Multi-GPU Support on Single Node Using Directive-Based Programming Model

Algorithm 2

S3D thermodynamics kernel in multi-GPU using hybrid model.

call omp_set_num_threads(2)

!$omp parallel private(m)

!$omp sections

!$omp section

call acc_set_device_num(1, acc_device_not_host)

!$acc data copyout(c(1:np)) copyin(T(1:np),…)

do m = 1, MR

 call calc_mixcp(np, nslvs, T, …  ,  c)

end do

!$acc end data

!$omp section

call acc_set_device_num(2, acc_device_not_host)

!$acc data copyout(h(1:np)) copyin(T(1:np),…)

do m = 1, MR

 call calc_mixenth(np, nslvs, T, …  ,  h)

end do

!$acc end data

!$omp end sections

!$omp end parallel