c_str(), is_blocking_aicpu_op_);
+
+  if (UpdateEventIdForAicpuBlockingOp(ext_handle) != SUCCESS) {
+    GELOGE(FAILED, "[Call][UpdateEventIdForAicpuBlockingOp] failed for op:%s(%s)",
+           op_desc_->GetName().c_str(), op_desc_->GetType().c_str());
+    return FAILED;
+  }
+
   auto rt_ret = rtMalloc(&aicpu_ext_info_addr_, ext_handle->GetExtInfoLen(), RT_MEMORY_HBM);
   if (rt_ret != RT_ERROR_NONE) {
     REPORT_CALL_ERROR("E19999", "Call rtMalloc failed for op:%s(%s), size:%zu, ret:0x%X",
diff --git a/ge/graph/load/model_manager/task_info/kernel_task_info.h b/ge/graph/load/model_manager/task_info/kernel_task_info.h
index d9dd30bb..59a91aee 100644
--- a/ge/graph/load/model_manager/task_info/kernel_task_info.h
+++ b/ge/graph/load/model_manager/task_info/kernel_task_info.h
@@ -24,6 +24,8 @@
 
 #include "graph/load/model_manager/task_info/task_info.h"
 #include "graph/op_desc.h"
+#include "hybrid/node_executor/aicpu/aicpu_ext_info.h"
+
 namespace ge {
 class KernelTaskInfo : public TaskInfo {
  public:
@@ -148,6 +150,11 @@ class KernelTaskInfo : public TaskInfo {
   bool DoubleCallSKTSaveCheck();
   void SetArgs();
 
+  // for blocking aicpu op
+  Status DistributeWaitTaskForAicpuBlockingOp();
+  Status CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support);
+  Status UpdateEventIdForAicpuBlockingOp(std::shared_ptr<ge::hybrid::AicpuExtInfoHandler> &ext_handle);
+
   void *stub_func_;
   void *args_;
   void *sm_desc_;
@@ -187,6 +194,7 @@ class KernelTaskInfo : public TaskInfo {
   uint32_t skt_dump_flag_ = RT_KERNEL_DEFAULT;
   void *superkernel_device_args_addr_ = nullptr;
   void *superkernel_dev_nav_table_ = nullptr;
+  bool is_blocking_aicpu_op_ = false;
 
   struct AICPUCustomInfo {
     void *input_descs = nullptr;
diff --git a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc
index c607a43e..6e8841b9 100644
--- a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc
+++ b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.cc
@@ -81,6 +81,9 @@ Status AicpuExtInfoHandler::Parse(const std::string &ext_info) {
       case aicpu::FWKAdapter::FWK_ADPT_EXT_TOPIC_TYPE:
         GE_CHK_STATUS_RET(ParseExtTopicType(aicpu_ext_info), "[Parse][ExtTopicType] failed.");
         break;
+      case aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT:
+        GE_CHK_STATUS_RET(ParseExtAsyncWait(aicpu_ext_info), "[Parse][ExtAsyncWait] failed.");
+        break;
       default:
         GELOGD("Node[%s] ignore infoType=%d, infoLen=%u.",
                node_name_.c_str(), aicpu_ext_info->infoType, aicpu_ext_info->infoLen);
@@ -101,6 +104,22 @@ Status AicpuExtInfoHandler::Parse(const std::string &ext_info) {
   return SUCCESS;
 }
 
+Status AicpuExtInfoHandler::ParseExtAsyncWait(AicpuExtInfo *aicpu_ext_info) {
+  if (aicpu_ext_info->infoLen != sizeof(AsyncWaitInfo)) {
+    REPORT_INNER_ERROR("E19999",
+                       "Node[%s] parse ext async wait info failed as infoLen must be %zu but %u.",
+                       node_name_.c_str(), sizeof(AsyncWaitInfo), aicpu_ext_info->infoLen);
+    GELOGE(ACL_ERROR_GE_PARAM_INVALID,
+           "[Check][DataLen]Node[%s] parse ext async wait info failed as infoLen must be %zu but %u.",
+           node_name_.c_str(), sizeof(AsyncWaitInfo), aicpu_ext_info->infoLen);
+    return ACL_ERROR_GE_PARAM_INVALID;
+  }
+
+  async_wait_ = reinterpret_cast<AsyncWaitInfo *>(aicpu_ext_info->infoMsg);
+  GELOGI("Node[%s] parse async wait info success infoLen=%u.", node_name_.c_str(), aicpu_ext_info->infoLen);
+  return SUCCESS;
+}
+
 Status AicpuExtInfoHandler::ParseExtShapeType(AicpuExtInfo *aicpu_ext_info) {
   GE_IF_BOOL_EXEC(aicpu_ext_info->infoLen != sizeof(int32_t),
                   REPORT_INNER_ERROR("E19999", "Node[%s] parse ext shape type failed as infoLen must be %zu but %u.",
@@ -280,6 +299,17 @@ Status AicpuExtInfoHandler::UpdateSessionInfo(uint64_t session_id, uint64_t kern
   return SUCCESS;
 }
 
+Status AicpuExtInfoHandler::UpdateEventId(uint32_t event_id) {
+  if (async_wait_ == nullptr) {
+    REPORT_INNER_ERROR("E19999", "async_wait_ is nullptr.");
+    GELOGE(FAILED, "[Check][async_wait_] async_wait_ is nullptr.");
+    return FAILED;
+  }
+  async_wait_->waitType = 1;
+  async_wait_->waitId = event_id;
+  return SUCCESS;
+}
+
 Status AicpuExtInfoHandler::UpdateSessionInfoSessionId(uint64_t session_id) {
   if (session_info_ == nullptr) {
     GELOGD("There is no session info in ext_info, no need update.");
diff --git a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h
index 46fb7c05..80e3bb92 100644
--- a/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h
+++ b/ge/hybrid/node_executor/aicpu/aicpu_ext_info.h
@@ -27,6 +27,7 @@ namespace ge {
 namespace hybrid {
 using AicpuShapeAndType = aicpu::FWKAdapter::ShapeAndType;
 using AicpuExtInfo = aicpu::FWKAdapter::ExtInfo;
+using AsyncWaitInfo = aicpu::FWKAdapter::AsyncWait;
 using AicpuSessionInfo = SessionInfo;
 
 class AicpuExtInfoHandler {
@@ -59,6 +60,8 @@ class AicpuExtInfoHandler {
 
   Status UpdateExecuteMode(bool flag);
 
+  Status UpdateEventId(uint32_t event_id);
+
   Status GetOutputShapeAndType(uint32_t output_index, GeShape &shape, DataType &data_type);
 
   bool IsNeedRefreshIOAddr();
@@ -73,6 +76,7 @@ class AicpuExtInfoHandler {
   Status ParseExtBitMap(AicpuExtInfo *aicpu_ext_info);
   Status ParseExtUpdateAddr(AicpuExtInfo *aicpu_ext_info);
   Status ParseExtTopicType(AicpuExtInfo *aicpu_ext_info);
+  Status ParseExtAsyncWait(AicpuExtInfo *aicpu_ext_info);
 
   static Status UpdateShapeAndType(const GeShape &shape,
                                    DataType data_type,
@@ -90,6 +94,7 @@ class AicpuExtInfoHandler {
   const uint32_t output_num_;
   UnknowShapeOpType unknown_type_;
   AicpuSessionInfo *session_info_ = nullptr;
+  AsyncWaitInfo *async_wait_ = nullptr;
   uint64_t *bit_map_ = nullptr;
   uint32_t *update_addr_ = nullptr;
   int32_t topic_type_flag_ = -1;
diff --git a/ge/hybrid/node_executor/aicpu/aicpu_node_executor.cc b/ge/hybrid/node_executor/aicpu/aicpu_node_executor.cc
index cf20303c..f309ebd0 100755
--- a/ge/hybrid/node_executor/aicpu/aicpu_node_executor.cc
+++ b/ge/hybrid/node_executor/aicpu/aicpu_node_executor.cc
@@ -22,6 +22,7 @@
 #include "graph/utils/node_utils.h"
 #include "hybrid/executor/hybrid_execution_context.h"
 #include "hybrid/model/hybrid_model.h"
+#include "runtime/rt.h"
 
 namespace ge {
 namespace hybrid {
@@ -33,6 +34,12 @@ const char *const kAicpuAllshape = "_AllShape";
 REGISTER_NODE_EXECUTOR_BUILDER(NodeExecutorManager::ExecutorType::AICPU_TF, AiCpuNodeExecutor);
 REGISTER_NODE_EXECUTOR_BUILDER(NodeExecutorManager::ExecutorType::AICPU_CUSTOM, AiCpuNodeExecutor);
 
