How to pass num_act_out_real from get_indice_pairs TensorRT plugin to implicit_gemm plugin

[ArseniuML]

From main.cu:

  // after get pair datas, we can start to do real convolution!
  // in static inference engine, you need to split pair-gen and conv to
  // different layers to reuse pair data
  // here we just use previous result.
  int num_act_out_real = std::get<1>(pair_res);

The question is - how to pass this data-dependent CPU variable from enqueue(..) function of get_indice_pairs plugin to enqueue(..) function of implicit_gemm plugin? TensorRT seems not to support passing CPU variables from one layer from another (or I missed something..) I don't want to copy num_act_out_real to GPU and back to CPU in another plugin...