Method Merge

tensorflow/lite/delegates/gpu/cl/inference_context.cc:220–262

}

void InferenceContext::Merge() {
  std::unordered_set<ValueId> ready_tensors;
  for (const auto& input_id : input_ids_) {
    ready_tensors.insert(input_id);
  }
  for (int i = 0; i < nodes_.size(); ++i) {
    auto& node = nodes_[i];
    for (const auto& out_id : node.outputs) {
      ready_tensors.insert(out_id);
    }
    if (node.outputs.size() != 1) {
      continue;
    }
    std::vector<int> next_nodes;
    for (int j = i + 1; j < nodes_.size(); ++j) {
      for (int k = 0; k < nodes_[j].inputs.size(); ++k) {
        if (nodes_[j].inputs[k] == node.outputs[0]) {
          next_nodes.push_back(j);
        }
      }
    }
    if (next_nodes.size() != 1) {
      continue;
    }
    auto& linkable_node = nodes_[next_nodes[0]];
    auto* elementwise =
        dynamic_cast<ElementwiseOperation*>(linkable_node.operations[0].get());
    if (!elementwise || linkable_node.outputs.size() != 1 ||
        !IsReady(ready_tensors, linkable_node)) {
      continue;
    }
    MergeCLNodes(&linkable_node, &node);
    nodes_.erase(nodes_.begin() + next_nodes[0]);
    i -= 1;
  }
  for (auto& node : nodes_) {
    for (int j = 1; j < node.operations.size(); ++j) {
      auto* elementwise =
          dynamic_cast<ElementwiseOperation*>(node.operations[j].get());
      node.operations[0]->AddOperation(elementwise);
    }
  }
}

Status InferenceContext::AllocateMemory(const GraphFloat32& graph,
                                        const CLDevice& device,