fix

ge/hybrid/node_executor/aicpu/aicpu_node_executor.cc

@@ -441,8 +441,8 @@ Status AicpuTfNodeTask::EnsureSessionCreated(uint64_t session_id) {
   return SUCCESS;
 }
 
-Status AicpuTfNodeTask::ReadResultSummaryAndPrepareMemory(TaskContext &context,
-                                                          std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
+Status AicpuNodeTaskBase::ReadResultSummaryAndPrepareMemory(TaskContext &context,
+                                                            std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
   for (auto i = 0; i < node_item_->num_outputs; ++i) {
     auto &result_summary = output_summary_host_[i];
     GE_CHK_RT_RET(rtMemcpy(&result_summary, sizeof(aicpu::FWKAdapter::ResultSummary),
@@ -467,6 +467,31 @@ Status AicpuTfNodeTask::ReadResultSummaryAndPrepareMemory(TaskContext &context,
   return SUCCESS;
 }
 
+Status AicpuNodeTask::CopyDataToHbm(TaskContext &context,
+                                    const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
+  GE_CHK_BOOL_RET_STATUS(out_shape_hbm.size() == static_cast<std::size_t>(node_item_->num_outputs),
+                         INTERNAL_ERROR,
+                         "[Check][Size]Node[%s] has %d outputs but out shape is %zu not equal.",
+                         node_name_.c_str(), node_item_->num_outputs, out_shape_hbm.size());
+
+  GE_CHK_STATUS_RET_NOLOG(PrepareCopyInputs(context, out_shape_hbm));
+
+  RECORD_CALLBACK_EVENT(context.GetExecutionContext(), node_name_.c_str(), "[LaunchCopy] Start");
+  uint32_t flag = RT_KERNEL_DEFAULT;
+  auto rt_ret = rtCpuKernelLaunchWithFlag(reinterpret_cast<const void *>(memcpy_so_name_.c_str()),
+                                          reinterpret_cast<const void *>(memcpy_kernel_name_.c_str()),
+                                          1, // default core dim is 1
+                                          memcpy_args_.get(), memcpy_args_size_,
+                                          nullptr, context.GetStream(), flag);
+  GE_CHK_RT_RET(rt_ret);
+
+  RECORD_CALLBACK_EVENT(context.GetExecutionContext(), node_name_.c_str(), "[LaunchCopy] End");
+
+  GE_CHK_RT_RET(rtStreamSynchronize(context.GetStream()));
+  RECORD_CALLBACK_EVENT(context.GetExecutionContext(), node_name_.c_str(), "[SynchronizeCopy] End");
+  return SUCCESS;
+}
+
 Status AicpuTfNodeTask::CopyDataToHbm(TaskContext &context,
                                       const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
   GE_CHK_BOOL_RET_STATUS(out_shape_hbm.size() == static_cast<std::size_t>(node_item_->num_outputs),
@@ -486,8 +511,8 @@ Status AicpuTfNodeTask::CopyDataToHbm(TaskContext &context,
   return SUCCESS;
 }
 
-Status AicpuTfNodeTask::PrepareCopyInputs(const TaskContext &context,
-                                          const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
+Status AicpuNodeTaskBase::PrepareCopyInputs(const TaskContext &context,
+                                            const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
   std::vector<uint64_t> copy_input_release_flag;
   std::vector<uint64_t> copy_input_data_size;
   std::vector<uint64_t> copy_input_src;
@@ -528,8 +553,8 @@ Status AicpuTfNodeTask::PrepareCopyInputs(const TaskContext &context,
   return SUCCESS;
 }
 
-Status AicpuTfNodeTask::UpdateShapeByHbmBuffer(TaskContext &context,
-                                               const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
+Status AicpuNodeTaskBase::UpdateShapeByHbmBuffer(TaskContext &context,
+                                                 const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) {
   GE_CHK_BOOL_RET_STATUS(out_shape_hbm.size() == static_cast<std::size_t>(node_item_->num_outputs),
                          INTERNAL_ERROR,
                          "Node[%s] has %d outputs but out shape is %zu",
@@ -560,7 +585,7 @@ Status AicpuTfNodeTask::UpdateShapeByHbmBuffer(TaskContext &context,
   return SUCCESS;
 }
 
-Status AicpuTfNodeTask::UpdateShapeAndDataByResultSummary(TaskContext &context) {
+Status AicpuNodeTaskBase::UpdateShapeAndDataByResultSummary(TaskContext &context) {
   GELOGD("Node[%s] update shape and data by result summary begin.", node_name_.c_str());
 
   std::vector<std::unique_ptr<TensorBuffer>> out_shape_hbm;
@@ -649,7 +674,7 @@ Status AicpuTfNodeTask::LaunchTask(TaskContext &context) {
   return SUCCESS;
 }
 
-Status AicpuTfNodeTask::TaskCallback(TaskContext &context) {
+Status AicpuNodeTaskBase::TaskCallback(TaskContext &context) {
   GELOGD("Node[%s] task callback start. is_dynamic=%s, unknown_type=%d.",
          node_name_.c_str(), node_item_->is_dynamic ? "true" : "false", unknown_type_);
   Status callback_ret = SUCCESS;
@@ -666,14 +691,110 @@ Status AicpuTfNodeTask::TaskCallback(TaskContext &context) {
   return callback_ret;
 }
 