+AicpuNodeTaskBase::~AicpuNodeTaskBase() {
+  if (rt_event_ != nullptr) {
+    (void)rtEventDestroy(rt_event_);
+  }
+}
+
 Status AicpuNodeTaskBase::AllocTensorBuffer(size_t size, std::unique_ptr<TensorBuffer> &tensor_buffer) {
   auto allocator = NpuMemoryAllocator::GetAllocator();
   GE_CHECK_NOTNULL(allocator);
@@ -64,6 +71,13 @@ Status AicpuNodeTaskBase::InitExtInfo(const std::string &kernel_ext_info, int64_
   GE_CHK_STATUS_RET(aicpu_ext_handle_.UpdateSessionInfoSessionId(session_id),
                     "[Update][SessionInfoSessionId] failed, session_id:%ld.", session_id);
 
+  if (is_blocking_aicpu_op_) {
+    if (UpdateEventIdForBlockingAicpuOp() != SUCCESS) {
+      GELOGE(FAILED, "[Call][UpdateEventIdForBlockingAicpuOp] Call UpdateEventIdForBlockingAicpuOp failed");
+      return FAILED;
+    }
+  }
+
   // copy task args buf
   GE_CHK_STATUS_RET(AllocTensorBuffer(aicpu_ext_handle_.GetExtInfoLen(), ext_info_addr_dev_),
                     "[Invoke][AllocTensorBuffer]Node[%s] alloc kernel_ext_info buf failed, size=%zu",
@@ -230,6 +244,96 @@ Status AicpuNodeTaskBase::ExecuteAsync(TaskContext &context, std::function<void(
   return SUCCESS;
 }
 
+Status AicpuNodeTaskBase::UpdateEventIdForBlockingAicpuOp() {
+  bool is_support = false;
+  if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+    GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+    return FAILED;
+  }
+  if (!is_support) {
+    GELOGD("Device not support blocking aicpu op process");
+    return SUCCESS;
+  }
+  uint32_t event_id = 0;
+  auto rt_ret = rtEventCreateWithFlag(&rt_event_, RT_EVENT_WITH_FLAG);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtEventCreateWithFlag failed for node:%s, ret:0x%X", node_name_.c_str(),
+                      rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtEventCreateWithFlag] failed for node:%s, ret:0x%X", node_name_.c_str(), rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  rt_ret = rtGetEventID(rt_event_, &event_id);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetEventID failed for node:%s, ret:0x%X", node_name_.c_str(), rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetEventID] failed for node:%s, ret:0x%X", node_name_.c_str(), rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  if (aicpu_ext_handle_.UpdateEventId(event_id) != SUCCESS) {
+    REPORT_CALL_ERROR("E19999", "Update event id failed for node:%s.", node_name_.c_str());
+    GELOGE(FAILED, "[Update][EventId] Update event id failed for node:%s", node_name_.c_str());
+    return FAILED;
+  }
+  GELOGI("Update event_id=%u success", event_id);
+  return SUCCESS;
+}
+
+Status AicpuNodeTaskBase::CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support) {
+  int32_t device_id = 0;
+  auto rt_ret = rtGetDevice(&device_id);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDevice failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDevice] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  int32_t value = 0;
+  rt_ret = rtGetDeviceCapability(device_id, FEATURE_TYPE_BLOCKING_OPERATOR, RT_MODULE_TYPE_AICPU, &value);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDeviceCapability failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDeviceCapability] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  if (value != RT_AICPU_BLOCKING_OP_NOT_SUPPORT && value != RT_AICPU_BLOCKING_OP_SUPPORT) {
+    REPORT_INNER_ERROR("E19999", "Value should be %d or %d but %d",
+                       RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    GELOGE(FAILED, "[Check][Value] Value should be %d or %d but %d",
+           RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    return FAILED;
+  }
+  is_support = (value == RT_AICPU_BLOCKING_OP_SUPPORT ? true : false);
+  return SUCCESS;
+}
+
+Status AicpuNodeTaskBase::DistributeWaitTaskForAicpuBlockingOp(rtStream_t stream) {
+  bool is_support = false;
+  if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+    GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+    return FAILED;
+  }
+  if (!is_support) {
+    GELOGD("Device not support blocking aicpu op process.");
+    return SUCCESS;
+  }
+  GELOGD("Distribute queue task begin");
+  if (rt_event_ == nullptr) {
+    REPORT_INNER_ERROR("E19999", "rt_event_ is nullptr");
+    GELOGE(FAILED, "[Check][rt_event_] rt_event_ is nullptr");
+    return FAILED;
+  }
+  auto rt_ret = rtStreamWaitEvent(stream, rt_event_);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtStreamWaitEvent failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  rt_ret = rtEventReset(rt_event_, stream);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtEventReset failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  return SUCCESS;
+}
+
 Status AicpuTfNodeTask::InitForDependComputeTask() {
   if ((unknown_type_ != DEPEND_COMPUTE) || (node_item_->num_outputs == 0)) {
     GELOGD("Node[%s] type[%s] unknown_type is %d, output num is %d.",
@@ -325,6 +429,9 @@ Status AicpuTfNodeTask::Init(const HybridModel &model) {
 
   // init ext info
   uint64_t ext_session_id = model.GetSessionId();
+  const OpDescPtr op_desc = node_item_->GetOpDesc();
+  AttrUtils::GetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, is_blocking_aicpu_op_);
+  GELOGD("Get op:%s attribute(is_blocking_op), value:%d", op_desc->GetName().c_str(), is_blocking_aicpu_op_);
   GE_CHK_STATUS_RET(InitExtInfo(kernel_ext_info, ext_session_id), "[Init][ExtInfo] failed for Node[%s].",
                     node_name_.c_str());
   GE_CHK_STATUS_RET(InitForDependComputeTask(), "[Init][DependComputeTask] failed for Node[%s].", node_name_.c_str());
@@ -642,6 +749,12 @@ Status AicpuTfNodeTask::LaunchTask(TaskContext &context) {
                                   kernel_buf_->GetSize(), flag, context.GetStream()));
   RECORD_EXECUTION_EVENT(context.GetExecutionContext(), node_name_.c_str(), "[AicpuTfNodertKernelLaunchEx] End");
   GELOGD("Node[%s] launch end.", node_name_.c_str());
+  if (is_blocking_aicpu_op_) {
+    if (DistributeWaitTaskForAicpuBlockingOp(context.GetStream()) != SUCCESS) {
+      GELOGE(FAILED, "[Call][DistributeWaitTaskForAicpuBlockingOp] Call DistributeWaitTaskForAicpuBlockingOp failed");
+      return FAILED;
+    }
+  }
   if (need_sync_) {
     GELOGD("[%s] Task needs sync", node_name_.c_str());
     GE_CHK_STATUS_RET_NOLOG(context.Synchronize());
@@ -760,6 +873,8 @@ Status AicpuNodeTask::Init(const HybridModel &model) {
                   return FAILED;);
 
   uint64_t ext_session_id = model.GetSessionId();
+  AttrUtils::GetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, is_blocking_aicpu_op_);
+  GELOGD("Get op:%s attribute(is_blocking_op), value:%d", op_desc->GetName().c_str(), is_blocking_aicpu_op_);
   GE_CHK_STATUS_RET(InitExtInfo(kernel_ext_info, ext_session_id),
                     "[Init][ExtInfo] failed for Node[%s].", node_name.c_str());
 
