diff --git a/tests/gpu/Makefile b/tests/gpu/Makefile new file mode 100644 index 00000000..5e12f775 --- /dev/null +++ b/tests/gpu/Makefile @@ -0,0 +1,52 @@ +AMGDIR=../.. +AMGINCDIR=$(AMGDIR)/include +include $(AMGINCDIR)/Make.inc.amg4psblas +AMGMODDIR=$(AMGDIR)/modules +AMGLIBDIR=$(AMGDIR)/lib +AMG_LIBS=-L$(AMGLIBDIR) -lpsb_krylov -lamg_prec -lpsb_prec +FINCLUDES=$(FMFLAG). $(FMFLAG)$(AMGMODDIR) $(FMFLAG)$(AMGINCDIR) $(PSBLAS_INCLUDES) $(FIFLAG). + +LINKOPT= +EXEDIR=./runs + +all: amg_s_pde3d amg_d_pde3d amg_s_pde2d amg_d_pde2d + +amg_d_pde3d: amg_d_pde3d.o amg_d_genpde_mod.o amg_d_pde3d_base_mod.o amg_d_pde3d_exp_mod.o amg_d_pde3d_gauss_mod.o data_input.o + $(FLINK) $(LINKOPT) amg_d_pde3d.o amg_d_genpde_mod.o amg_d_pde3d_base_mod.o amg_d_pde3d_exp_mod.o amg_d_pde3d_gauss_mod.o data_input.o -o amg_d_pde3d $(AMG_LIBS) $(PSBLAS_LIBS) $(LDLIBS) + /bin/mv amg_d_pde3d $(EXEDIR) + +amg_s_pde3d: amg_s_pde3d.o amg_s_genpde_mod.o amg_s_pde3d_base_mod.o amg_s_pde3d_exp_mod.o amg_s_pde3d_gauss_mod.o data_input.o + $(FLINK) $(LINKOPT) amg_s_pde3d.o amg_s_genpde_mod.o amg_s_pde3d_base_mod.o amg_s_pde3d_exp_mod.o amg_s_pde3d_gauss_mod.o data_input.o -o amg_s_pde3d $(AMG_LIBS) $(PSBLAS_LIBS) $(LDLIBS) + /bin/mv amg_s_pde3d $(EXEDIR) + +amg_d_pde2d: amg_d_pde2d.o amg_d_genpde_mod.o amg_d_pde2d_base_mod.o amg_d_pde2d_exp_mod.o amg_d_pde2d_box_mod.o data_input.o + $(FLINK) $(LINKOPT) amg_d_pde2d.o amg_d_genpde_mod.o amg_d_pde2d_base_mod.o amg_d_pde2d_exp_mod.o amg_d_pde2d_box_mod.o data_input.o -o amg_d_pde2d $(AMG_LIBS) $(PSBLAS_LIBS) $(LDLIBS) + /bin/mv amg_d_pde2d $(EXEDIR) + +amg_s_pde2d: amg_s_pde2d.o amg_s_genpde_mod.o amg_s_pde2d_base_mod.o amg_s_pde2d_exp_mod.o amg_s_pde2d_box_mod.o data_input.o + $(FLINK) $(LINKOPT) amg_s_pde2d.o amg_s_genpde_mod.o amg_s_pde2d_base_mod.o amg_s_pde2d_exp_mod.o amg_s_pde2d_box_mod.o data_input.o -o amg_s_pde2d $(AMG_LIBS) $(PSBLAS_LIBS) $(LDLIBS) + /bin/mv amg_s_pde2d $(EXEDIR) + +amg_d_pde3d_rebld: amg_d_pde3d_rebld.o data_input.o + $(FLINK) $(LINKOPT) amg_d_pde3d_rebld.o data_input.o -o amg_d_pde3d_rebld $(AMG_LIBS) $(PSBLAS_LIBS) $(LDLIBS) + /bin/mv amg_d_pde3d_rebld $(EXEDIR) + +amg_d_pde3d.o amg_s_pde3d.o amg_d_pde2d.o amg_s_pde2d.o: data_input.o + +amg_d_pde3d.o: amg_d_genpde_mod.o amg_d_pde3d_base_mod.o amg_d_pde3d_exp_mod.o amg_d_pde3d_gauss_mod.o +amg_s_pde3d.o: amg_s_genpde_mod.o amg_s_pde3d_base_mod.o amg_s_pde3d_exp_mod.o amg_s_pde3d_gauss_mod.o +amg_d_pde2d.o: amg_d_genpde_mod.o amg_d_pde2d_base_mod.o amg_d_pde2d_exp_mod.o amg_d_pde2d_box_mod.o +amg_s_pde2d.o: amg_s_genpde_mod.o amg_s_pde2d_base_mod.o amg_s_pde2d_exp_mod.o amg_s_pde2d_box_mod.o + +check: all + cd runs && ./amg_d_pde2d f + else + f_ => d_null_func_3d + end if + + if (present(partition)) then + if ((1<= partition).and.(partition <= 3)) then + partition_ = partition + else + write(*,*) 'Invalid partition choice ',partition,' defaulting to 3' + partition_ = 3 + end if + else + partition_ = 3 + end if + deltah = done/(idim+2) + sqdeltah = deltah*deltah + deltah2 = 2.0_psb_dpk_* deltah + + if (present(partition)) then + if ((1<= partition).and.