+Status AicpuNodeTask::SetMemCopyTask(const domi::TaskDef &task_def) {
+  if (node_item_->num_outputs == 0) {
+    GELOGD("Node[%s] type[%s] has no output, no need set mem_copy task.",
+           node_name_.c_str(), node_item_->node_type.c_str());
+    return SUCCESS;
+  }
+
+  GELOGD("Start to set memcpy task for node[%s].", node_name_.c_str());
+  const domi::KernelDef &kernel_def = task_def.kernel();
+  auto &memcpy_args = kernel_def.args();
+  memcpy_args_size_ = kernel_def.args_size();
+  memcpy_so_name_ = kernel_def.so_name();
+  memcpy_kernel_name_ = kernel_def.kernel_name();
+  GE_IF_BOOL_EXEC(memcpy_args.size() != memcpy_args_size_,
+                  REPORT_INNER_ERROR("E19999", "MemCopy task def args.size=%zu, but args_size=%u not equal.",
+                                     memcpy_args.size(), memcpy_args_size_);
+                  GELOGE(FAILED, "[Check][Size]MemCopy task def args.size=%zu, but args_size=%u not equal.",
+                         memcpy_args.size(), memcpy_args_size_);
+                  return FAILED;);
+  GE_IF_BOOL_EXEC(memcpy_args_size_ < sizeof(aicpu::AicpuParamHead),
+                  REPORT_INNER_ERROR("E19999",
+                                     "Task def args_size=%u is less than aicpu param head len=%zu.",
+                                     memcpy_args_size_, sizeof(aicpu::AicpuParamHead));
+                  GELOGE(FAILED,
+                         "[Check][Size] Task def args_size=%u is less than aicpu param head len=%zu.",
+                         memcpy_args_size_, sizeof(aicpu::AicpuParamHead));
+                  return FAILED;);
+
+  memcpy_args_.reset(new(std::nothrow) uint8_t[memcpy_args_size_]());
+  GE_IF_BOOL_EXEC(memcpy_args_ == nullptr,
+                  REPORT_INNER_ERROR("E19999", "new memory failed for Node[MemCopy], task_size[%u].",
+                                     memcpy_args_size_);
+                  GELOGE(FAILED, "[Malloc][Memory] failed for Node[MemCopy], task_size[%u].",
+                         memcpy_args_size_);
+                  return FAILED;);
+
+  errno_t sec_ret = memcpy_s(memcpy_args_.get(), memcpy_args_size_, memcpy_args.c_str(), memcpy_args.size());
+  GE_IF_BOOL_EXEC(sec_ret != EOK,
+                  REPORT_INNER_ERROR("E19999",
+                                     "memcpy_s argc_ failed for Node[MemCopy], ret: %d", sec_ret);
+                  GELOGE(INTERNAL_ERROR,
+                         "[Update][args] failed for Node[MemCopy], ret: %d", sec_ret);
+                  return sec_ret;);
+  auto memcpy_param_head = reinterpret_cast<aicpu::AicpuParamHead *>(memcpy_args_.get());
+  uint32_t memcpy_io_num = memcpy_param_head->ioAddrNum;
+  auto memcpy_io_addr = memcpy_args_.get() + sizeof(aicpu::AicpuParamHead);
+  // if has input and output, need copy to ioaddr
+  int cpy_ret = memcpy_s(memcpy_io_addr, memcpy_args_size_ - sizeof(aicpu::AicpuParamHead),
+                         &copy_io_addr_[0], sizeof(uint64_t) * memcpy_io_num);
+  GE_IF_BOOL_EXEC(cpy_ret != 0,
+                  REPORT_INNER_ERROR("E19999", "Node[Memcpoy] memcpy io addr to AicpuParamHead failed,"
+                         "ret=%d, args_size=%u, io nums=%u.",
+                         cpy_ret, memcpy_args_size_, memcpy_io_num);
+                  GELOGE(INTERNAL_ERROR, "[Update][io_addr]Node[MemCopy] memcpy io addr to AicpuParamHead failed,"
+                         "ret=%d, args_size=%u, io nums=%u.",
+                         cpy_ret, memcpy_args_size_, memcpy_io_num);
+                  return INTERNAL_ERROR;);
+  GELOGD("Set memcpy task for node[MemCopy] successfully.");
+  return SUCCESS;
+}
+
+Status AicpuNodeTask::InitForDependComputeTask() {
+  if ((unknown_type_ != DEPEND_COMPUTE) || (node_item_->num_outputs == 0)) {
+    GELOGD("Node[%s] type[%s] unknown_type is %d, output num is %d.",
+           node_name_.c_str(), node_item_->node_type.c_str(), unknown_type_, node_item_->num_outputs);
+    return SUCCESS;
+  }
+
+  output_summary_.resize(node_item_->num_outputs);
+  constexpr auto result_summary_size = sizeof(aicpu::FWKAdapter::ResultSummary);
+  for (auto i = 0; i < node_item_->num_outputs; ++i) {
+    GE_CHK_STATUS_RET(AllocTensorBuffer(result_summary_size, output_summary_[i]),
+                      "[Alloc][TensorBuffer] failed for Node[%s] to copy result summary info, size=%zu.",
+                      node_name_.c_str(), result_summary_size);
+  }
+  output_summary_host_.resize(node_item_->num_outputs);
+
+  // init for mem copy task
+  // copy task need copy output_data and output_shape, max len is 2 * output_num
+  const size_t copy_input_buf_len = node_item_->num_outputs * 2 * sizeof(uint64_t);
+  GE_CHK_STATUS_RET(AllocTensorBuffer(copy_input_buf_len, copy_input_release_flag_dev_),
+                    "[Alloc][TensorBuffer] failed for Node[%s] to copy task input release_flag, size=%zu",
+                    node_name_.c_str(), copy_input_buf_len);
+  GE_CHK_STATUS_RET(AllocTensorBuffer(copy_input_buf_len, copy_input_data_size_dev_),
+                    "[Alloc][TensorBuffer] failed for Node[%s] to copy task input data_size, size=%zu",
+                    node_name_.c_str(), copy_input_buf_len);
+  GE_CHK_STATUS_RET(AllocTensorBuffer(copy_input_buf_len, copy_input_src_dev_),
+                    "[Alloc][TensorBuffer] failed for Node[%s] to copy task input src, size=%zu",
+                    node_name_.c_str(), copy_input_buf_len);
+  GE_CHK_STATUS_RET(AllocTensorBuffer(copy_input_buf_len, copy_input_dst_dev_),
+                    "[Alloc][TensorBuffer] failed for Node[%s] to copy task input dst, size=%zu",
+                    node_name_.c_str(), copy_input_buf_len);
+
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_release_flag_dev_->GetData()));
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_data_size_dev_->GetData()));
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_src_dev_->GetData()));
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_dst_dev_->GetData()));
+  return SUCCESS;
+}
+
 Status AicpuNodeTask::Init(const HybridModel &model) {
   auto node_name = node_name_;
   GELOGD("Node[%s] init start.", node_name.c_str());
 