@@ -826,6 +941,12 @@ Status AicpuNodeTask::LaunchTask(TaskContext &context) {
                                             args_.get(), args_size_,
                                             nullptr, context.GetStream(), flag);
   GE_CHK_RT_RET(rt_ret);
+  if (is_blocking_aicpu_op_) {
+    if (DistributeWaitTaskForAicpuBlockingOp(context.GetStream()) != SUCCESS) {
+      GELOGE(FAILED, "[Call][DistributeWaitTaskForAicpuBlockingOp] Call DistributeWaitTaskForAicpuBlockingOp failed");
+      return FAILED;
+    }
+  }
   GELOGD("Node[%s] launch task end.", node_name_.c_str());
   return SUCCESS;
 }
diff --git a/ge/hybrid/node_executor/aicpu/aicpu_node_executor.h b/ge/hybrid/node_executor/aicpu/aicpu_node_executor.h
index 14bc8fcc..3911e090 100644
--- a/ge/hybrid/node_executor/aicpu/aicpu_node_executor.h
+++ b/ge/hybrid/node_executor/aicpu/aicpu_node_executor.h
@@ -35,7 +35,7 @@ class AicpuNodeTaskBase : public NodeTask {
                           node_item->num_outputs,
                           node_item->shape_inference_type) {}
 
-  ~AicpuNodeTaskBase() override = default;
+  ~AicpuNodeTaskBase() override;
 
   using NodeTask::Init;
 
@@ -61,6 +61,10 @@ class AicpuNodeTaskBase : public NodeTask {
 
   static Status AllocTensorBuffer(size_t size, std::unique_ptr<TensorBuffer> &tensor_buffer);
 
+  Status DistributeWaitTaskForAicpuBlockingOp(rtStream_t stream);
+  Status CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support);
+  Status UpdateEventIdForBlockingAicpuOp();
+
  protected:
   const NodeItem *node_item_;
   // just reference.
@@ -78,6 +82,10 @@ class AicpuNodeTaskBase : public NodeTask {
 
   // ext info addr, device mem
   std::unique_ptr<TensorBuffer> ext_info_addr_dev_;
+
+  // for blocking aicpu op
+  bool is_blocking_aicpu_op_ = false;
+  rtEvent_t rt_event_ = nullptr;
 };
 
 class AicpuTfNodeTask : public AicpuNodeTaskBase {
diff --git a/ge/single_op/task/op_task.cc b/ge/single_op/task/op_task.cc
index dbc90ac5..83cb0529 100755
--- a/ge/single_op/task/op_task.cc
+++ b/ge/single_op/task/op_task.cc
@@ -564,6 +564,41 @@ AiCpuBaseTask::~AiCpuBaseTask() {
   if (ext_info_addr_dev_ != nullptr) {
     (void)rtFree(ext_info_addr_dev_);
   }
+  if (rt_event_ != nullptr) {
+    (void)rtEventDestroy(rt_event_);
+  }
+}
+
+Status AiCpuBaseTask::UpdateEventIdForBlockingAicpuOp() {
+  bool is_support = false;
+  if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+    GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+    return FAILED;
+  }
+  if (!is_support) {
+    GELOGD("Device not support blocking aicpu op process");
+    return SUCCESS;
+  }
+  uint32_t event_id = 0;
+  auto rt_ret = rtEventCreateWithFlag(&rt_event_, RT_EVENT_WITH_FLAG);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtEventCreateWithFlag failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtEventCreateWithFlag] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  rt_ret = rtGetEventID(rt_event_, &event_id);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetEventID failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetEventID] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  if (aicpu_ext_handle_->UpdateEventId(event_id) != SUCCESS) {
+    REPORT_CALL_ERROR("E19999", "Update event id=%u failed.", event_id);
+    GELOGE(FAILED, "[Update][EventId] Update event id failed", event_id);
+    return FAILED;
+  }
+  GELOGI("Update event_id=%u success", event_id);
+  return SUCCESS;
 }
 
 Status AiCpuBaseTask::SetExtInfoAndType(const std::string &kernel_ext_info, uint64_t kernel_id) {
@@ -577,6 +612,9 @@ Status AiCpuBaseTask::SetExtInfoAndType(const std::string &kernel_ext_info, uint
   GELOGD("Get unknown_type is %d.", unknown_shape_type_val);
   unknown_type_ = static_cast<UnknowShapeOpType>(unknown_shape_type_val);
 
+  AttrUtils::GetBool(op_desc_, ATTR_NAME_IS_BLOCKING_OP, is_blocking_aicpu_op_);
+  GELOGD("Get op:%s attribute(is_blocking_op), value:%d", op_desc_->GetName().c_str(), is_blocking_aicpu_op_);
+
   aicpu_ext_handle_.reset(new(std::nothrow) ::ge::hybrid::AicpuExtInfoHandler(op_desc_->GetName(),
                                                                               num_inputs_,
                                                                               num_outputs_,
@@ -595,6 +633,13 @@ Status AiCpuBaseTask::SetExtInfoAndType(const std::string &kernel_ext_info, uint
   GE_CHK_STATUS_RET(aicpu_ext_handle_->UpdateSessionInfo(ULLONG_MAX, kernel_id, false),
                     "[Update][SessionInfo] failed.");
 
+  if (is_blocking_aicpu_op_) {
+    if (UpdateEventIdForBlockingAicpuOp() != SUCCESS) {
+      GELOGE(FAILED, "[Call][UpdateEventIdForBlockingAicpuOp] Call UpdateEventIdForBlockingAicpuOp failed");
+      return FAILED;
+    }
+  }
+
   GE_CHK_RT_RET(rtMalloc(&ext_info_addr_dev_, aicpu_ext_handle_->GetExtInfoLen(), RT_MEMORY_HBM));
   GE_CHK_RT_RET(rtMemcpy(ext_info_addr_dev_, aicpu_ext_handle_->GetExtInfoLen(),
                          aicpu_ext_handle_->GetExtInfo(), aicpu_ext_handle_->GetExtInfoLen(),
@@ -770,6 +815,63 @@ Status AiCpuBaseTask::UpdateIoAddr(const vector<DataBuffer> &inputs, const vecto
   return SUCCESS;
 }
 
+Status AiCpuBaseTask::CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support) {
+  int32_t device_id = 0;
+  auto rt_ret = rtGetDevice(&device_id);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDevice failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDevice] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  int32_t value = 0;
+  rt_ret = rtGetDeviceCapability(device_id, FEATURE_TYPE_BLOCKING_OPERATOR, RT_MODULE_TYPE_AICPU, &value);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtGetDeviceCapability failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][rtGetDeviceCapability] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  if (value != RT_AICPU_BLOCKING_OP_NOT_SUPPORT && value != RT_AICPU_BLOCKING_OP_SUPPORT) {
+    REPORT_INNER_ERROR("E19999", "Value should be %d or %d but %d",
+                       RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    GELOGE(FAILED, "[Check][Value] Value should be %d or %d but %d",
+           RT_AICPU_BLOCKING_OP_NOT_SUPPORT, RT_AICPU_BLOCKING_OP_SUPPORT, value);
+    return FAILED;
+  }
+  is_support = (value == RT_AICPU_BLOCKING_OP_SUPPORT ? true : false);
+  return SUCCESS;
+}
+
+Status AiCpuBaseTask::DistributeWaitTaskForAicpuBlockingOp(rtStream_t stream) {
+  bool is_support = false;
+  if (CheckDeviceSupportBlockingAicpuOpProcess(is_support) != SUCCESS) {
+    GELOGE(FAILED, "[Call][CheckDeviceSupportBlockingAicpuOpProcess] Call CheckDeviceSupportBlockingAicpuOpProcess failed");
+    return FAILED;
+  }
+  if (!is_support) {
+    GELOGD("Device not support blocking aicpu op process.");
+    return SUCCESS;
+  }
+  GELOGI("Distribute queue task begin");
+  if (rt_event_ == nullptr) {
+    REPORT_INNER_ERROR("E19999", "rt_event_ is nullptr");
+    GELOGE(FAILED, "[Check][rt_event_] rt_event_ is nullptr");
+    return FAILED;
+  }
+  auto rt_ret = rtStreamWaitEvent(stream, rt_event_);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtStreamWaitEvent failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  rt_ret = rtEventReset(rt_event_, stream);
+  if (rt_ret != RT_ERROR_NONE) {
+    REPORT_CALL_ERROR("E19999", "Call rtEventReset failed, ret:0x%X", rt_ret);
+    GELOGE(RT_FAILED, "[Call][RtApi] failed, ret:0x%X", rt_ret);
+    return RT_ERROR_TO_GE_STATUS(rt_ret);
+  }
+  return SUCCESS;
+}
+
 AiCpuTask::~AiCpuTask() {
   FreeHbm(args_);
   FreeHbm(io_addr_);
@@ -813,6 +915,14 @@ Status AiCpuTask::LaunchKernel(rtStream_t stream) {
   GELOGI("[TASK_INFO] %lu/%s", kernel_id_, op_type_.c_str());
 