(partition <= 3)) then + partition_ = partition + else + write(*,*) 'Invalid partition choice ',partition,' defaulting to 3' + partition_ = 3 + end if + else + partition_ = 3 + end if + + ! initialize array descriptor and sparse matrix storage. provide an + ! estimate of the number of non zeroes + + m = (1_psb_lpk_*idim)*idim*idim + n = m + nnz = 7*((n+np-1)/np) + if(iam == psb_root_) write(psb_out_unit,'("Generating Matrix (size=",i0,")...")')n + t0 = psb_wtime() + select case(partition_) + case(1) + ! A BLOCK partition + if (present(nrl)) then + nr = nrl + else + ! + ! Using a simple BLOCK distribution. + ! + nt = (m+np-1)/np + nr = max(0,min(nt,m-(iam*nt))) + end if + + nt = nr + call psb_sum(ctxt,nt) + if (nt /= m) then + write(psb_err_unit,*) iam, 'Initialization error ',nr,nt,m + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end if + + ! + ! First example of use of CDALL: specify for each process a number of + ! contiguous rows + ! + call psb_cdall(ctxt,desc_a,info,nl=nr) + myidx = desc_a%get_global_indices() + nlr = size(myidx) + + case(2) + ! A partition defined by the user through IV + + if (present(iv)) then + if (size(iv) /= m) then + write(psb_err_unit,*) iam, 'Initialization error: wrong IV size',size(iv),m + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end if + else + write(psb_err_unit,*) iam, 'Initialization error: IV not present' + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end if + + ! + ! Second example of use of CDALL: specify for each row the + ! process that owns it + ! + call psb_cdall(ctxt,desc_a,info,vg=iv) + myidx = desc_a%get_global_indices() + nlr = size(myidx) + + case(3) + ! A 3-dimensional partition + + ! A nifty MPI function will split the process list + npdims = 0 + call mpi_dims_create(np,3,npdims,info) + npx = npdims(1) + npy = npdims(2) + npz = npdims(3) + + allocate(bndx(0:npx),bndy(0:npy),bndz(0:npz)) + ! We can reuse idx2ijk for process indices as well. + call idx2ijk(iamx,iamy,iamz,iam,npx,npy,npz,base=0) + ! Now let's split the 3D cube in hexahedra + call dist1Didx(bndx,idim,npx) + mynx = bndx(iamx+1)-bndx(iamx) + call dist1Didx(bndy,idim,npy) + myny = bndy(iamy+1)-bndy(iamy) + call dist1Didx(bndz,idim,npz) + mynz = bndz(iamz+1)-bndz(iamz) + + ! How many indices do I own? + nlr = mynx*myny*mynz + allocate(myidx(nlr)) + ! Now, let's generate the list of indices I own + nr = 0 + do i=bndx(iamx),bndx(iamx+1)-1 + do j=bndy(iamy),bndy(iamy+1)-1 + do k=bndz(iamz),bndz(iamz+1)-1 + nr = nr + 1 + call ijk2idx(myidx(nr),i,j,k,idim,idim,idim) + end do + end do + end do + if (nr /= nlr) then + write(psb_err_unit,*) iam,iamx,iamy,iamz, 'Initialization error: NR vs NLR ',& + & nr,nlr,mynx,myny,mynz + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + end if + + ! + ! Third example of use of CDALL: specify for each process + ! the set of global indices it owns. + ! + call psb_cdall(ctxt,desc_a,info,vl=myidx) + + ! + ! Specify process topology + ! + block + ! + ! Use adjcncy methods + ! + integer(psb_mpk_), allocatable :: neighbours(:) + integer(psb_mpk_) :: cnt + logical, parameter :: debug_adj=.true. + if (debug_adj.and.(np > 1)) then + cnt = 0 + allocate(neighbours(np)) + if (iamx < npx-1) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx+1,iamy,iamz,npx,npy,npz,base=0) + end if + if (iamy < npy-1) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx,iamy+1,iamz,npx,npy,npz,base=0) + end if + if (iamz < npz-1) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx,iamy,iamz+1,npx,npy,npz,base=0) + end if + if (iamx >0) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx-1,iamy,iamz,npx,npy,npz,base=0) + end if + if (iamy >0) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx,iamy-1,iamz,npx,npy,npz,base=0) + end if + if (iamz >0) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx,iamy,iamz-1,npx,npy,npz,base=0) + end if + call psb_realloc(cnt, neighbours,info) + call desc_a%set_p_adjcncy(neighbours) + !write(0,*) iam,' Check on neighbours: ',desc_a%get_p_adjcncy() + end if + end block + + case default + write(psb_err_unit,*) iam, 'Initialization error: should not get here' + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end select + + + if (info == psb_success_) call psb_spall(a,desc_a,info,nnz=nnz) + ! define rhs from boundary conditions; also build initial guess + if (info == psb_success_) call psb_geall(xv,desc_a,info) + if (info == psb_success_) call psb_geall(bv,desc_a,info) + + call psb_barrier(ctxt) + talc = psb_wtime()-t0 + + if (info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='allocation rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + + ! we build an auxiliary matrix consisting of one row at a + ! time; just a small matrix. might be extended to generate + ! a bunch of rows per call. + ! + allocate(val(20*nb),irow(20*nb),& + &icol(20*nb),stat=info) + if (info /= psb_success_ ) then + info=psb_err_alloc_dealloc_ + call psb_errpush(info,name) + goto 9999 + endif + + + ! loop over rows belonging to current process in a block + ! distribution. + + call psb_barrier(ctxt) + t1 = psb_wtime() + do ii=1, nlr,nb + ib = min(nb,nlr-ii+1) + icoeff = 1 + do k=1,ib + i=ii+k-1 + ! local matrix pointer + glob_row=myidx(i) + ! compute gridpoint coordinates + call idx2ijk(ix,iy,iz,glob_row,idim,idim,idim) + ! x, y, z coordinates + x = (ix-1)*deltah + y = (iy-1)*deltah + z = (iz-1)*deltah + zt(k) = f_(x,y,z) + ! internal point: build discretization + ! + ! term depending on (x-1,y,z) + ! + val(icoeff) = -a1(x,y,z)/sqdeltah-b1(x,y,z)/deltah2 + if (ix == 1) then + zt(k) = g(dzero,y,z)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix-1,iy,iz,idim,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + ! term depending on (x,y-1,z) + val(icoeff) = -a2(x,y,z)/sqdeltah-b2(x,y,z)/deltah2 + if (iy == 1) then + zt(k) = g(x,dzero,z)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix,iy-1,iz,idim,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + ! term depending on (x,y,z-1) + val(icoeff)=-a3(x,y,z)/sqdeltah-b3(x,y,z)/deltah2 + if (iz == 1) then + zt(k) = g(x,y,dzero)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix,iy,iz-1,idim,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + + ! term depending on (x,y,z) + val(icoeff)=(2*done)*(a1(x,y,z)+a2(x,y,z)+a3(x,y,z))/sqdeltah & + & + c(x,y,z) + call ijk2idx(icol(icoeff),ix,iy,iz,idim,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + ! term depending on (x,y,z+1) + val(icoeff)=-a3(x,y,z)/sqdeltah+b3(x,y,z)/deltah2 + if (iz == idim) then + zt(k) = g(x,y,done)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix,iy,iz+1,idim,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + ! term depending on (x,y+1,z) + val(icoeff)=-a2(x,y,z)/sqdeltah+b2(x,y,z)/deltah2 + if (iy == idim) then + zt(k) = g(x,done,z)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix,iy+1,iz,idim,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + ! term depending on (x+1,y,z) + val(icoeff)=-a1(x,y,z)/sqdeltah+b1(x,y,z)/deltah2 + if (ix==idim) then + zt(k) = g(done,y,z)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix+1,iy,iz,idim,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + + end do + call psb_spins(icoeff-1,irow,icol,val,a,desc_a,info) + if(info /= psb_success_) exit + call psb_geins(ib,myidx(ii:ii+ib-1),zt(1:ib),bv,desc_a,info) + if(info /= psb_success_) exit + zt(:)=dzero + call psb_geins(ib,myidx(ii:ii+ib-1),zt(1:ib),xv,desc_a,info) + if(info /= psb_success_) exit + end do + + tgen = psb_wtime()-t1 + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='insert rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + + deallocate(val,irow,icol) + + call psb_barrier(ctxt) + t1 = psb_wtime() + call psb_cdasb(desc_a,info) + tcdasb = psb_wtime()-t1 + call psb_barrier(ctxt) + t1 = psb_wtime() + if (info == psb_success_) then + if (present(amold)) then + call psb_spasb(a,desc_a,info,dupl=psb_dupl_err_,mold=amold) + else + call psb_spasb(a,desc_a,info,dupl=psb_dupl_err_,afmt=afmt) + end if + end if + call psb_barrier(ctxt) + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='asb rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + if (info == psb_success_) call psb_geasb(xv,desc_a,info,mold=vmold) + if (info == psb_success_) call psb_geasb(bv,desc_a,info,mold=vmold) + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='asb rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + tasb = psb_wtime()-t1 + call psb_barrier(ctxt) + ttot = psb_wtime() - t0 + + call psb_amx(ctxt,talc) + call psb_amx(ctxt,tgen) + call psb_amx(ctxt,tasb) + call psb_amx(ctxt,ttot) + if(iam == psb_root_) then + tmpfmt = a%get_fmt() + write(psb_out_unit,'("The matrix has been generated and assembled in ",a3," format.")')& + & tmpfmt + write(psb_out_unit,'("-allocation time : ",es12.5)') talc + write(psb_out_unit,'("-coeff. gen. time : ",es12.5)') tgen + write(psb_out_unit,'("-desc asbly time : ",es12.5)') tcdasb + write(psb_out_unit,'("- mat asbly time : ",es12.5)') tasb + write(psb_out_unit,'("-total time : ",es12.5)') ttot + + end if + call psb_erractionrestore(err_act) + return + +9999 call psb_error_handler(ctxt,err_act) + + return + end subroutine amg_d_gen_pde3d + + + + ! + ! subroutine to allocate and fill in the coefficient matrix and + ! the rhs. + ! + subroutine amg_d_gen_pde2d(ctxt,idim,a,bv,xv,desc_a,afmt,& + & a1,a2,b1,b2,c,g,info,f,amold,vmold,partition, nrl,iv) + use psb_base_mod + use psb_util_mod + ! + ! Discretizes the partial differential equation + ! + ! d d(u) d d(u) b1 d(u) b2 d(u) + ! - -- a1 ---- - -- a1 ---- + ----- + ------ + c u = f + ! dx dx dy dy dx dy + ! + ! with Dirichlet boundary conditions + ! u = g + ! + ! on the unit square 0<=x,y<=1. + ! + ! + ! Note that if b1=b2=c=0., the PDE is the Laplace equation. + ! + implicit none + procedure(d_func_2d) :: b1,b2,c,a1,a2,g + integer(psb_ipk_) :: idim + type(psb_dspmat_type) :: a + type(psb_d_vect_type) :: xv,bv + type(psb_desc_type) :: desc_a + integer(psb_ipk_) :: info + type(psb_ctxt_type) :: ctxt + character :: afmt*5 + procedure(d_func_2d), optional :: f + class(psb_d_base_sparse_mat), optional :: amold + class(psb_d_base_vect_type), optional :: vmold + integer(psb_ipk_), optional :: partition, nrl,iv(:) + ! Local variables. + + integer(psb_ipk_), parameter :: nb=20 + type(psb_d_csc_sparse_mat) :: acsc + type(psb_d_coo_sparse_mat) :: acoo + type(psb_d_csr_sparse_mat) :: acsr + real(psb_dpk_) :: zt(nb),x,y,z,xph,xmh,yph,ymh,zph,zmh + integer(psb_ipk_) :: nnz,nr,nlr,i,j,ii,ib,k, partition_ + integer(psb_lpk_) :: m,n,glob_row,nt + integer(psb_ipk_) :: ix,iy,iz,ia,indx_owner + ! For 2D partition + ! Note: integer control variables going directly into an MPI call + ! must be 4 bytes, i.e. psb_mpk_ + integer(psb_mpk_) :: npdims(2), npp, minfo + integer(psb_ipk_) :: npx,npy,iamx,iamy,mynx,myny + integer(psb_ipk_), allocatable :: bndx(:),bndy(:) + ! Process grid + integer(psb_ipk_) :: np, iam + integer(psb_ipk_) :: icoeff + integer(psb_lpk_), allocatable :: irow(:),icol(:),myidx(:) + real(psb_dpk_), allocatable :: val(:) + ! deltah dimension of each grid cell + ! deltat