-  GE_CHK_BOOL_RET_STATUS(unknown_type_ != DEPEND_COMPUTE, FAILED,
-                         "[Check][Type]Node[%s] unknown type[%d] is depend compute, it's not supported now.",
-                         node_name.c_str(), unknown_type_);
-
   GE_CHK_BOOL_RET_STATUS(task_def_.has_kernel(), FAILED,
                          "[Check][task_def_]Node[%s] task def does not has kernel.", node_name.c_str());
   auto &kernel_def = task_def_.kernel();
@@ -762,7 +883,10 @@ Status AicpuNodeTask::Init(const HybridModel &model) {
   uint64_t ext_session_id = model.GetSessionId();
   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());
   if (ext_info_addr_dev_ == nullptr) {
     aicpu_param_head->extInfoLength = 0;
     aicpu_param_head->extInfoAddr = 0;
@@ -770,7 +894,11 @@ Status AicpuNodeTask::Init(const HybridModel &model) {
     aicpu_param_head->extInfoLength = ext_info_addr_dev_->GetSize();
     aicpu_param_head->extInfoAddr = reinterpret_cast<uintptr_t>(ext_info_addr_dev_->GetData());
   }
-
+  auto task_defs = model.GetTaskDefs(node_item_->node);
+  GE_CHECK_NOTNULL(task_defs);
+  if (unknown_type_ == DEPEND_COMPUTE) {
+    GE_CHK_STATUS_RET_NOLOG(SetMemCopyTask((*task_defs)[1]));
+  }
   GELOGD("Node[%s] init end.", node_name.c_str());
   return SUCCESS;
 }
@@ -786,13 +914,29 @@ Status AicpuNodeTask::UpdateIoAddr(TaskContext &context) {
     io_addrs.emplace_back(reinterpret_cast<uintptr_t>(inputData->GetData()));
   }
 
-  GE_CHK_STATUS_RET_NOLOG(context.AllocateOutputs());
-  for (auto j = 0; j < node_item_->num_outputs; ++j) {
-    auto outputData = context.GetOutput(j);
-    GE_CHECK_NOTNULL(outputData);
-    GELOGD("Node[%s] output[%d] addr = %p, size = %zu", node_name_.c_str(), j,
-           outputData->GetData(), outputData->GetSize());
-    io_addrs.emplace_back(reinterpret_cast<uintptr_t>(outputData->GetData()));
+  // known shape or not depend compute
+  if (!node_item_->is_dynamic || unknown_type_ != DEPEND_COMPUTE) {
+    // unknown type 4 do this in call back.
+    GE_CHK_STATUS_RET_NOLOG(context.AllocateOutputs());
+    for (auto j = 0; j < node_item_->num_outputs; ++j) {
+      auto outputData = context.GetOutput(j);
+      GE_CHECK_NOTNULL(outputData);
+      GELOGD("Node[%s] output[%d] addr = %p, size = %zu",
+             node_name_.c_str(), j, outputData->GetData(), outputData->GetSize());
+      io_addrs.emplace_back(reinterpret_cast<uintptr_t>(outputData->GetData()));
+    }
+  } else {
+    // unknown type 4 use result summary update ioaddr.
+    GELOGD("Node[%s] is depend compute node, use result summary as out addr.", node_name_.c_str());
+    GE_CHK_BOOL_RET_STATUS(output_summary_.size() == static_cast<std::size_t>(node_item_->num_outputs),
+                           INTERNAL_ERROR,
+                           "[Check][Size]Node[%s] has %d output but %zu output summary not equal.",
+                           node_name_.c_str(), node_item_->num_outputs, output_summary_.size());
+
+    for (auto j = 0; j < node_item_->num_outputs; ++j) {
+      void *summary_addr = output_summary_[j]->GetData();
+      io_addrs.emplace_back(reinterpret_cast<uintptr_t>(summary_addr));
+    }
   }
 
   auto io_addr = args_.get() + sizeof(aicpu::AicpuParamHead);
@@ -830,23 +974,6 @@ Status AicpuNodeTask::LaunchTask(TaskContext &context) {
   return SUCCESS;
 }
 
-Status AicpuNodeTask::TaskCallback(TaskContext &context) {
-  GELOGD("Node[%s] task callback start, is_dynamic = %s, unknown_type=%d.",
-         node_name_.c_str(), node_item_->is_dynamic ? "true" : "false", unknown_type_);
-  Status callback_ret = SUCCESS;
-
-  // check need update shape, call update shape.
-  if (node_item_->is_dynamic && unknown_type_ == DEPEND_SHAPE_RANGE) {
-    // check result
-    callback_ret = UpdateOutputShapeFromExtInfo(context);
-  } else {
-    GELOGD("Node[%s] unknown shape type is %d no need update output shape.",
-           node_name_.c_str(), unknown_type_);
-  }
-  GELOGD("Node[%s] task callback end.", node_name_.c_str());
-  return callback_ret;
-}
-
 Status AiCpuNodeExecutor::PrepareTask(NodeTask &task, TaskContext &context) const {
   // malloc HBM memory at Init, here just update them
   RECORD_EXECUTION_EVENT(context.GetExecutionContext(), context.GetNodeName(), "[AiCpuNodeExecutorPrepareTask] Start");

ge/hybrid/node_executor/aicpu/aicpu_node_executor.h

@@ -55,12 +55,34 @@ class AicpuNodeTaskBase : public NodeTask {
 
   virtual Status LaunchTask(TaskContext &context) = 0;
 
-  virtual Status TaskCallback(TaskContext &context) = 0;
+  virtual Status InitForDependComputeTask() = 0;
+
+  Status TaskCallback(TaskContext &context);
+
+  virtual Status UpdateShapeAndDataByResultSummary(TaskContext &context);
 
   virtual Status UpdateIoAddr(TaskContext &context) = 0;
 
   static Status AllocTensorBuffer(size_t size, std::unique_ptr<TensorBuffer> &tensor_buffer);
 
+  virtual Status CopyDataToHbm(TaskContext &context,
+                               const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) = 0;
+
+  ///
+  /// read result summary and prepare copy task memory.
+  /// @param context task context
+  /// @param out_shape_hbm if scalar, TensorBuffer->data is null, size=0
+  /// @return SUCCESS:success other:failed
+  ///
+  Status ReadResultSummaryAndPrepareMemory(TaskContext &context,
+                                           std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm);
+
+  Status UpdateShapeByHbmBuffer(TaskContext &context,
+                                const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm);
+
+  Status PrepareCopyInputs(const TaskContext &context,
+                           const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm);
+
  protected:
   const NodeItem *node_item_;
   // just reference.
@@ -78,6 +100,15 @@ class AicpuNodeTaskBase : public NodeTask {
 