   GELOGD("Done launch kernel successfully. task = %s", this->op_type_.c_str());
+
+  if (is_blocking_aicpu_op_) {
+    if (DistributeWaitTaskForAicpuBlockingOp(stream) != SUCCESS) {
+      GELOGE(FAILED, "[Call][DistributeWaitTaskForAicpuBlockingOp] Call DistributeWaitTaskForAicpuBlockingOp failed");
+      return FAILED;
+    }
+  }
+
   return SUCCESS;
 }
 
@@ -1089,6 +1199,13 @@ Status AiCpuCCTask::LaunchKernel(rtStream_t stream) {
   }
   GELOGI("[TASK_INFO] %lu/%s", kernel_id_, op_type_.c_str());
   GELOGD("Invoke rtCpuKernelLaunch succeeded");
+
+  if (is_blocking_aicpu_op_) {
+    if (DistributeWaitTaskForAicpuBlockingOp(stream) != SUCCESS) {
+      GELOGE(FAILED, "[Call][DistributeWaitTaskForAicpuBlockingOp] Call DistributeWaitTaskForAicpuBlockingOp failed");
+      return FAILED;
+    }
+  }
   return SUCCESS;
 }
 
diff --git a/ge/single_op/task/op_task.h b/ge/single_op/task/op_task.h
index 132672b0..adf51dba 100644
--- a/ge/single_op/task/op_task.h
+++ b/ge/single_op/task/op_task.h
@@ -178,6 +178,10 @@ class AiCpuBaseTask : public OpTask {
                        rtStream_t stream);
   Status UpdateOutputShape(vector<GeTensorDesc> &output_desc);
   Status UpdateShapeToOutputDesc(const GeShape &shape_new, GeTensorDesc &output_desc);
+  // for blocking aicpu op
+  Status DistributeWaitTaskForAicpuBlockingOp(rtStream_t stream);
+  Status UpdateEventIdForBlockingAicpuOp();
+  Status CheckDeviceSupportBlockingAicpuOpProcess(bool &is_support);
 
  protected:
   size_t num_inputs_ = 0;
@@ -186,6 +190,9 @@ class AiCpuBaseTask : public OpTask {
   std::unique_ptr<ge::hybrid::AicpuExtInfoHandler> aicpu_ext_handle_;
   void *ext_info_addr_dev_ = nullptr;
   vector<bool> input_is_const_;
+  // for blocking aicpu op
+  bool is_blocking_aicpu_op_ = false;
+  rtEvent_t rt_event_ = nullptr;
 };
 
 class AiCpuTask : public AiCpuBaseTask {
diff --git a/tests/depends/runtime/src/runtime_stub.cc b/tests/depends/runtime/src/runtime_stub.cc
index 25d6c2d3..32df7552 100644
--- a/tests/depends/runtime/src/runtime_stub.cc
+++ b/tests/depends/runtime/src/runtime_stub.cc
@@ -16,12 +16,94 @@
 
 #include <cce/dnn.h>
 #include <securec.h>
+#include "runtime_stub.h"
+#include "runtime/rt.h"
+
+#define ADD_STUB_RETURN_VALUE(FUNC, TYPE) std::vector<TYPE> g_Stub_##FUNC##_RETURN
+
+#define GET_STUB_RETURN_VALUE(FUNC, TYPE, DEFAULT) ({   \
+  TYPE result = DEFAULT;                                \
+  if (!g_Stub_##FUNC##_RETURN.empty()) {                \
+    result = g_Stub_##FUNC##_RETURN.back();             \
+    g_Stub_##FUNC##_RETURN.pop_back();                  \
+  }                                                     \
+  result;                                               \
+})
+
+#define DEL_STUB_RETURN_VALUE(FUNC, TYPE)           \
+do {                                                \
+  extern std::vector<TYPE> g_Stub_##FUNC##_RETURN;  \
+  g_Stub_##FUNC##_RETURN.clear();                   \
+} while (0)
+
+
+#define ADD_STUB_OUTBOUND_VALUE(FUNC, TYPE, NAME) std::vector<TYPE> g_Stub_##FUNC##_OUT_##NAME
+
+#define GET_STUB_OUTBOUND_VALUE(FUNC, TYPE, NAME, DEFAULT) ({ \
+  TYPE value;                                                 \
+  if (!g_Stub_##FUNC##_OUT_##NAME.empty()) {                  \
+    value = g_Stub_##FUNC##_OUT_##NAME.back();                \
+    g_Stub_##FUNC##_OUT_##NAME.pop_back();                    \
+  } else {                                                    \
+    value = DEFAULT;                                          \
+  }                                                           \
+  value;                                                      \
+})
+
+#define DEL_STUB_OUTBOUND_VALUE(FUNC, TYPE, NAME)       \
+do {                                                    \
+  extern std::vector<TYPE> g_Stub_##FUNC##_OUT_##NAME;  \
+  g_Stub_##FUNC##_OUT_##NAME.clear();                   \
+} while (0)
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 #define EVENT_LENTH 10
 
+void rtStubTearDown() {
+  DEL_STUB_RETURN_VALUE(rtGetDevice, rtError_t);
+  DEL_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t);
+  DEL_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t);
+  DEL_STUB_RETURN_VALUE(rtEventReset, rtError_t);
+  DEL_STUB_RETURN_VALUE(rtEventCreate, rtError_t);
+  DEL_STUB_RETURN_VALUE(rtGetEventID, rtError_t);
+}
+
+ADD_STUB_RETURN_VALUE(rtGetDevice, rtError_t);
+rtError_t rtGetDevice(int32_t *device) {
+  return GET_STUB_RETURN_VALUE(rtGetDevice, rtError_t, RT_ERROR_NONE);
+}
+
+ADD_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t);
+ADD_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value);
+rtError_t rtGetDeviceCapability(int32_t device, int32_t moduleType, int32_t featureType, int32_t *value) {
+  *value = GET_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_SUPPORT);
+  return GET_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+}
+
+ADD_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t);
+rtError_t rtStreamWaitEvent(rtStream_t stream, rtEvent_t event) {
+  return GET_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t, RT_ERROR_NONE);
+}
+
+ADD_STUB_RETURN_VALUE(rtEventReset, rtError_t);
+rtError_t rtEventReset(rtEvent_t event, rtStream_t stream) {
+  return GET_STUB_RETURN_VALUE(rtEventReset, rtError_t, RT_ERROR_NONE);
+}
+
+ADD_STUB_RETURN_VALUE(rtEventCreate, rtError_t);
+rtError_t rtEventCreate(rtEvent_t *event) {
+  *event = new int[EVENT_LENTH];
+  return GET_STUB_RETURN_VALUE(rtEventCreate, rtError_t, RT_ERROR_NONE);
+}
+
+ADD_STUB_RETURN_VALUE(rtGetEventID, rtError_t);
+rtError_t rtGetEventID(rtEvent_t event, uint32_t *event_id) {
+  *event_id = 0;
+  return GET_STUB_RETURN_VALUE(rtEventCreate, rtError_t, RT_ERROR_NONE);
+}
+
 rtError_t rtCtxSetCurrent(rtContext_t ctx) { return RT_ERROR_NONE; }
 
 rtError_t rtGetStreamId(rtStream_t stream, int32_t *stream_id) {
@@ -42,11 +124,6 @@ rtError_t rtEventGetTimeStamp(uint64_t *time, rtEvent_t event) {
   return RT_ERROR_NONE;
 }
 
