fix

ge/hybrid/node_executor/aicpu/aicpu_node_executor.cc

@@ -759,6 +759,62 @@ Status AicpuNodeTask::Init(const HybridModel &model) {
   return SUCCESS;
 }
 
+Status AicpuNodeTask::InitForDependComputeTask() override {
+  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 task args buf
+  GE_CHK_STATUS_RET(AllocTensorBuffer(sizeof(STR_FWK_OP_KERNEL), copy_task_args_buf_),
+                    "[Alloc][TensorBuffer] failed for Node[%s] to copy task args, size=%zu",
+                    node_name_.c_str(), sizeof(STR_FWK_OP_KERNEL));
+
+  std::vector<uint64_t> copy_io_addr;
+  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()));
+
+  // mem copy op has 4 inputs and 0 output.
+  const auto copy_io_addr_size = sizeof(uint64_t) * copy_io_addr.size();
+
+  // can alloc in init, it can reuse
+  GE_CHK_STATUS_RET(AllocTensorBuffer(copy_io_addr_size, copy_ioaddr_dev_),
+                    "[Alloc][TensorBuffer] failed for Node[%s] to copy task ioaddr, size=%zu",
+                    node_name_.c_str(), copy_io_addr_size);
+
+  GE_CHK_RT_RET(rtMemcpy(copy_ioaddr_dev_->GetData(), copy_io_addr_size,
+                         &copy_io_addr[0], copy_io_addr_size, RT_MEMCPY_HOST_TO_DEVICE));
+  return SUCCESS;
+}
+
 Status AicpuNodeTask::UpdateIoAddr(TaskContext &context) {
   vector<uint64_t> io_addrs;
   io_addrs.reserve(node_item_->num_inputs + node_item_->num_outputs);