typedef void (*func_t)(const PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<D> dst, int kWidth, int kHeight, int anchorX, int anchorY, const float* borderValue, cudaStream_t stream);
typedef void (*func_t)(const PtrStepSz<T> srcWhole, int xoff, int yoff, PtrStepSz<D> dst, int kWidth, int kHeight, int anchorX, int anchorY, const float* borderValue, cudaStream_t stream);