-rtError_t rtEventCreate(rtEvent_t *event) {
-  *event = new int[EVENT_LENTH];
-  return RT_ERROR_NONE;
-}
-
 rtError_t rtEventCreateWithFlag(rtEvent_t *event, uint32_t flag) {
   return rtEventCreate(event);
 }
@@ -112,8 +189,6 @@ rtError_t rtMemcpyAsync(void *dst, uint64_t dest_max, const void *src, uint64_t
   return RT_ERROR_NONE;
 }
 
-rtError_t rtStreamWaitEvent(rtStream_t stream, rtEvent_t event) { return RT_ERROR_NONE; }
-
 rtError_t rtSetTSDevice(uint32_t tsId) {
   return RT_ERROR_NONE;
 }
@@ -347,10 +422,6 @@ rtError_t rtStreamSwitchEx(void *ptr, rtCondition_t condition, void *value_ptr,
 
 rtError_t rtStreamActive(rtStream_t active_stream, rtStream_t stream) { return RT_ERROR_NONE; }
 
-rtError_t rtEventReset(rtEvent_t event, rtStream_t stream) { return RT_ERROR_NONE; }
-
-rtError_t rtGetDevice(int32_t *device) { return RT_ERROR_NONE; }
-
 rtError_t rtDatadumpInfoLoad(const void *dump_info, uint32_t length) { return RT_ERROR_NONE; }
 
 rtError_t rtKernelLaunchWithFlag(const void *stub_func, uint32_t block_dim, void *args, uint32_t args_size,
diff --git a/tests/depends/runtime/src/runtime_stub.h b/tests/depends/runtime/src/runtime_stub.h
new file mode 100644
index 00000000..b693b9ea
--- /dev/null
+++ b/tests/depends/runtime/src/runtime_stub.h
@@ -0,0 +1,70 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __INC_LLT_RUNTIME_STUB_H
+#define __INC_LLT_RUNTIME_STUB_H
+
+#include <vector>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+void rtStubTearDown();
+
+#define RTS_STUB_SETUP()    \
+do {                        \
+  rtStubTearDown();         \
+} while (0)
+
+#define RTS_STUB_TEARDOWN() \
+do {                        \
+  rtStubTearDown();         \
+} while (0)
+
+#define RTS_STUB_RETURN_VALUE(FUNC, TYPE, VALUE)                          \
+do {                                                                      \
+  g_Stub_##FUNC##_RETURN.emplace(g_Stub_##FUNC##_RETURN.begin(), VALUE);  \
+} while (0)
+
+#define RTS_STUB_OUTBOUND_VALUE(FUNC, TYPE, NAME, VALUE)                          \
+do {                                                                              \
+  g_Stub_##FUNC##_OUT_##NAME.emplace(g_Stub_##FUNC##_OUT_##NAME.begin(), VALUE);  \
+} while (0)
+
+
+#define RTS_STUB_RETURN_EXTERN(FUNC, TYPE) extern std::vector<TYPE> g_Stub_##FUNC##_RETURN;
+#define RTS_STUB_OUTBOUND_EXTERN(FUNC, TYPE, NAME) extern std::vector<TYPE> g_Stub_##FUNC##_OUT_##NAME;
+
+RTS_STUB_RETURN_EXTERN(rtGetDevice, rtError_t);
+RTS_STUB_OUTBOUND_EXTERN(rtGetDevice, int32_t, device)
+
+RTS_STUB_RETURN_EXTERN(rtGetDeviceCapability, rtError_t);
+RTS_STUB_OUTBOUND_EXTERN(rtGetDeviceCapability, int32_t, value);
+
+RTS_STUB_RETURN_EXTERN(rtStreamWaitEvent, rtError_t);
+
+RTS_STUB_RETURN_EXTERN(rtEventReset, rtError_t);
+
+RTS_STUB_RETURN_EXTERN(rtEventCreate, rtError_t);
+RTS_STUB_OUTBOUND_EXTERN(rtEventCreate, rtEvent_t, event);
+
+RTS_STUB_RETURN_EXTERN(rtGetEventID, rtError_t);
+RTS_STUB_OUTBOUND_EXTERN(rtEventCreate, uint32_t, event_id);
+
+#ifdef __cplusplus
+}
+#endif
+#endif // __INC_LLT_RUNTIME_STUB_H
diff --git a/tests/ut/ge/CMakeLists.txt b/tests/ut/ge/CMakeLists.txt
index a0790cf2..f9d9e857 100755
--- a/tests/ut/ge/CMakeLists.txt
+++ b/tests/ut/ge/CMakeLists.txt
@@ -935,6 +935,7 @@ target_link_libraries(ge_single_op PRIVATE
     ascend_protobuf
     json
     c_sec
+    runtime_stub
 )
 
 # ut binary
diff --git a/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc b/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc
index 327dd248..86569789 100644
--- a/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc
+++ b/tests/ut/ge/graph/load/kernel_ex_task_info_unittest.cc
@@ -23,15 +23,20 @@
 
 #include "graph/load/model_manager/task_info/kernel_ex_task_info.h"
 #include "cce/aicpu_engine_struct.h"
+#include "tests/depends/runtime/src/runtime_stub.h"
 
 namespace ge {
 extern OpDescPtr CreateOpDesc(string name, string type);
 
 class UtestKernelExTaskInfo : public testing::Test {
  protected:
-  void SetUp() {}
+  void SetUp() {
+    RTS_STUB_SETUP();
+  }
 
-  void TearDown() {}
+  void TearDown() {
+    RTS_STUB_TEARDOWN();
+  }
 };
 
 // test kernel_ex_task_Release
@@ -209,4 +214,136 @@ TEST_F(UtestKernelExTaskInfo, parse_topic_type_failed_2) {
   KernelExTaskInfo kernel_ex_task_info;
   EXPECT_NE(kernel_ex_task_info.InitTaskExtInfo(ext_info, op_desc), SUCCESS);
 }
+
+TEST_F(UtestKernelExTaskInfo, blocking_aicpu_op) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::TaskDef task_def;
+  domi::KernelExDef kernel_ex_def;
+  kernel_ex_def.set_kernel_ext_info(buf, len);
+  kernel_ex_def.set_kernel_ext_info_size(len);
+  domi::KernelExDef *kernel_ex_def_tmp = task_def.mutable_kernel_ex();
+  *kernel_ex_def_tmp = kernel_ex_def;
+
+  const OpDescPtr op_desc = CreateOpDesc("deque", "Deque");
+  ge::AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+
+  KernelExTaskInfo kernel_ex_task_info;
+  kernel_ex_task_info.op_desc_ = op_desc;
+  DavinciModel davinci_model(0, nullptr);
+  kernel_ex_task_info.davinci_model_ = &davinci_model;
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), SUCCESS);
+  EXPECT_EQ(kernel_ex_task_info.Distribute(), SUCCESS);
+  kernel_ex_task_info.op_desc_ = op_desc;
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), SUCCESS);
+  EXPECT_EQ(kernel_ex_task_info.Distribute(), SUCCESS);
+}
+
+TEST_F(UtestKernelExTaskInfo, blocking_aicpu_op_fail_01) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::TaskDef task_def;
+  domi::KernelExDef kernel_ex_def;
+  kernel_ex_def.set_kernel_ext_info(buf, len);
+  kernel_ex_def.set_kernel_ext_info_size(len);
+  domi::KernelExDef *kernel_ex_def_tmp = task_def.mutable_kernel_ex();
+  *kernel_ex_def_tmp = kernel_ex_def;
+
+  const OpDescPtr op_desc = CreateOpDesc("deque", "Deque");
+
+  KernelExTaskInfo kernel_ex_task_info;
+  kernel_ex_task_info.op_desc_ = op_desc;
+  DavinciModel davinci_model(0, nullptr);
+  kernel_ex_task_info.davinci_model_ = &davinci_model;
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), SUCCESS);
+
+  kernel_ex_task_info.is_blocking_aicpu_op_ = true;
+  EXPECT_EQ(kernel_ex_task_info.Distribute(), FAILED);
+}
+
+TEST_F(UtestKernelExTaskInfo, blocking_aicpu_op_fail_02) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::TaskDef task_def;
+  domi::KernelExDef kernel_ex_def;
+  kernel_ex_def.set_kernel_ext_info(buf, len);
+  kernel_ex_def.set_kernel_ext_info_size(len);
+  domi::KernelExDef *kernel_ex_def_tmp = task_def.mutable_kernel_ex();
+  *kernel_ex_def_tmp = kernel_ex_def;
+
+  const OpDescPtr op_desc = CreateOpDesc("deque", "Deque");
+  ge::AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+  KernelExTaskInfo kernel_ex_task_info;
+  kernel_ex_task_info.op_desc_ = op_desc;
+  DavinciModel davinci_model(0, nullptr);
+  kernel_ex_task_info.davinci_model_ = &davinci_model;
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_SUPPORT + 1);
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_ex_task_info.Distribute(), FAILED);
+
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_ex_task_info.Distribute(), FAILED);
+
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtEventReset, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_ex_task_info.Distribute(), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(kernel_ex_task_info.InitTaskExtInfo(kernel_ex_def.kernel_ext_info(), op_desc), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(kernel_ex_task_info.Distribute(), SUCCESS);
+}
+
 }  // namespace ge
diff --git a/tests/ut/ge/graph/load/kernel_task_info_unittest.cc b/tests/ut/ge/graph/load/kernel_task_info_unittest.cc
index 0c8da4b5..45ae7853 100644
--- a/tests/ut/ge/graph/load/kernel_task_info_unittest.cc
+++ b/tests/ut/ge/graph/load/kernel_task_info_unittest.cc
@@ -22,15 +22,20 @@
 #include "graph/load/model_manager/davinci_model.h"
 #include "graph/load/model_manager/task_info/kernel_task_info.h"
 #include "graph/load/model_manager/task_info/hccl_task_info.h"
+#include "tests/depends/runtime/src/runtime_stub.h"
 
