fix

ge/hybrid/node_executor/aicpu/aicpu_node_executor.cc

@@ -754,12 +754,59 @@ 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;
 }
 
-Status AicpuNodeTask::InitForDependComputeTask() override {
+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::KernelExDef &kernel_def = task_def.kernel_ex();
+  if (kernel_def.args_size() > sizeof(STR_FWK_OP_KERNEL)) {
+    GELOGE(PARAM_INVALID, "[Check][Size]sizeof STR_FWK_OP_KERNEL is:%lu, but args_size:%d is bigger",
+           sizeof(STR_FWK_OP_KERNEL), kernel_def.args_size());
+    REPORT_INNER_ERROR("E19999", "sizeof STR_FWK_OP_KERNEL is:%lu, but args_size:%d is bigger.",
+                       sizeof(STR_FWK_OP_KERNEL), kernel_def.args_size());
+    return PARAM_INVALID;
+  }
+  STR_FWK_OP_KERNEL aicpu_task = {0};
+  auto sec_ret = memcpy_s(&aicpu_task, sizeof(STR_FWK_OP_KERNEL),
+                          kernel_def.args().data(), kernel_def.args_size());
+  if (sec_ret != EOK) {
+    GELOGE(FAILED, "[Update][aicpu_task] failed, ret: %d", sec_ret);
+    REPORT_CALL_ERROR("E19999", "update aicpu_task failed, ret: %d.", sec_ret);
+    return FAILED;
+  }
+
+  GE_CHK_STATUS_RET(AllocTensorBuffer(kernel_def.task_info_size(), copy_workspace_buf_),
+                    "[Alloc][TensorBuffer] for Node[%s] to copy task workspace buf, size=%u.",
+                    node_name_.c_str(), kernel_def.task_info_size());
+
+  GE_CHK_RT_RET(rtMemcpy(copy_workspace_buf_->GetData(), kernel_def.task_info_size(),
+                         kernel_def.task_info().data(), kernel_def.task_info_size(), RT_MEMCPY_HOST_TO_DEVICE));
+
+  aicpu_task.fwkKernelBase.fwk_kernel.inputOutputAddr = reinterpret_cast<uintptr_t>(copy_ioaddr_dev_->GetData());
+  aicpu_task.fwkKernelBase.fwk_kernel.workspaceBaseAddr = reinterpret_cast<uintptr_t>(copy_workspace_buf_->GetData());
+  aicpu_task.fwkKernelBase.fwk_kernel.extInfoAddr = 0;
+  aicpu_task.fwkKernelBase.fwk_kernel.extInfoLen = 0;
+
+  GE_CHK_RT_RET(rtMemcpy(copy_task_args_buf_->GetData(), sizeof(STR_FWK_OP_KERNEL),
+                         &aicpu_task, sizeof(STR_FWK_OP_KERNEL), RT_MEMCPY_HOST_TO_DEVICE));
+  GELOGD("Set memcpy task for node[%s] successfully.", node_name_.c_str());
+  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);
@@ -899,6 +946,31 @@ Status AiCpuNodeExecutor::PrepareTask(NodeTask &task, TaskContext &context) cons
   return status;
 }
 
+Status AicpuNodeTask::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;
+  GE_CHK_STATUS_RET(ReadResultSummaryAndPrepareMemory(context, out_shape_hbm),
+                    "[Invoke][ReadResultSummaryAndPrepareMemory] failed for Node[%s].",
+                    node_name_.c_str());
+
+  RECORD_CALLBACK_EVENT(context.GetExecutionContext(), node_name_.c_str(),
+                        "[ReadResultSummaryAndPrepareMemory] End");
+
+  GE_CHK_STATUS_RET(CopyDataToHbm(context, out_shape_hbm),
+                    "[Invoke][CopyDataToHbm] failed for Node[%s] copy data to output.",
+                    node_name_.c_str());
+
+  RECORD_CALLBACK_EVENT(context.GetExecutionContext(), node_name_.c_str(), "[CopyDataToHbm] End");
+
+  GE_CHK_STATUS_RET(UpdateShapeByHbmBuffer(context, out_shape_hbm),
+                    "[Update][ShapeByHbmBuffer] failed for Node[%s].",
+                    node_name_.c_str());
+
+  GELOGD("Node[%s] update shape and data by result summary end.", node_name_.c_str());
+  return SUCCESS;
+}
+
 Status AiCpuNodeExecutor::LoadTask(const HybridModel &model,
                                    const NodePtr &node,
                                    std::shared_ptr<NodeTask> &task) const {

ge/hybrid/node_executor/aicpu/aicpu_node_executor.h

@@ -99,8 +99,6 @@ class AicpuTfNodeTask : public AicpuNodeTaskBase {
 
   Status LaunchTask(TaskContext &context) override;
 
-  Status TaskCallback(TaskContext &context) override;
-
   Status UpdateIoAddr(TaskContext &context) override;
   
   Status UpdateShapeAndDataByResultSummary(TaskContext &context) override;
@@ -167,15 +165,13 @@ class AicpuNodeTask : public AicpuNodeTaskBase {
 
   Status LaunchTask(TaskContext &context) override;
 
-  Status TaskCallback(TaskContext &context) override;
-
   Status UpdateIoAddr(TaskContext &context) override;
   
   Status UpdateShapeAndDataByResultSummary(TaskContext &context) override;
 
   Status InitForDependComputeTask() override;
 
-  Status SetMemCopyTask(const domi::TaskDef &task_def);
+  Status SetMemCopyTask(const domi::TaskDef &task_def) override;
  protected:
   // host mem
   std::unique_ptr<uint8_t[]> args_;