GPU implementation of the mortar exchange (see the base class for the algorithm) : traces are staged, reoriented, restricted, and projected in device memory with the SELF_Mortar kernels.
| Type | Intent | Optional | Attributes | Name | ||
|---|---|---|---|---|---|---|
| class(MappedScalar2D), | intent(inout) | :: | this | |||
| type(Mesh2D), | intent(inout) | :: | mesh |
subroutine MortarExchange_MappedScalar2D(this,mesh)
!! GPU implementation of the mortar exchange (see the base class for the
!! algorithm) : traces are staged, reoriented, restricted, and projected in
!! device memory with the SELF_Mortar kernels.
implicit none
class(MappedScalar2D),intent(inout) :: this
type(Mesh2D),intent(inout) :: mesh
! Local
integer :: offset
integer(c_size_t) :: buffSize
offset = mesh%decomp%offsetElem(mesh%decomp%rankId+1)
if(.not. c_associated(this%mortarBuff_gpu)) then
buffSize = int(this%interp%N+1,c_size_t)*4*mesh%nMortars*this%nvar*prec
call gpuCheck(hipMalloc(this%mortarBuff_gpu,buffSize))
endif
if(mesh%decomp%mpiEnabled) then
call this%MPIMortarExchangeAsync(mesh)
endif
! Stage rank-local traces in the big side's edge orientation
call MortarGather_2D_gpu(this%mortarBuff_gpu,this%boundary_gpu, &
mesh%mortarInfo_gpu,mesh%decomp%elemToRank_gpu, &
mesh%decomp%rankId,offset,this%interp%N,this%nvar, &
mesh%nMortars,this%nelem)
if(mesh%decomp%mpiEnabled) then
call mesh%decomp%FinalizeMPIExchangeAsync()
! Reorient small-side traces received over MPI
call MortarFlip_2D_gpu(this%mortarBuff_gpu,mesh%mortarInfo_gpu, &
mesh%decomp%elemToRank_gpu,mesh%decomp%rankId, &
this%interp%N,this%nvar,mesh%nMortars)
endif
! Small sides get the restricted big-side trace; the big side gets the L2
! projection of the small-side traces
call MortarScatter_2D_gpu(this%extBoundary_gpu,this%mortarBuff_gpu, &
this%interp%mortarR_gpu,this%interp%mortarP_gpu, &
mesh%mortarInfo_gpu,mesh%decomp%elemToRank_gpu, &
mesh%decomp%rankId,offset,this%interp%N,this%nvar, &
mesh%nMortars,this%nelem)
endsubroutine MortarExchange_MappedScalar2D