fix

ge/single_op/single_op_model.cc

@@ -328,7 +328,7 @@ 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));

ge/single_op/task/op_task.cc

@@ -786,7 +786,7 @@ Status AiCpuCCTask::CopyDataToHbm(vector<DataBuffer> &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_arg_size_),
+                                       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));
@@ -1040,8 +1040,6 @@ Status AiCpuCCTask::InitForSummaryAndCopy() {
   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));
 
-  GE_CHK_RT_RET(rtMalloc(&copy_task_args_buf_, sizeof(STR_FWK_OP_KERNEL), RT_MEMORY_HBM));
-
   std::vector<uint64_t> copy_io_addr;
   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_));

ge/single_op/task/op_task.h

@@ -161,6 +161,8 @@ class AiCpuBaseTask : public OpTask {
   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;
@@ -201,7 +203,7 @@ class AiCpuTask : public AiCpuBaseTask {
   // for copy task.
   Status InitForSummaryAndCopy();
 
-  Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream);
+  Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream) override;
  private:
   friend class AiCpuTaskBuilder;
   void *workspace_addr_ = nullptr;
@@ -247,7 +249,7 @@ class AiCpuCCTask : public AiCpuBaseTask {
  private:
   Status InitForSummaryAndCopy();
 
-  Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream);
+  Status CopyDataToHbm(vector<DataBuffer> &outputs, rtStream_t stream) override;
 private:
   friend class AiCpuCCTaskBuilder;
   std::string so_name_;
@@ -266,7 +268,7 @@ private:
   std::string memcpy_so_name_;
   std::string memcpy_kernel_name_;
   // args size
-  size_t memcpy_args_size_ = 0;  
+  size_t memcpy_args_size_ = 0;
 };
 
 class MemcpyAsyncTask : public OpTask {