   // ext info addr, device mem
   std::unique_ptr<TensorBuffer> ext_info_addr_dev_;
+
+  std::vector<std::unique_ptr<TensorBuffer>> output_summary_;
+  std::vector<aicpu::FWKAdapter::ResultSummary> output_summary_host_;
+
+  std::unique_ptr<TensorBuffer> copy_input_release_flag_dev_;
+  std::unique_ptr<TensorBuffer> copy_input_data_size_dev_;
+  std::unique_ptr<TensorBuffer> copy_input_src_dev_;
+  std::unique_ptr<TensorBuffer> copy_input_dst_dev_;
+  bool need_sync_ = false;
 };
 
 class AicpuTfNodeTask : public AicpuNodeTaskBase {
@@ -93,33 +124,15 @@ class AicpuTfNodeTask : public AicpuNodeTaskBase {
 
   Status LaunchTask(TaskContext &context) override;
 
-  Status TaskCallback(TaskContext &context) override;
-
   Status UpdateIoAddr(TaskContext &context) override;
 
- private:
-  Status SetMemCopyTask(const domi::TaskDef &task_def);
-
-  Status InitForDependComputeTask();
-
-  Status UpdateShapeAndDataByResultSummary(TaskContext &context);
-
-  ///
-  /// read result summary and prepare copy task memory.
-  /// @param context task context
-  /// @param out_shape_hbm if scalar, TensorBuffer->data is null, size=0
-  /// @return SUCCESS:success other:failed
-  ///
-  Status ReadResultSummaryAndPrepareMemory(TaskContext &context,
-                                           std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm);
   Status CopyDataToHbm(TaskContext &context,
-                       const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm);
+                       const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) override;
 
-  Status UpdateShapeByHbmBuffer(TaskContext &context,
-                                const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm);
+  Status InitForDependComputeTask() override;
 
-  Status PrepareCopyInputs(const TaskContext &context,
-                           const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm);
+ private:
+  Status SetMemCopyTask(const domi::TaskDef &task_def);
 
   static Status EnsureSessionCreated(uint64_t session_id);
   static uint64_t GetStepIdAddr(const HybridModel &model);
@@ -135,17 +148,8 @@ class AicpuTfNodeTask : public AicpuNodeTaskBase {
   // just used for depend DEPEND_COMPUTE op
   std::unique_ptr<TensorBuffer> copy_task_args_buf_;
 
-  std::vector<std::unique_ptr<TensorBuffer>> output_summary_;
-  std::vector<aicpu::FWKAdapter::ResultSummary> output_summary_host_;
-
   std::unique_ptr<TensorBuffer> copy_ioaddr_dev_;
 
-  std::unique_ptr<TensorBuffer> copy_input_release_flag_dev_;
-  std::unique_ptr<TensorBuffer> copy_input_data_size_dev_;
-  std::unique_ptr<TensorBuffer> copy_input_src_dev_;
-  std::unique_ptr<TensorBuffer> copy_input_dst_dev_;
-  bool need_sync_ = false;
-
   std::unique_ptr<TensorBuffer> copy_workspace_buf_;
 };
 
@@ -162,16 +166,30 @@ class AicpuNodeTask : public AicpuNodeTaskBase {
 
   Status LaunchTask(TaskContext &context) override;
 
-  Status TaskCallback(TaskContext &context) override;
+  Status CopyDataToHbm(TaskContext &context,
+                       const std::vector<std::unique_ptr<TensorBuffer>> &out_shape_hbm) override;
 
   Status UpdateIoAddr(TaskContext &context) override;
 
+  Status InitForDependComputeTask() override;
+private:
+  Status SetMemCopyTask(const domi::TaskDef &task_def);
+
  protected:
   // host mem
   std::unique_ptr<uint8_t[]> args_;
+    // host memcpy mem
+  std::unique_ptr<uint8_t[]> memcpy_args_;
 
+  std::string memcpy_so_name_;
+
+  std::string memcpy_kernel_name_;
+  // args size
+  uint32_t memcpy_args_size_ = 0;  
   // args size
   uint32_t args_size_ = 0;
+
+  std::vector<uint64_t> copy_io_addr_;
 };
 
 class AiCpuNodeExecutor : public NodeExecutor {

ge/single_op/single_op.cc

@@ -433,13 +433,11 @@ Status DynamicSingleOp::ExecuteAsync(const vector<GeTensorDesc> &input_desc,
   if (!inputs_size.empty()) {
     StreamResource *stream_resource  = SingleOpManager::GetInstance().GetResource(resource_id_, stream_);
     GE_CHK_STATUS_RET_NOLOG(UpdateInputsBufferAddr(stream_resource, stream_, inputs_size, update_buffers));
+    GE_CHK_STATUS_RET_NOLOG(SetHostTensorValue(input_desc, input_buffers));
   }
 
   if (hybrid_model_executor_ != nullptr) {
     GELOGD("Execute multi-task dynamic single op by hybrid model executor");
-    if (!inputs_size.empty()) {
-      GE_CHK_STATUS_RET_NOLOG(SetHostTensorValue(input_desc, input_buffers));
-    }
     hybrid::HybridModelExecutor::ExecuteArgs args;
     GE_CHK_STATUS_RET_NOLOG(InitHybridModelArgs(update_buffers, output_buffers, input_desc, args));
 

ge/single_op/single_op_model.cc

@@ -333,7 +333,7 @@ Status SingleOpModel::BuildTaskList(StreamResource *stream_resource, SingleOp &s
         single_op.tasks_.emplace_back(tbe_task);
       } else if (kernel_type == ccKernelType::AI_CPU || kernel_type == ccKernelType::CUST_AI_CPU) {
         GELOGD("Building AICPU_CC task");
-        OpTask *task = nullptr;
+        AiCpuCCTask *task = nullptr;
         uint64_t singleop_kernel_id = aicpu_kernel_id++;
         GELOGI("Build singleOp CCTask, kernel_id = %lu", singleop_kernel_id);
         GE_CHK_STATUS_RET_NOLOG(BuildCpuKernelTask(task_def.kernel(), &task, singleop_kernel_id));
@@ -489,7 +489,7 @@ Status SingleOpModel::BuildKernelExTask(const domi::KernelExDef &kernel_def, AiC
   return SUCCESS;
 }
 
