warpReduce         62 modules/cudev/include/opencv2/cudev/warp/reduce.hpp __device__ __forceinline__ void warpReduce(volatile T* smem, T& val, uint tid, const Op& op)
warpReduce         70 modules/cudev/include/opencv2/cudev/warp/reduce.hpp __device__ __forceinline__ void warpReduce(const tuple<P0, P1, P2, P3, P4, P5, P6, P7, P8, P9>& smem,