Asynchronous execution

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi,

I am working on an asynchronous execution using a subroutine similar to this:

subroutine sub1(stream)

  $acc data create () copy() async(stream)
  $acc do loop  async(stream)
    ...
  $acc end do loop
  $acc data end

end subroutine sub1

The subroutine is called several times by a code like this:

do i =1 , num_chunks

streamid= mod(i,2) +1                  !  create the ids of two streams: 1,2

call sub1(streamid)

end do

The idea is to create two (or more) streams to have a pipelined execution. My doubt is if the acc data end clause performs implicitly a cudaEventSynchonize call that will not enable the concurrent execution.
I hope you can clarify this.

Regards,

Guillermo

2

Hi Guillermo,

Structured data regions don’t allow for the use of the async clause.

Try using unstructured data regions instead.

subroutine sub1(stream) 

  $acc enter data create () copyin() async(stream) 
  $acc do loop  async(stream) 
    ... 
  $acc end do loop 
  $acc exit data delete() copyout() async(stream)

end subroutine sub1

-Mat