8#ifndef ENCAP_SHMEM_HPP_
9#define ENCAP_SHMEM_HPP_
16template<
unsigned int sz>
19 static const int size = (sz / 4)*4 + (sz % 4 != 0)*4;
20 static const int nthr = (sz / 4) + (sz % 4 != 0);
22 __device__
static void copy(
int * src,
int * dst)
24 if (threadIdx.x < nthr)
25 {dst[threadIdx.x] = src[threadIdx.x];}
memcpy it split the copy across threads