 namespace ge {
 extern OpDescPtr CreateOpDesc(string name, string type);
 
 class UtestKernelTaskInfo : public testing::Test {
  protected:
-  void SetUp() {}
+  void SetUp() {
+    RTS_STUB_SETUP();
+  }
 
-  void TearDown() {}
+  void TearDown() {
+    RTS_STUB_TEARDOWN();
+  }
 };
 
 // test KernelTaskInfo Init.
@@ -1240,4 +1245,135 @@ TEST_F(UtestKernelTaskInfo, kernel_task_info_super_kernel_info) {
   EXPECT_EQ(kernel_task_info.SKTFinalize(), SUCCESS);
 }
 
+TEST_F(UtestKernelTaskInfo, blocking_aicpu_op) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::TaskDef task_def;
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+
+  const OpDescPtr op_desc = CreateOpDesc("deque", "Deque");
+  op_desc->SetId(0);
+  ge::AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+  DavinciModel davinci_model(0, nullptr);
+  davinci_model.op_list_.emplace(0, op_desc);
+
+  KernelTaskInfo kernel_task_info;
+  kernel_task_info.op_desc_ = op_desc;
+  kernel_task_info.davinci_model_ = &davinci_model;
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), SUCCESS);
+  EXPECT_EQ(kernel_task_info.Distribute(), SUCCESS);
+  kernel_task_info.op_desc_ = op_desc;
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), SUCCESS);
+  EXPECT_EQ(kernel_task_info.Distribute(), SUCCESS);
+}
+
+TEST_F(UtestKernelTaskInfo, blocking_aicpu_op_fail_01) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+
+  const OpDescPtr op_desc = CreateOpDesc("deque", "Deque");
+  op_desc->SetId(0);
+  DavinciModel davinci_model(0, nullptr);
+  davinci_model.op_list_.emplace(0, op_desc);
+
+  KernelTaskInfo kernel_task_info;
+  kernel_task_info.davinci_model_ = &davinci_model;
+  kernel_task_info.op_desc_ = op_desc;
+
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), SUCCESS);
+
+  kernel_task_info.is_blocking_aicpu_op_ = true;
+  EXPECT_EQ(kernel_task_info.Distribute(), FAILED);
+}
+
+TEST_F(UtestKernelTaskInfo, blocking_aicpu_op_fail_02) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+
+  const OpDescPtr op_desc = CreateOpDesc("deque", "Deque");
+  ge::AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+  op_desc->SetId(0);
+  DavinciModel davinci_model(0, nullptr);
+  davinci_model.op_list_.emplace(0, op_desc);
+
+  KernelTaskInfo kernel_task_info;
+  kernel_task_info.davinci_model_ = &davinci_model;
+  kernel_task_info.op_desc_ = op_desc;
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_SUPPORT + 1);
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_task_info.Distribute(), FAILED);
+
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_task_info.Distribute(), FAILED);
+
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtEventReset, rtError_t, 0x78000001);
+  EXPECT_EQ(kernel_task_info.Distribute(), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(kernel_task_info.InitAicpuTaskExtInfo(kernel_def.kernel_ext_info()), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(kernel_task_info.Distribute(), SUCCESS);
+}
+
 }  // namespace ge
diff --git a/tests/ut/ge/hybrid/node_executor/aicpu/aicpu_node_executor_unittest.cc b/tests/ut/ge/hybrid/node_executor/aicpu/aicpu_node_executor_unittest.cc
index b225949b..034b3f47 100644
--- a/tests/ut/ge/hybrid/node_executor/aicpu/aicpu_node_executor_unittest.cc
+++ b/tests/ut/ge/hybrid/node_executor/aicpu/aicpu_node_executor_unittest.cc
@@ -27,7 +27,7 @@
 #include "hybrid/node_executor/aicpu/aicpu_node_executor.h"
 #undef protected
 #undef private
-
+#include "tests/depends/runtime/src/runtime_stub.h"
 using namespace std;
 using namespace testing;
 
@@ -43,8 +43,12 @@ using namespace hybrid;
 
 class UtestAicpuNodeExecutor : public testing::Test {
  protected:
-  void SetUp() {}
-  void TearDown() {}
+  void SetUp() {
+    RTS_STUB_SETUP();
+  }
+  void TearDown() {
+    RTS_STUB_TEARDOWN();
+  }
 };
 
 static NodePtr CreateNode(ComputeGraphPtr graph, const string &name, const string &type, int in_num, int out_num) {
@@ -164,5 +168,222 @@ TEST_F(UtestAicpuNodeExecutor, aicpu_tf_node_task) {
 