discretization time + real(psb_dpk_) :: deltah, sqdeltah, deltah2, dd + real(psb_dpk_), parameter :: rhs=0.d0,one=done,zero=0.d0 + real(psb_dpk_) :: t0, t1, t2, t3, tasb, talc, ttot, tgen, tcdasb + integer(psb_ipk_) :: err_act + procedure(d_func_2d), pointer :: f_ + character(len=20) :: name, ch_err,tmpfmt + + info = psb_success_ + name = 'create_matrix' + call psb_erractionsave(err_act) + + call psb_info(ctxt, iam, np) + + + if (present(f)) then + f_ => f + else + f_ => d_null_func_2d + end if + + deltah = done/(idim+2) + sqdeltah = deltah*deltah + deltah2 = 2.0_psb_dpk_* deltah + + + if (present(partition)) then + if ((1<= partition).and.(partition <= 3)) then + partition_ = partition + else + write(*,*) 'Invalid partition choice ',partition,' defaulting to 3' + partition_ = 3 + end if + else + partition_ = 3 + end if + + ! initialize array descriptor and sparse matrix storage. provide an + ! estimate of the number of non zeroes + + m = (1_psb_lpk_)*idim*idim + n = m + nnz = 7*((n+np-1)/np) + if(iam == psb_root_) write(psb_out_unit,'("Generating Matrix (size=",i0,")...")')n + t0 = psb_wtime() + select case(partition_) + case(1) + ! A BLOCK partition + if (present(nrl)) then + nr = nrl + else + ! + ! Using a simple BLOCK distribution. + ! + nt = (m+np-1)/np + nr = max(0,min(nt,m-(iam*nt))) + end if + + nt = nr + call psb_sum(ctxt,nt) + if (nt /= m) then + write(psb_err_unit,*) iam, 'Initialization error ',nr,nt,m + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end if + + ! + ! First example of use of CDALL: specify for each process a number of + ! contiguous rows + ! + call psb_cdall(ctxt,desc_a,info,nl=nr) + myidx = desc_a%get_global_indices() + nlr = size(myidx) + + case(2) + ! A partition defined by the user through IV + + if (present(iv)) then + if (size(iv) /= m) then + write(psb_err_unit,*) iam, 'Initialization error: wrong IV size',size(iv),m + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end if + else + write(psb_err_unit,*) iam, 'Initialization error: IV not present' + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end if + + ! + ! Second example of use of CDALL: specify for each row the + ! process that owns it + ! + call psb_cdall(ctxt,desc_a,info,vg=iv) + myidx = desc_a%get_global_indices() + nlr = size(myidx) + + case(3) + ! A 2-dimensional partition + + ! A nifty MPI function will split the process list + npdims = 0 + call mpi_dims_create(np,2,npdims,info) + npx = npdims(1) + npy = npdims(2) + + allocate(bndx(0:npx),bndy(0:npy)) + ! We can reuse idx2ijk for process indices as well. + call idx2ijk(iamx,iamy,iam,npx,npy,base=0) + ! Now let's split the 2D square in rectangles + call dist1Didx(bndx,idim,npx) + mynx = bndx(iamx+1)-bndx(iamx) + call dist1Didx(bndy,idim,npy) + myny = bndy(iamy+1)-bndy(iamy) + + ! How many indices do I own? + nlr = mynx*myny + allocate(myidx(nlr)) + ! Now, let's generate the list of indices I own + nr = 0 + do i=bndx(iamx),bndx(iamx+1)-1 + do j=bndy(iamy),bndy(iamy+1)-1 + nr = nr + 1 + call ijk2idx(myidx(nr),i,j,idim,idim) + end do + end do + if (nr /= nlr) then + write(psb_err_unit,*) iam,iamx,iamy, 'Initialization error: NR vs NLR ',& + & nr,nlr,mynx,myny + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + end if + + ! + ! Third example of use of CDALL: specify for each process + ! the set of global indices it owns. + ! + call psb_cdall(ctxt,desc_a,info,vl=myidx) + + ! + ! Specify process topology + ! + block + ! + ! Use adjcncy methods + ! + integer(psb_mpk_), allocatable :: neighbours(:) + integer(psb_mpk_) :: cnt + logical, parameter :: debug_adj=.true. + if (debug_adj.and.