GPU implementation of the EC-DG 2-D tendency.
TwoPointFluxMethod is computed on the host (user Fortran function) and uploaded before the GPU kernel sequence. The surface term uses the dedicated ECDGSurfaceContribution_2D_gpu kernel which applies the Jacobian weighting consistently with the volume term.
| Type | Intent | Optional | Attributes | Name | ||
|---|---|---|---|---|---|---|
| class(ECDGModel2D), | intent(inout) | :: | this |
subroutine CalculateTendency_ECDGModel2D(this)
!! GPU implementation of the EC-DG 2-D tendency.
!!
!! TwoPointFluxMethod is computed on the host (user Fortran function) and
!! uploaded before the GPU kernel sequence. The surface term uses the
!! dedicated ECDGSurfaceContribution_2D_gpu kernel which applies the
!! Jacobian weighting consistently with the volume term.
implicit none
class(ECDGModel2D),intent(inout) :: this
! Local
integer :: ndof
call this%solution%BoundaryInterp()
! Post the halo exchange for the prognostic variables; the MPI messages
! are in flight while the hooks, boundary conditions, source, and the
! two-point volume flux and its divergence below execute.
call this%solution%SideExchangeStart(this%mesh,this%nstepped)
call this%PreTendencyHook()
call this%SetBoundaryCondition()
if(this%gradient_enabled) then
! The BR gradient consumes extBoundary (through the side averages), so
! the exchange must complete before the gradient is computed.
call this%solution%SideExchangeFinish(this%mesh)
call this%CalculateSolutionGradient()
call this%SetGradientBoundaryCondition()
call this%solutionGradient%AverageSides()
endif
call this%SourceMethod()
! Two-point flux: concrete GPU models (e.g. ECAdvection2D) override
! TwoPointFluxMethod to run entirely on device. Models that compute
! on the host should call twoPointFlux%UpdateDevice() at the end of
! their TwoPointFluxMethod override.
call this%TwoPointFluxMethod()
! EC volume divergence (uses MappedTwoPointVectorDivergence_2D_gpu)
call this%twoPointFlux%MappedDivergence(this%fluxDivergence%interior_gpu)
! BoundaryFlux is the first consumer of extBoundary; the two-point volume
! flux and its divergence above overlap with the halo exchange (a no-op
! wait when gradients already finished it). BoundaryFlux writes only the
! Riemann trace (flux boundarynormal), disjoint from the volume-term
! outputs, so this reordering does not change any floating-point results.
call this%solution%SideExchangeFinish(this%mesh)
call this%BoundaryFlux()
! Add (1/J) * M^{-1} B^T f_Riemann
call ECDGSurfaceContribution_2D_gpu( &
this%flux%boundarynormal_gpu, &
this%geometry%J%interior_gpu, &
this%solution%interp%bMatrix_gpu, &
this%solution%interp%qWeights_gpu, &
this%fluxDivergence%interior_gpu, &
this%solution%interp%N,this%solution%nVar,this%mesh%nElem)
! dSdt = source - fluxDivergence
ndof = this%solution%nVar* &
this%mesh%nElem* &
(this%solution%interp%N+1)* &
(this%solution%interp%N+1)
call CalculateDSDt_gpu(this%fluxDivergence%interior_gpu, &
this%source%interior_gpu, &
this%dSdt%interior_gpu,ndof)
endsubroutine CalculateTendency_ECDGModel2D