Method ConvertOperations

tensorflow/lite/delegates/gpu/cl/inference_context.cc:181–218

}

Status InferenceContext::ConvertOperations(
    const CreationContext& creation_context, const GraphFloat32& graph,
    ModelHints hints) {
  std::vector<Node*> graph_nodes = graph.nodes();
  for (int i = 0; i < graph_nodes.size(); ++i) {
    const Node& node = *graph_nodes[i];
    auto inputs = graph.FindInputs(node.id);
    auto outputs = graph.FindOutputs(node.id);
    OperationDef op_def;
    op_def.precision = precision_;
    auto data_type = DeduceDataTypeFromPrecision(precision_);
    for (int j = 0; j < inputs.size(); ++j) {
      op_def.src_tensors.push_back({data_type, storage_type_});
    }
    for (int j = 0; j < outputs.size(); ++j) {
      op_def.dst_tensors.push_back({data_type, storage_type_});
    }
    std::unique_ptr<GPUOperation> gpu_op;
    RETURN_IF_ERROR(GPUOperationFromNode(creation_context, op_def, hints, graph,
                                         node, &gpu_op));
    CLNode cl_node;
    cl_node.operations.push_back(std::move(gpu_op));
    cl_node.ranges.push_back(int2(0, static_cast<int>(inputs.size())));
    cl_node.inputs.resize(inputs.size());
    for (int j = 0; j < inputs.size(); ++j) {
      cl_node.inputs[j] = inputs[j]->id;
    }
    cl_node.outputs.resize(outputs.size());
    for (int j = 0; j < outputs.size(); ++j) {
      cl_node.outputs[j] = outputs[j]->id;
    }
    cl_node.name = node.operation.type + " " + std::to_string(node.id) + " " +
                   std::to_string(i);
    nodes_.push_back(std::move(cl_node));
  }

  return OkStatus();
}

void InferenceContext::Merge() {
  std::unordered_set<ValueId> ready_tensors;