(np > 1)) then + cnt = 0 + allocate(neighbours(np)) + if (iamx < npx-1) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx+1,iamy,npx,npy,base=0) + end if + if (iamy < npy-1) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx,iamy+1,npx,npy,base=0) + end if + if (iamx >0) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx-1,iamy,npx,npy,base=0) + end if + if (iamy >0) then + cnt = cnt + 1 + call ijk2idx(neighbours(cnt),iamx,iamy-1,npx,npy,base=0) + end if + call psb_realloc(cnt, neighbours,info) + call desc_a%set_p_adjcncy(neighbours) + !write(0,*) iam,' Check on neighbours: ',desc_a%get_p_adjcncy() + end if + end block + + case default + write(psb_err_unit,*) iam, 'Initialization error: should not get here' + info = -1 + call psb_barrier(ctxt) + call psb_abort(ctxt) + return + end select + + + if (info == psb_success_) call psb_spall(a,desc_a,info,nnz=nnz) + ! define rhs from boundary conditions; also build initial guess + if (info == psb_success_) call psb_geall(xv,desc_a,info) + if (info == psb_success_) call psb_geall(bv,desc_a,info) + + call psb_barrier(ctxt) + talc = psb_wtime()-t0 + + if (info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='allocation rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + + ! we build an auxiliary matrix consisting of one row at a + ! time; just a small matrix. might be extended to generate + ! a bunch of rows per call. + ! + allocate(val(20*nb),irow(20*nb),& + &icol(20*nb),stat=info) + if (info /= psb_success_ ) then + info=psb_err_alloc_dealloc_ + call psb_errpush(info,name) + goto 9999 + endif + + + ! loop over rows belonging to current process in a block + ! distribution. + + call psb_barrier(ctxt) + t1 = psb_wtime() + do ii=1, nlr,nb + ib = min(nb,nlr-ii+1) + icoeff = 1 + do k=1,ib + i=ii+k-1 + ! local matrix pointer + glob_row=myidx(i) + ! compute gridpoint coordinates + call idx2ijk(ix,iy,glob_row,idim,idim) + ! x, y coordinates + x = (ix-1)*deltah + y = (iy-1)*deltah + + zt(k) = f_(x,y) + ! internal point: build discretization + ! + ! term depending on (x-1,y) + ! + val(icoeff) = -a1(x,y)/sqdeltah-b1(x,y)/deltah2 + if (ix == 1) then + zt(k) = g(dzero,y)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix-1,iy,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + ! term depending on (x,y-1) + val(icoeff) = -a2(x,y)/sqdeltah-b2(x,y)/deltah2 + if (iy == 1) then + zt(k) = g(x,dzero)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix,iy-1,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + + ! term depending on (x,y) + val(icoeff)=(2*done)*(a1(x,y) + a2(x,y))/sqdeltah + c(x,y) + call ijk2idx(icol(icoeff),ix,iy,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + ! term depending on (x,y+1) + val(icoeff)=-a2(x,y)/sqdeltah+b2(x,y)/deltah2 + if (iy == idim) then + zt(k) = g(x,done)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix,iy+1,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + ! term depending on (x+1,y) + val(icoeff)=-a1(x,y)/sqdeltah+b1(x,y)/deltah2 + if (ix==idim) then + zt(k) = g(done,y)*(-val(icoeff)) + zt(k) + else + call ijk2idx(icol(icoeff),ix+1,iy,idim,idim) + irow(icoeff) = glob_row + icoeff = icoeff+1 + endif + + end do + call psb_spins(icoeff-1,irow,icol,val,a,desc_a,info) + if(info /= psb_success_) exit + call psb_geins(ib,myidx(ii:ii+ib-1),zt(1:ib),bv,desc_a,info) + if(info /= psb_success_) exit + zt(:)=dzero + call psb_geins(ib,myidx(ii:ii+ib-1),zt(1:ib),xv,desc_a,info) + if(info /= psb_success_) exit + end do + + tgen = psb_wtime()-t1 + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='insert rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + + deallocate(val,irow,icol) + + call psb_barrier(ctxt) + t1 = psb_wtime() + call psb_cdasb(desc_a,info) + tcdasb = psb_wtime()-t1 + call psb_barrier(ctxt) + t1 = psb_wtime() + if (info == psb_success_) then + if (present(amold)) then + call psb_spasb(a,desc_a,info,dupl=psb_dupl_err_,mold=amold) + else + call psb_spasb(a,desc_a,info,dupl=psb_dupl_err_,afmt=afmt) + end if + end if + call psb_barrier(ctxt) + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='asb rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + if (info == psb_success_) call psb_geasb(xv,desc_a,info,mold=vmold) + if (info == psb_success_) call psb_geasb(bv,desc_a,info,mold=vmold) + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='asb rout.' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + tasb = psb_wtime()-t1 + call psb_barrier(ctxt) + ttot = psb_wtime() - t0 + + call psb_amx(ctxt,talc) + call psb_amx(ctxt,tgen) + call psb_amx(ctxt,tasb) + call psb_amx(ctxt,ttot) + if(iam == psb_root_) then + tmpfmt = a%get_fmt() + write(psb_out_unit,'("The matrix has been generated and assembled in ",a3," format.")')& + & tmpfmt + write(psb_out_unit,'("-allocation time : ",es12.5)') talc + write(psb_out_unit,'("-coeff. gen. time : ",es12.5)') tgen + write(psb_out_unit,'("-desc asbly time : ",es12.5)') tcdasb + write(psb_out_unit,'("- mat asbly time : ",es12.5)') tasb + write(psb_out_unit,'("-total time : ",es12.5)') ttot + + end if + call psb_erractionrestore(err_act) + return + +9999 continue + call psb_erractionrestore(err_act) + if (err_act == psb_act_abort_) then + call psb_error(ctxt) + return + end if + return + end subroutine amg_d_gen_pde2d +end module amg_d_genpde_mod diff --git a/tests/gpu/amg_d_pde3d.f90 b/tests/gpu/amg_d_pde3d.f90 new file mode 100644 index 00000000..c7526f80 --- /dev/null +++ b/tests/gpu/amg_d_pde3d.f90 @@ -0,0 +1,671 @@ +! +! +! AMG4PSBLAS version 1.0 +! Algebraic Multigrid Package +! based on PSBLAS (Parallel Sparse BLAS version 3.7) +! +! (C) Copyright 2021 +! +! Salvatore Filippone +! Pasqua D'Ambra +! Fabio Durastante +! +! Redistribution and use in source and binary forms, with or without +! modification, are permitted provided that the following conditions +! are met: +! 1. Redistributions of source code must retain the above copyright +! notice, this list of conditions and the following disclaimer. +! 2. Redistributions in binary form must reproduce the above copyright +! notice, this list of conditions, and the following disclaimer in the +! documentation and/or other materials provided with the distribution. +! 3. The name of the AMG4PSBLAS group or the names of its contributors may +! not be used to endorse or promote products derived from this +! software without specific written permission. +! +! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED +! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS +! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +! POSSIBILITY OF SUCH DAMAGE. +! +! +! +! File: amg_d_pde3d.f90 +! +! Program: amg_d_pde3d +! This sample program solves a linear system obtained by discretizing a +! PDE with Dirichlet BCs. +! +! +! The PDE is a general second order equation in 3d +! +! a1 dd(u) a2 dd(u) a3 dd(u) b1 d(u) b2 d(u) b3 d(u) +! - ------ - ------ - ------ + ----- + ------ + ------ + c u = f +! dxdx dydy dzdz dx dy dz +! +! with Dirichlet boundary conditions +! u = g +! +! on the unit cube 0<=x,y,z<=1. +! +! +! Note that if b1=b2=b3=c=0., the PDE is the Laplace equation. +! +! There are three choices available for data distribution: +! 