Stream Management

Accelerated Computing HPC Compilers Legacy PGI Compilers
#1

I am trying to use concurrent streams but I am having problems with it. Maybe someone has already tried to use streams and can give me a piece of code so I can see how it works. First of all I start with:

istat = cudaStreamCreate(1)

this should create a new stream, but still the compiler states:‘Argument number 1 to cudastreamcreate does not match INTENT (OUT)’. Am I already missing something? Am I correct that I then have to create for each stream a data copy? eg.

istat = cudaMemcpyAsync(adev,a,100,0)
istat = cudaMemcpyAsync(adev,a,100,1)

Thanks, thomi

#2

Hi thomi,

The function cudaStreamCreate expects an integer variable as it’s argument. The function will then assign it a stream id.

% cat teststream.cuf

program teststream
   use cudafor

   integer :: strm1, istat, N
   real, dimension(:),pinned,allocatable :: A
   real, dimension(:),allocatable, device :: Adev
   N=100
   allocate(A(N),Adev(N))
   A=1.1
   istat = cudaStreamCreate(strm1)
   print *, strm1
   istat = cudaMemcpyAsync(Adev,A,N,cudaMemcpyHostToDevice,strm1)

end program teststream
% pgf90 teststream.cuf -V10.9 ; a.out
            1

Note that the host array must be and allocatable and located in pinned memory.

Hope this helps,
Mat