-Status SingleOpModel::BuildCpuKernelTask(const domi::KernelDef &kernel_def, OpTask **task, uint64_t kernel_id) {
+Status SingleOpModel::BuildCpuKernelTask(const domi::KernelDef &kernel_def, AiCpuCCTask **task, uint64_t kernel_id) {
   const auto &context = kernel_def.context();
   auto iter = op_list_.find(context.op_index());
   if (iter == op_list_.end()) {
@@ -611,10 +611,19 @@ Status SingleOpModel::BuildTaskListForDynamicOp(StreamResource *stream_resource,
   } else if (lib_name == kEngineNameAiCpu) {
     const auto &task_def = task_defs[0];
     GELOGD("Building AICPU_CC task");
-    OpTask *task = nullptr;
+    AiCpuCCTask *task = nullptr;
     uint64_t dynamic_singleop_kernel_id = aicpu_kernel_id++;
     GELOGI("Build dynamic singleOp CCTask, kernel_id = %lu", dynamic_singleop_kernel_id);
     GE_CHK_STATUS_RET_NOLOG(BuildCpuKernelTask(task_def.kernel(), &task, dynamic_singleop_kernel_id));
+    if (task->GetUnknownType() == DEPEND_COMPUTE) {
+      if (task_defs.size() < kNumTaskWithMemCpyTask) {
+        GELOGE(ACL_ERROR_GE_PARAM_INVALID, "[Check][Task]The copy task of the fourth operator was not found.");
+        REPORT_INNER_ERROR("E19999", "The copy task of the fourth operator was not found.");
+        return ACL_ERROR_GE_PARAM_INVALID;
+      }
+      const TaskDef &copy_task_def = task_defs[1];
+      GE_CHK_STATUS_RET_NOLOG(task->SetMemCopyTask(copy_task_def.kernel()));
+    }
     task->SetModelArgs(model_name_, model_id_);
     single_op.op_task_.reset(task);
   } else if (lib_name == kEngineNameAiCpuTf) {

ge/single_op/single_op_model.h

@@ -71,7 +71,7 @@ class SingleOpModel {
   Status BuildKernelTask(const domi::TaskDef &task_def, TbeOpTask **task);
   Status BuildAtomicTask(const domi::TaskDef &task_def, AtomicAddrCleanOpTask **task);
   Status BuildKernelExTask(const domi::KernelExDef &kernel_def, AiCpuTask **task, uint64_t kernel_id);
-  Status BuildCpuKernelTask(const domi::KernelDef &kernel_def, OpTask **task, uint64_t kernel_id);
+  Status BuildCpuKernelTask(const domi::KernelDef &kernel_def, AiCpuCCTask **task, uint64_t kernel_id);
 
   static void ParseOpModelParams(ModelHelper &model_helper, SingleOpModelParam &param);
   void ParseArgTable(OpTask *task, SingleOp &op);

ge/single_op/task/aicpu_kernel_task_builder.cc

@@ -102,11 +102,8 @@ Status AiCpuCCTaskBuilder::BuildTask(AiCpuCCTask &task, uint64_t kernel_id, cons
     return ret;
   }
   GE_CHK_STATUS_RET(task.SetInputConst(), "[Set][InputConst] failed.");
+  GE_CHK_STATUS_RET(task.InitForSummaryAndCopy(), "[Init][SummaryAndCopy] failed.");
 
-  if (task.GetUnknownType() == DEPEND_COMPUTE) {
-    GELOGE(FAILED, "[Get][UnknownType] is depend compute, it's not supported now.");
-    return FAILED;
-  }
   auto aicpu_param_head = reinterpret_cast<aicpu::AicpuParamHead *>(task.args_.get());
   if (task.ext_info_addr_dev_ != nullptr) {
     aicpu_param_head->extInfoLength = kernel_ext_info.size();

ge/single_op/task/op_task.cc

@@ -293,9 +293,6 @@ Status TbeOpTask::UpdateNodeByShape(const vector<GeTensorDesc> &input_desc, cons
 }
 
 Status TbeOpTask::EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size) {
-  node_ = node;
-  tiling_buffer_ = tiling_buffer;
-  max_tiling_size_ = max_tiling_size;
   if (tiling_buffer != nullptr) {
     uintptr_t *arg_base = nullptr;
     size_t arg_num = 0;
@@ -313,6 +310,9 @@ Status TbeOpTask::EnableDynamicSupport(const NodePtr &node, void *tiling_buffer,
     }
     arg_base[tiling_index] = reinterpret_cast<uintptr_t>(tiling_buffer);
   }
+  node_ = node;
+  tiling_buffer_ = tiling_buffer;
+  max_tiling_size_ = max_tiling_size;
   return SUCCESS;
 }
 
@@ -481,25 +481,6 @@ void TbeOpTask::GetIoAddr(uintptr_t *&arg_base, size_t &arg_count) {
   }
 }
 
-Status AtomicAddrCleanOpTask::EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size) {
-  node_ = node;
-  tiling_buffer_ = tiling_buffer;
-  max_tiling_size_ = max_tiling_size;
-  if (tiling_buffer != nullptr) {
-    uintptr_t *arg_base = nullptr;
-    size_t arg_num = 0;
-    GetIoAddr(arg_base, arg_num);
-    uint32_t tiling_index = atomic_output_indices_.size();
-    if (arg_num == 0 || arg_num < tiling_index) {
-      GELOGE(ACL_ERROR_GE_INTERNAL_ERROR, "[Check][Size]Tiling index %u, arg number %zu is invalid.",
-             tiling_index, arg_num);
-      return ACL_ERROR_GE_INTERNAL_ERROR;
-    }
-    arg_base[tiling_index] = reinterpret_cast<uintptr_t>(tiling_buffer);
-  }
-  return SUCCESS;
-}
-
 Status AtomicAddrCleanOpTask::UpdateNodeByShape(const vector<GeTensorDesc> &input_desc,
                                                 const vector<GeTensorDesc> &output_desc) {
   return SUCCESS;
@@ -583,6 +564,17 @@ AiCpuBaseTask::~AiCpuBaseTask() {
   if (ext_info_addr_dev_ != nullptr) {
     (void)rtFree(ext_info_addr_dev_);
   }
+
+  FreeHbm(copy_input_release_flag_dev_);
+  FreeHbm(copy_input_data_size_dev_);
+  FreeHbm(copy_input_src_dev_);
+  FreeHbm(copy_input_dst_dev_);
+  for (auto summary : output_summary_) {
+    FreeHbm(summary);
+  }
+  for (auto out_shape : out_shape_hbm_) {
+    FreeHbm(out_shape);
+  }
 }
 