 }
 
+TEST_F(UtestAicpuNodeExecutor, aicpu_blocking_node_task) {
+  ComputeGraphPtr graph = std::make_shared<ComputeGraph>("test");
+  GeRootModelPtr ge_root_model = std::make_shared<GeRootModel>(graph);
+  ge_root_model->SetModelName("test_name");
+  HybridModel hybrid_model(ge_root_model);
+
+  NodePtr node = CreateNode(graph, "deque", FRAMEWORK_OP_TYPE, 1, 1);
+  ge::AttrUtils::SetBool(node->GetOpDesc(), ATTR_NAME_IS_BLOCKING_OP, true);
+  std::unique_ptr<NodeItem> new_node;
+  ASSERT_EQ(NodeItem::Create(node, new_node), SUCCESS);
+  NodeItem *node_item = new_node.get();
+  node_item->input_start = 0;
+  node_item->output_start = 0;
+  node_item->is_dynamic = true;
+  node_item->shape_inference_type = DEPEND_SHAPE_RANGE;
+
+  GraphItem graph_item;
+  graph_item.node_items_.emplace_back(node_item);
+  graph_item.total_inputs_ = 1;
+  graph_item.total_outputs_ = 1;
+
+  GraphExecutionContext graph_execution_context;
+  SubgraphContext subgraph_context(&graph_item, &graph_execution_context);
+  ASSERT_EQ(subgraph_context.Init(), SUCCESS);
+  graph_execution_context.callback_manager = std::unique_ptr<CallbackManager>(new CallbackManager());
+
+  auto node_state = subgraph_context.GetOrCreateNodeState(node_item);
+  ASSERT_NE(node_state, nullptr);
+
+  uint64_t value_0 = 512;
+  TensorValue in_tensor0(&value_0, sizeof(value_0));
+  subgraph_context.SetInput(*node_item, 0, in_tensor0);
+
+  TensorValue out_tensor0(&value_0, sizeof(value_0));
+  subgraph_context.SetOutput(*node_item, 0, out_tensor0);
+
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+  domi::TaskDef task_def;
+
+  AicpuTaskStruct args;
+  args.head.length = sizeof(args);
+  args.head.ioAddrNum = 2;
+
+  kernel_def.set_args(reinterpret_cast<const char *>(&args), args.head.length);
+  kernel_def.set_args_size(args.head.length);
+  domi::KernelDef *kernel_def_tmp = task_def.mutable_kernel();
+  *kernel_def_tmp = kernel_def;
+
+  AicpuNodeTask aicpu_node_task(node_item, task_def);
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), SUCCESS);
+  ASSERT_EQ(aicpu_node_task.LaunchTask(*node_state->GetTaskContext()), SUCCESS);
+
+  node_item->shape_inference_type = DEPEND_COMPUTE;
+  domi::KernelExDef kernel_ex_def;
+  kernel_ex_def.set_kernel_ext_info(buf, len);
+  kernel_ex_def.set_kernel_ext_info_size(len);
+  kernel_ex_def.set_args(reinterpret_cast<const char *>(&args), args.head.length);
+  kernel_ex_def.set_args_size(args.head.length);
+  domi::KernelExDef *kernel_ex_def_tmp = task_def.mutable_kernel_ex();
+  *kernel_ex_def_tmp = kernel_ex_def;
+  hybrid_model.task_defs_[node] = std::vector<domi::TaskDef>({task_def, task_def});
+
+  AicpuTfNodeTask aicpu_tf_node_task(node_item, task_def);
+  ASSERT_EQ(aicpu_tf_node_task.Init(hybrid_model), SUCCESS);
+  ASSERT_EQ(aicpu_tf_node_task.LaunchTask(*node_state->GetTaskContext()), SUCCESS);
+}
+
+TEST_F(UtestAicpuNodeExecutor, aicpu_blocking_node_task_fail) {
+  ComputeGraphPtr graph = std::make_shared<ComputeGraph>("test");
+  GeRootModelPtr ge_root_model = std::make_shared<GeRootModel>(graph);
+  ge_root_model->SetModelName("test_name");
+  HybridModel hybrid_model(ge_root_model);
+
+  NodePtr node = CreateNode(graph, "deque", FRAMEWORK_OP_TYPE, 1, 1);
+  ge::AttrUtils::SetBool(node->GetOpDesc(), ATTR_NAME_IS_BLOCKING_OP, true);
+  std::unique_ptr<NodeItem> new_node;
+  ASSERT_EQ(NodeItem::Create(node, new_node), SUCCESS);
+  NodeItem *node_item = new_node.get();
+  node_item->input_start = 0;
+  node_item->output_start = 0;
+  node_item->is_dynamic = true;
+  node_item->shape_inference_type = DEPEND_SHAPE_RANGE;
+
+  GraphItem graph_item;
+  graph_item.node_items_.emplace_back(node_item);
+  graph_item.total_inputs_ = 1;
+  graph_item.total_outputs_ = 1;
+
+  GraphExecutionContext graph_execution_context;
+  SubgraphContext subgraph_context(&graph_item, &graph_execution_context);
+  ASSERT_EQ(subgraph_context.Init(), SUCCESS);
+  graph_execution_context.callback_manager = std::unique_ptr<CallbackManager>(new CallbackManager());
+
+  auto node_state = subgraph_context.GetOrCreateNodeState(node_item);
+  ASSERT_NE(node_state, nullptr);
+
+  uint64_t value_0 = 512;
+  TensorValue in_tensor0(&value_0, sizeof(value_0));
+  subgraph_context.SetInput(*node_item, 0, in_tensor0);
+
+  TensorValue out_tensor0(&value_0, sizeof(value_0));
+  subgraph_context.SetOutput(*node_item, 0, out_tensor0);
+
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+  domi::TaskDef task_def;
+
+  AicpuTaskStruct args;
+  args.head.length = sizeof(args);
+  args.head.ioAddrNum = 2;
+
+  kernel_def.set_args(reinterpret_cast<const char *>(&args), args.head.length);
+  kernel_def.set_args_size(args.head.length);
+  domi::KernelDef *kernel_def_tmp = task_def.mutable_kernel();
+  *kernel_def_tmp = kernel_def;
+
+  AicpuNodeTask aicpu_node_task(node_item, task_def);
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_SUPPORT + 1);
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_node_task.LaunchTask(*node_state->GetTaskContext()), FAILED);
+
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_node_task.LaunchTask(*node_state->GetTaskContext()), FAILED);
+
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtEventReset, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_node_task.LaunchTask(*node_state->GetTaskContext()), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  ASSERT_EQ(aicpu_node_task.Init(hybrid_model), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  ASSERT_EQ(aicpu_node_task.LaunchTask(*node_state->GetTaskContext()), SUCCESS);
+
+  node_item->shape_inference_type = DEPEND_COMPUTE;
+  domi::KernelExDef kernel_ex_def;
+  kernel_ex_def.set_kernel_ext_info(buf, len);
+  kernel_ex_def.set_kernel_ext_info_size(len);
+  kernel_ex_def.set_args(reinterpret_cast<const char *>(&args), args.head.length);
+  kernel_ex_def.set_args_size(args.head.length);
+  domi::KernelExDef *kernel_ex_def_tmp = task_def.mutable_kernel_ex();
+  *kernel_ex_def_tmp = kernel_ex_def;
+  hybrid_model.task_defs_[node] = std::vector<domi::TaskDef>({task_def, task_def});
+
+  AicpuTfNodeTask aicpu_tf_node_task(node_item, task_def);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_tf_node_task.Init(hybrid_model), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_tf_node_task.Init(hybrid_model), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_SUPPORT + 1);
+  ASSERT_EQ(aicpu_tf_node_task.Init(hybrid_model), FAILED);
+
+  ASSERT_EQ(aicpu_tf_node_task.Init(hybrid_model), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_tf_node_task.LaunchTask(*node_state->GetTaskContext()), FAILED);
+
+  ASSERT_EQ(aicpu_tf_node_task.Init(hybrid_model), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtEventReset, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_tf_node_task.LaunchTask(*node_state->GetTaskContext()), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(aicpu_tf_node_task.Init(hybrid_model), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(aicpu_tf_node_task.LaunchTask(*node_state->GetTaskContext()), SUCCESS);
+}
 }  // namespace ge
 
diff --git a/tests/ut/ge/single_op/single_op_task_unittest.cc b/tests/ut/ge/single_op/single_op_task_unittest.cc
index 8964df74..52091856 100644
--- a/tests/ut/ge/single_op/single_op_task_unittest.cc
+++ b/tests/ut/ge/single_op/single_op_task_unittest.cc
@@ -19,6 +19,7 @@
 
 #include "graph/load/model_manager/model_utils.h"
 #include "graph/utils/graph_utils.h"
+#include "hybrid/node_executor/aicpu/aicpu_ext_info.h"
 #include "runtime/rt.h"
 
 #define protected public
@@ -30,6 +31,7 @@
 #include "external/register/op_tiling_registry.h"
 #undef private
 #undef protected
+#include "tests/depends/runtime/src/runtime_stub.h"
 
 using namespace std;
 using namespace testing;
@@ -38,9 +40,13 @@ using namespace optiling;
 
 class UtestSingleOpTask : public testing::Test {
  protected:
-  void SetUp() {}
+  void SetUp() {
+    RTS_STUB_SETUP();
+  }
 
-  void TearDown() {}
+  void TearDown() {
+    RTS_STUB_TEARDOWN();
+  }
 };
 
