inline DeepCopy( const ExecutionSpace& exec, void * dst , const void * src , size_t n ) { exec.fence(); hc::completion_future fut = DeepCopyAsyncROCm (dst,src,n); fut.wait(); // DeepCopyROCm (dst,src,n); }
DeepCopy( const ExecutionSpace& exec, void * dst , const void * src , size_t n ) { exec.fence(); memcpy( dst , src , n ); }
inline DeepCopy( const ExecutionSpace& exec, void * dst , const void * src , size_t n ) { exec.fence(); DeepCopyAsyncCuda (dst,src,n); }