 Status AiCpuBaseTask::SetExtInfoAndType(const std::string &kernel_ext_info, uint64_t kernel_id) {
@@ -795,17 +787,7 @@ AiCpuTask::~AiCpuTask() {
   FreeHbm(workspace_addr_);
   FreeHbm(copy_workspace_buf_);
   FreeHbm(copy_ioaddr_dev_);
-  FreeHbm(copy_input_release_flag_dev_);
-  FreeHbm(copy_input_data_size_dev_);
-  FreeHbm(copy_input_src_dev_);
-  FreeHbm(copy_input_dst_dev_);
   FreeHbm(copy_task_args_buf_);
-  for (auto summary : output_summary_) {
-    FreeHbm(summary);
-  }
-  for (auto out_shape : out_shape_hbm_) {
-    FreeHbm(out_shape);
-  }
 }
 
 Status AiCpuTask::LaunchKernel(rtStream_t stream) {
@@ -835,7 +817,7 @@ Status AiCpuTask::LaunchKernel(rtStream_t stream) {
   return SUCCESS;
 }
 
-Status AiCpuTask::PrepareCopyInputs(vector<DataBuffer> &outputs) {
+Status AiCpuBaseTask::PrepareCopyInputs(vector<DataBuffer> &outputs) {
   std::vector<uint64_t> copy_input_release_flag;
   std::vector<uint64_t> copy_input_data_size;
   std::vector<uint64_t> copy_input_src;
@@ -876,7 +858,7 @@ Status AiCpuTask::PrepareCopyInputs(vector<DataBuffer> &outputs) {
   return SUCCESS;
 }
 
-Status AiCpuTask::ReadResultSummaryAndPrepareMemory() {
+Status AiCpuBaseTask::ReadResultSummaryAndPrepareMemory() {
   for (size_t i = 0; i < num_outputs_; ++i) {
     auto &result_summary = output_summary_host_[i];
 
@@ -903,7 +885,20 @@ Status AiCpuTask::CopyDataToHbm(vector<DataBuffer> &outputs,
   return SUCCESS;
 }
 
-Status AiCpuTask::UpdateShapeByHbmBuffer(vector<GeTensorDesc> &output_desc) {
+Status AiCpuCCTask::CopyDataToHbm(vector<DataBuffer> &outputs,
+                                  rtStream_t stream) {
+  GE_CHK_STATUS_RET_NOLOG(PrepareCopyInputs(outputs));
+  
+  auto ret = rtCpuKernelLaunchWithFlag(static_cast<const void *>(memcpy_so_name_.data()),
+                                       static_cast<const void *>(memcpy_kernel_name_.data()),
+                                       block_dim_, memcpy_args_.get(), static_cast<uint32_t>(memcpy_args_size_),
+                                       nullptr, stream, dump_flag_);
+  GE_CHK_RT_RET(ret);
+  GE_CHK_RT_RET(rtStreamSynchronize(stream));
+  return SUCCESS;
+}
+
+Status AiCpuBaseTask::UpdateShapeByHbmBuffer(vector<GeTensorDesc> &output_desc) {
   for (size_t i = 0; i < num_outputs_; ++i) {
     const auto &result_summary = output_summary_host_[i];
     std::vector<int64_t> shape_dims;
@@ -932,7 +927,7 @@ Status AiCpuTask::UpdateShapeByHbmBuffer(vector<GeTensorDesc> &output_desc) {
 }
 
 
-Status AiCpuTask::UpdateShapeAndDataByResultSummary(vector<GeTensorDesc> &output_desc,
+Status AiCpuBaseTask::UpdateShapeAndDataByResultSummary(vector<GeTensorDesc> &output_desc,
                                                     vector<DataBuffer> &outputs,
                                                     rtStream_t stream) {
   if (num_outputs_ == 0) {
@@ -1122,8 +1117,91 @@ Status AiCpuCCTask::LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
   if (unknown_type_ == DEPEND_SHAPE_RANGE) {
     GE_CHK_RT_RET(rtStreamSynchronize(stream));
     GE_CHK_STATUS_RET_NOLOG(UpdateOutputShape(output_desc));
+  } else if (unknown_type_ == DEPEND_COMPUTE) {
+    GE_CHK_RT_RET(rtStreamSynchronize(stream));
+    GE_CHK_STATUS_RET_NOLOG(UpdateShapeAndDataByResultSummary(output_desc, output_buffers, stream));
+  }
+
+  return SUCCESS;
+}
+
+Status AiCpuCCTask::InitForSummaryAndCopy() {
+  if (unknown_type_ != DEPEND_COMPUTE || num_outputs_ == 0) {
+    GELOGI("Unknown_type is %d, output num is %zu.", unknown_type_, num_outputs_);
+    return SUCCESS;
+  }
+
+  output_summary_.resize(num_outputs_);
+  constexpr auto result_summary_size = sizeof(aicpu::FWKAdapter::ResultSummary);
+  for (size_t i = 0; i < num_outputs_; ++i) {
+    GE_CHK_RT_RET(rtMalloc(&output_summary_[i], result_summary_size, RT_MEMORY_HBM));
   }
+  output_summary_host_.resize(num_outputs_);
+
+  const size_t copy_input_buf_len = num_outputs_ * kCopyNum * sizeof(uint64_t);
+
+  GE_CHK_RT_RET(rtMalloc(&copy_input_release_flag_dev_, copy_input_buf_len, RT_MEMORY_HBM));
+  GE_CHK_RT_RET(rtMalloc(&copy_input_data_size_dev_, copy_input_buf_len, RT_MEMORY_HBM));
+  GE_CHK_RT_RET(rtMalloc(&copy_input_src_dev_, copy_input_buf_len, RT_MEMORY_HBM));
+  GE_CHK_RT_RET(rtMalloc(&copy_input_dst_dev_, copy_input_buf_len, RT_MEMORY_HBM));
+
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_release_flag_dev_));
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_data_size_dev_));
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_src_dev_));
+  copy_io_addr_.emplace_back(reinterpret_cast<uintptr_t>(copy_input_dst_dev_));
+  return SUCCESS;
+}
 