 TEST_F(UtestSingleOpTask, test_build_kernel_task) {
@@ -237,3 +243,124 @@ TEST_F(UtestSingleOpTask, test_aicpu_task_update_io_addr) {
     ASSERT_EQ(ret, PARAM_INVALID);
   }
 }
+
+TEST_F(UtestSingleOpTask, test_blocking_aicpu_op_01) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+
+  auto op_desc = make_shared<OpDesc>("deque", "Deque");
+  ge::AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+  AiCpuCCTask aicpu_task;
+  aicpu_task.SetOpDesc(op_desc);
+  rtStream_t stream;
+  ASSERT_EQ(rtStreamCreate(&stream, 0), RT_ERROR_NONE);
+
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), SUCCESS);
+  ASSERT_EQ(aicpu_task.LaunchKernel(stream), SUCCESS);
+}
+
+TEST_F(UtestSingleOpTask, test_blocking_aicpu_op_02) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+
+  auto op_desc = make_shared<OpDesc>("deque", "Deque");
+  ge::AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+  AiCpuTask aicpu_task;
+  aicpu_task.SetOpDesc(op_desc);
+  rtStream_t stream;
+  ASSERT_EQ(rtStreamCreate(&stream, 0), RT_ERROR_NONE);
+
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), SUCCESS);
+  ASSERT_EQ(aicpu_task.LaunchKernel(stream), SUCCESS);
+}
+
+TEST_F(UtestSingleOpTask, test_blocking_aicpu_op_fail) {
+  int len = sizeof(hybrid::AicpuExtInfo) + sizeof(hybrid::AsyncWaitInfo);
+  vector<char> aicpu_ext_info(len, 0);
+  char *buf = aicpu_ext_info.data();
+  int offset = 0;
+  hybrid::AicpuExtInfo *ext_info = reinterpret_cast<hybrid::AicpuExtInfo*>(buf + offset);
+  ext_info->infoType = aicpu::FWKAdapter::FWK_ADPT_EXT_ASYNCWAIT;
+  ext_info->infoLen = sizeof(hybrid::AsyncWaitInfo);
+  offset += sizeof(hybrid::AicpuExtInfo);
+  hybrid::AsyncWaitInfo *async_wait_info = reinterpret_cast<hybrid::AsyncWaitInfo*>(buf + offset);
+  async_wait_info->waitType = 0;
+  async_wait_info->waitId = 0;
+  async_wait_info->timeOut = 0;
+  async_wait_info->reserved = 0;
+
+  domi::KernelDef kernel_def;
+  kernel_def.set_kernel_ext_info(buf, len);
+  kernel_def.set_kernel_ext_info_size(len);
+
+  auto op_desc = make_shared<OpDesc>("deque", "Deque");
+  ge::AttrUtils::SetBool(op_desc, ATTR_NAME_IS_BLOCKING_OP, true);
+  AiCpuTask aicpu_task;
+  aicpu_task.SetOpDesc(op_desc);
+  rtStream_t stream;
+  ASSERT_EQ(rtStreamCreate(&stream, 0), RT_ERROR_NONE);
+
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), SUCCESS);
+  ASSERT_EQ(aicpu_task.LaunchKernel(stream), SUCCESS);
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_SUPPORT + 1);
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDevice, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_task.LaunchKernel(stream), FAILED);
+
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtStreamWaitEvent, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_task.LaunchKernel(stream), FAILED);
+
+  ASSERT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtEventReset, rtError_t, 0x78000001);
+  ASSERT_EQ(aicpu_task.LaunchKernel(stream), FAILED);
+
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(aicpu_task.SetExtInfoAndType(kernel_def.kernel_ext_info(), 0), SUCCESS);
+  RTS_STUB_RETURN_VALUE(rtGetDeviceCapability, rtError_t, RT_ERROR_NONE);
+  RTS_STUB_OUTBOUND_VALUE(rtGetDeviceCapability, int32_t, value, RT_AICPU_BLOCKING_OP_NOT_SUPPORT);
+  EXPECT_EQ(aicpu_task.LaunchKernel(stream), SUCCESS);
+}
diff --git a/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h b/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h
index df57c82e..5733d68f 100644
--- a/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h
+++ b/third_party/fwkacllib/inc/cce/fwk_adpt_struct.h
@@ -62,6 +62,7 @@ enum FWKTaskExtInfoType {
   FWK_ADPT_EXT_SESSION_INFO,
   FWK_ADPT_EXT_BITMAP,
   FWK_ADPT_EXT_TOPIC_TYPE,
+  FWK_ADPT_EXT_ASYNCWAIT,
   FWK_ADPT_EXT_INVALID
 };
 
@@ -80,6 +81,12 @@ enum FWKExtUpdateAddrType {
   FWK_ADPT_UPDATE_INPUT_OUTPUT
 };
 
+enum FWKExtWaitType {
+  FWK_ADPT_WAIT_TYPE_NULL = 0,
+  FWK_ADPT_WAIT_TYPE_EVENT,
+  FWK_ADPT_WAIT_TYPE_INVALID
+};
+
 #pragma pack(push, 1)
 // API Parameter Structure
 struct StrFWKKernel {
@@ -133,6 +140,15 @@ struct ResultSummary {
   uint64_t raw_data_size;    // size of raw data
 };
 #pragma pack(pop)
+
+#pragma pack(push, 1)
+struct AsyncWait {
+  uint8_t waitType; // wait type, FWK_ADPT_WAIT_TYPE_EVENT: event wait
+  uint32_t waitId; // wait id, GE refresh
+  uint32_t timeOut; // reserved
+  uint64_t reserved;
+};
+#pragma pack(pop)
 }  // end  namespace FWKAdapter
 }  // namespace aicpu
 
diff --git a/third_party/fwkacllib/inc/runtime/config.h b/third_party/fwkacllib/inc/runtime/config.h
index c1327c45..a244c793 100644
--- a/third_party/fwkacllib/inc/runtime/config.h
+++ b/third_party/fwkacllib/inc/runtime/config.h
@@ -52,6 +52,14 @@ typedef enum tagRtAicpuScheType {
     SCHEDULE_HARDWARE, /* HWTS Schedule */
 } rtAicpuScheType;
 
+typedef enum tagRtDeviceCapabilityType {
+  RT_SCHEDULE_SOFTWARE = 0, // SoftWare Schedule
+  RT_SCHEDULE_SOFTWARE_OPT,
+  RT_SCHEDULE_HARDWARE, // HWTS Schedule
+  RT_AICPU_BLOCKING_OP_NOT_SUPPORT,
+  RT_AICPU_BLOCKING_OP_SUPPORT, // 1910/1980/1951 ts support AICPU blocking operation
+} rtDeviceCapabilityType;
+
 typedef enum tagRtVersion {
     VER_BEGIN = 0,
     VER_NA = VER_BEGIN,
diff --git a/third_party/fwkacllib/inc/runtime/dev.h b/third_party/fwkacllib/inc/runtime/dev.h
index 2cf6712f..18d837eb 100644
--- a/third_party/fwkacllib/inc/runtime/dev.h
+++ b/third_party/fwkacllib/inc/runtime/dev.h
@@ -65,6 +65,7 @@ typedef enum tagRtFeatureType {
 
 typedef enum tagRtDeviceFeatureType {
   FEATURE_TYPE_SCHE,
+  FEATURE_TYPE_BLOCKING_OPERATOR,
   FEATURE_TYPE_END,
 } rtDeviceFeatureType_t;
 
@@ -78,6 +79,17 @@ typedef enum tagMemoryInfo {
     MEMORY_INFO_RSV
 } rtMemoryInfo_t;
 
+typedef enum tagRtDeviceModuleType {
+  RT_MODULE_TYPE_SYSTEM = 0,
+  RT_MODULE_TYPE_AICPU,
+  RT_MODULE_TYPE_CCPU,
+  RT_MODULE_TYPE_DCPU,
+  RT_MODULE_TYPE_AICORE,
+  RT_MODULE_TYPE_TSCPU,
+  RT_MODULE_TYPE_PCIE,
+  RT_MODULE_TYPE_VECTOR_CORE
+} tagRtDeviceModuleType_t;
+
 /**
  * @ingroup dvrt_dev
  * @brief get total device number.