Method ForEachEdgeWarp

grape/cuda/parallel/parallel_engine.h:1276–1307

  template <typename FRAG_T, typename WORK_SOURCE_T, typename ASSIGN_OP,
            typename EDGE_OP, EdgeDirection ed>
  inline void ForEachEdgeWarp(const Stream& stream, const FRAG_T& dev_frag,
                              const WORK_SOURCE_T& ws, ASSIGN_OP assign_op,
                              EDGE_OP op) {
    int grid_size, block_size;
    auto size = ws.size();
    if (size == 0) {
      return;
    }
    auto calc_shmem_size = [] DEV_HOST(int block_size) -> int {
      using vid_t = typename FRAG_T::vid_t;
      using vertex_t = typename FRAG_T::vertex_t;
      using nbr_t = typename FRAG_T::nbr_t;
      using metadata_t = typename std::result_of<ASSIGN_OP&(vertex_t&)>::type;

      int warp_size = 32;
      auto n_warp = (block_size + warp_size - 1) / warp_size;

      auto smem_size = block_size * (sizeof(VertexMetadata<vid_t, metadata_t>) +
                                     sizeof(const nbr_t*)) +
                       n_warp * warp_size * sizeof(vid_t);
      return smem_size;
    };

    auto lb_wrapper = [=] __device__() {
      LBWARP<FRAG_T, WORK_SOURCE_T, ASSIGN_OP, EDGE_OP, ed>(dev_frag, ws,
                                                            assign_op, op);
    };  // NOLINT

    KernelSizing(grid_size, block_size, size);
    KernelWrapper<<<grid_size, block_size, calc_shmem_size(block_size),
                    stream.cuda_stream()>>>(lb_wrapper);
  }

  template <typename FRAG_T, typename WORK_SOURCE_T, typename ASSIGN_OP,
            typename EDGE_OP, EdgeDirection ed>