+Status AiCpuCCTask::SetMemCopyTask(const domi::KernelDef &kernel_def) {
+  auto &memcpy_args = kernel_def.args();
+  memcpy_args_size_ = kernel_def.args_size();
+  memcpy_so_name_ = kernel_def.so_name();
+  memcpy_kernel_name_ = kernel_def.kernel_name();
+  GE_IF_BOOL_EXEC(memcpy_args.size() != memcpy_args_size_,
+                  REPORT_INNER_ERROR("E19999", "MemCopy task def args.size=%zu, but args_size=%u not equal.",
+                                     memcpy_args.size(), memcpy_args_size_);
+                  GELOGE(FAILED, "[Check][Size]MemCopy task def args.size=%zu, but args_size=%u not equal.",
+                         memcpy_args.size(), memcpy_args_size_);
+                  return FAILED;);
+  GE_IF_BOOL_EXEC(memcpy_args_size_ < sizeof(aicpu::AicpuParamHead),
+                  REPORT_INNER_ERROR("E19999",
+                                     "Task def args_size=%u is less than aicpu param head len=%zu.",
+                                     memcpy_args_size_, sizeof(aicpu::AicpuParamHead));
+                  GELOGE(FAILED,
+                         "[Check][Size] Task def args_size=%u is less than aicpu param head len=%zu.",
+                         memcpy_args_size_, sizeof(aicpu::AicpuParamHead));
+                  return FAILED;);
+
+  memcpy_args_.reset(new(std::nothrow) uint8_t[memcpy_args_size_]());
+  GE_IF_BOOL_EXEC(memcpy_args_ == nullptr,
+                  REPORT_INNER_ERROR("E19999", "new memory failed for Node[MemCopy], task_size[%u].",
+                                     memcpy_args_size_);
+                  GELOGE(FAILED, "[Malloc][Memory] failed for Node[MemCopy], task_size[%u].",
+                         memcpy_args_size_);
+                  return FAILED;);
+
+  errno_t sec_ret = memcpy_s(memcpy_args_.get(), memcpy_args_size_, memcpy_args.c_str(), memcpy_args.size());
+  GE_IF_BOOL_EXEC(sec_ret != EOK,
+                  REPORT_INNER_ERROR("E19999",
+                                     "memcpy_s argc_ failed for Node[MemCopy], ret: %d", sec_ret);
+                  GELOGE(INTERNAL_ERROR,
+                         "[Update][args] failed for Node[MemCopy], ret: %d", sec_ret);
+                  return sec_ret;);
+  auto memcpy_param_head = reinterpret_cast<aicpu::AicpuParamHead *>(memcpy_args_.get());
+  uint32_t memcpy_io_num = memcpy_param_head->ioAddrNum;
+  auto memcpy_io_addr = memcpy_args_.get() + sizeof(aicpu::AicpuParamHead);
+  // if has input and output, need copy to ioaddr
+  int cpy_ret = memcpy_s(memcpy_io_addr, memcpy_args_size_ - sizeof(aicpu::AicpuParamHead),
+                         &copy_io_addr_[0], sizeof(uint64_t) * memcpy_io_num);
+  GE_IF_BOOL_EXEC(cpy_ret != 0,
+                  REPORT_INNER_ERROR("E19999", "Node[Memcpoy] memcpy io addr to AicpuParamHead failed,"
+                         "ret=%d, args_size=%u, io nums=%u.",
+                         cpy_ret, memcpy_args_size_, memcpy_io_num);
+                  GELOGE(INTERNAL_ERROR, "[Update][io_addr]Node[MemCopy] memcpy io addr to AicpuParamHead failed,"
+                         "ret=%d, args_size=%u, io nums=%u.",
+                         cpy_ret, memcpy_args_size_, memcpy_io_num);
+                  return INTERNAL_ERROR;);
+  GELOGD("Set memcpy task for node[MemCopy] successfully.");
   return SUCCESS;
 }
 

ge/single_op/task/op_task.h

@@ -97,7 +97,7 @@ class TbeOpTask : public OpTask {
   const void *GetArgs() const;
   size_t GetArgSize() const;
   const std::string &GetStubName() const;
-  virtual Status EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size);
+  Status EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size);
   const std::string &GetTaskType() const override;
   void SetHandle(void *handle);
 
@@ -149,7 +149,6 @@ class TbeOpTask : public OpTask {
 class AtomicAddrCleanOpTask : public TbeOpTask {
  public:
   Status InitAtomicAddrCleanIndices();
-  Status EnableDynamicSupport(const NodePtr &node, void *tiling_buffer, uint32_t max_tiling_size) override;
 
  private:
   Status UpdateNodeByShape(const vector<GeTensorDesc> &input_desc,
@@ -157,8 +156,8 @@ class AtomicAddrCleanOpTask : public TbeOpTask {
   Status UpdateIoAddr(const vector<DataBuffer> &inputs, const vector<DataBuffer> &outputs) override;
   Status UpdateTilingArgs(rtStream_t stream) override;
   Status CalcTilingInfo(optiling::utils::OpRunInfo &run_info) override;
-
   std::vector<int> atomic_output_indices_;
+
 };
 
 class AiCpuBaseTask : public OpTask {
@@ -179,7 +178,17 @@ class AiCpuBaseTask : public OpTask {
                        rtStream_t stream);
   Status UpdateOutputShape(vector<GeTensorDesc> &output_desc);
   Status UpdateShapeToOutputDesc(const GeShape &shape_new, GeTensorDesc &output_desc);
+  
+  Status UpdateShapeAndDataByResultSummary(vector<GeTensorDesc> &output_desc,
+                                           vector<DataBuffer> &outputs,
+                                           rtStream_t stream);
+  Status ReadResultSummaryAndPrepareMemory();
 
+  Status PrepareCopyInputs(vector<DataBuffer> &outputs);
+
+  Status UpdateShapeByHbmBuffer(vector<GeTensorDesc> &output_desc);
+
+  virtual Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream) = 0;
  protected:
   size_t num_inputs_ = 0;
   size_t num_outputs_ = 0;
@@ -187,6 +196,16 @@ class AiCpuBaseTask : public OpTask {
   std::unique_ptr<ge::hybrid::AicpuExtInfoHandler> aicpu_ext_handle_;
   void *ext_info_addr_dev_ = nullptr;
   vector<bool> input_is_const_;
+  
+  std::vector<void *> output_summary_;
+  std::vector<aicpu::FWKAdapter::ResultSummary> output_summary_host_;
+
+  void *copy_input_release_flag_dev_ = nullptr;
+  void *copy_input_data_size_dev_ = nullptr;
+  void *copy_input_src_dev_ = nullptr;
+  void *copy_input_dst_dev_ = nullptr;
+
+  vector<void *> out_shape_hbm_;
 };
 
 class AiCpuTask : public AiCpuBaseTask {
@@ -207,16 +226,9 @@ class AiCpuTask : public AiCpuBaseTask {
  private:
   // for copy task.
   Status InitForSummaryAndCopy();
-  Status UpdateShapeAndDataByResultSummary(vector<GeTensorDesc> &output_desc,
-                                           vector<DataBuffer> &outputs,
-                                           rtStream_t stream);
-  Status ReadResultSummaryAndPrepareMemory();
-
-  Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream);
-  Status PrepareCopyInputs(vector<DataBuffer> &outputs);
-
-  Status UpdateShapeByHbmBuffer(vector<GeTensorDesc> &output_desc);
 
+  Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream) override;
+ private:
   friend class AiCpuTaskBuilder;
   void *workspace_addr_ = nullptr;
   std::string task_info_;
@@ -235,17 +247,7 @@ class AiCpuTask : public AiCpuBaseTask {
   void *copy_task_args_buf_ = nullptr;
   void *copy_workspace_buf_ = nullptr;
 
-  std::vector<void *> output_summary_;
-  std::vector<aicpu::FWKAdapter::ResultSummary> output_summary_host_;
-
   void *copy_ioaddr_dev_ = nullptr;
-
-  void *copy_input_release_flag_dev_ = nullptr;
-  void *copy_input_data_size_dev_ = nullptr;
-  void *copy_input_src_dev_ = nullptr;
-  void *copy_input_dst_dev_ = nullptr;
-
-  vector<void *> out_shape_hbm_;
   uint64_t kernel_id_ = 0;
 };
 
@@ -264,13 +266,16 @@ class AiCpuCCTask : public AiCpuBaseTask {
   void SetkernelName(const std::string &kernel_Name);
   void SetIoAddr(uintptr_t *io_addr);
   size_t GetArgSize() const;
-
+  Status SetMemCopyTask(const domi::KernelDef &kernel_def);
   Status LaunchKernel(const std::vector<GeTensorDesc> &input_desc,
                       const std::vector<DataBuffer> &input_buffers,
                       std::vector<GeTensorDesc> &output_desc,
                       std::vector<DataBuffer> &output_buffers,
                       rtStream_t stream)  override;
+ private:
+  Status InitForSummaryAndCopy();
 
+  Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream) override;
 private:
   friend class AiCpuCCTaskBuilder;
   std::string so_name_;
@@ -284,6 +289,13 @@ private:
   uint32_t dump_flag_ = RT_KERNEL_DEFAULT;
   std::string op_type_;
   uint64_t kernel_id_ = 0;
+  // host memcpy mem
+  std::unique_ptr<uint8_t[]> memcpy_args_;
+  std::string memcpy_so_name_;
+  std::string memcpy_kernel_name_;
+  std::vector<uint64_t> copy_io_addr_;
+  // args size
+  uint32_t memcpy_args_size_ = 0;
 };
 
 class MemcpyAsyncTask : public OpTask {

ge/single_op/task/tbe_task_builder.cc

@@ -425,7 +425,7 @@ Status TbeTaskBuilder::InitTilingInfo(TbeOpTask &task) {
     GELOGD("[%s] Done allocating tiling buffer, size=%ld.", op_desc_->GetName().c_str(), max_size);
   }
 
-  GE_CHK_STATUS_RET_NOLOG(task.EnableDynamicSupport(node_, tiling_buffer, static_cast<uint32_t>(max_size)));
+  task.EnableDynamicSupport(node_, tiling_buffer, static_cast<uint32_t>(max_size));
   return SUCCESS;
 }
 

tests/ut/ge/hybrid/node_executor/aicpu/aicpu_node_executor_unittest.cc

@@ -165,4 +165,3 @@ TEST_F(UtestAicpuNodeExecutor, aicpu_tf_node_task) {
 }
 
 }  // namespace ge
-

tests/ut/ge/single_op/single_op_task_unittest.cc

@@ -237,27 +237,3 @@ TEST_F(UtestSingleOpTask, test_aicpu_task_update_io_addr) {
     ASSERT_EQ(ret, PARAM_INVALID);
   }
 }
-
-TEST_F(UtestSingleOpTask, test_dynamic_support) {
-  auto graph = make_shared<ComputeGraph>("graph");
-  auto op_desc = make_shared<OpDesc>("Add", "Add");
-  auto node = graph->AddNode(op_desc);
-  AtomicAddrCleanOpTask atomic_task;
-  TbeOpTask tbe_task;
-
-  tbe_task.arg_size_ = sizeof(void *) * 1;
-  tbe_task.args_.reset(new (std::nothrow) uint8_t[tbe_task.arg_size_]);
-  atomic_task.arg_size_ = sizeof(void *) * 1;
-  atomic_task.args_.reset(new (std::nothrow) uint8_t[atomic_task.arg_size_]);
-  ASSERT_EQ(tbe_task.EnableDynamicSupport(node, (void *)0x0001, 1), ACL_ERROR_GE_INTERNAL_ERROR);
-  ASSERT_EQ(atomic_task.EnableDynamicSupport(node, (void *)0x0001, 1), ACL_ERROR_GE_INTERNAL_ERROR);
-
-  tbe_task.arg_size_ = sizeof(void *) * 2;
-  tbe_task.args_.reset(new (std::nothrow) uint8_t[tbe_task.arg_size_]);
-  atomic_task.arg_size_ = sizeof(void *) * 2;
-  atomic_task.args_.reset(new (std::nothrow) uint8_t[atomic_task.arg_size_]);
-  ASSERT_EQ(tbe_task.EnableDynamicSupport(node, (void *)0x0001, 1), SUCCESS);
-  ASSERT_EQ(atomic_task.EnableDynamicSupport(node, (void *)0x0001, 1), SUCCESS);
-  tbe_task.tiling_buffer_ = nullptr;
-  atomic_task.tiling_buffer_ = nullptr;
-}