1. A simple BLOCK distribution +! 2. A ditribution based on arbitrary assignment of indices to processes, +! typically from a graph partitioner +! 3. A 3D distribution in which the unit cube is partitioned +! into subcubes, each one assigned to a process. +! +program amg_d_pde3d + use psb_base_mod + use amg_prec_mod + use psb_krylov_mod + use psb_util_mod + use psb_gpu_mod + use data_input + use amg_d_pde3d_base_mod + use amg_d_pde3d_exp_mod + use amg_d_pde3d_gauss_mod + use amg_d_genpde_mod + implicit none + + ! input parameters + character(len=20) :: kmethd, ptype + character(len=5) :: afmt, pdecoeff + integer(psb_ipk_) :: idim + integer(psb_epk_) :: system_size + + ! miscellaneous + real(psb_dpk_) :: t1, t2, tprec, thier, tslv + + ! sparse matrix and preconditioner + type(psb_dspmat_type) :: a + type(amg_dprec_type) :: prec + ! GPU variables + type(psb_d_hlg_sparse_mat) :: agmold + type(psb_d_vect_gpu) :: vgmold + type(psb_i_vect_gpu) :: igmold + + ! descriptor + type(psb_desc_type) :: desc_a + ! dense vectors + type(psb_d_vect_type) :: x,b,r + ! parallel environment + type(psb_ctxt_type) :: ctxt + integer(psb_ipk_) :: iam, np + + ! solver parameters + integer(psb_ipk_) :: iter, itmax,itrace, istopc, irst, nlv + integer(psb_epk_) :: amatsize, precsize, descsize + real(psb_dpk_) :: err, resmx, resmxp + + ! Krylov solver data + type solverdata + character(len=40) :: kmethd ! Krylov solver + integer(psb_ipk_) :: istopc ! stopping criterion + integer(psb_ipk_) :: itmax ! maximum number of iterations + integer(psb_ipk_) :: itrace ! tracing + integer(psb_ipk_) :: irst ! restart + real(psb_dpk_) :: eps ! stopping tolerance + end type solverdata + type(solverdata) :: s_choice + + ! preconditioner data + type precdata + + ! preconditioner type + character(len=40) :: descr ! verbose description of the prec + character(len=10) :: ptype ! preconditioner type + + integer(psb_ipk_) :: outer_sweeps ! number of outer sweeps: sweeps for 1-level, + ! AMG cycles for ML + ! general AMG data + character(len=16) :: mlcycle ! AMG cycle type + integer(psb_ipk_) :: maxlevs ! maximum number of levels in AMG preconditioner + + ! AMG aggregation + character(len=16) :: aggr_prol ! aggregation type: SMOOTHED, NONSMOOTHED + character(len=16) :: par_aggr_alg ! parallel aggregation algorithm: DEC, SYMDEC + character(len=16) :: aggr_ord ! ordering for aggregation: NATURAL, DEGREE + character(len=16) :: aggr_filter ! filtering: FILTER, NO_FILTER + real(psb_dpk_) :: mncrratio ! minimum aggregation ratio + real(psb_dpk_), allocatable :: athresv(:) ! smoothed aggregation threshold vector + integer(psb_ipk_) :: thrvsz ! size of threshold vector + real(psb_dpk_) :: athres ! smoothed aggregation threshold + integer(psb_ipk_) :: csizepp ! minimum size of coarsest matrix per process + + ! AMG smoother or pre-smoother; also 1-lev preconditioner + character(len=16) :: smther ! (pre-)smoother type: BJAC, AS + integer(psb_ipk_) :: jsweeps ! (pre-)smoother / 1-lev prec. sweeps + integer(psb_ipk_) :: novr ! number of overlap layers + character(len=16) :: restr ! restriction over application of AS + character(len=16) :: prol ! prolongation over application of AS + character(len=16) :: solve ! local subsolver type: ILU, MILU, ILUT, + ! UMF, MUMPS, SLU, FWGS, BWGS, JAC + character(len=16) :: variant ! AINV variant: LLK, etc + integer(psb_ipk_) :: fill ! fill-in for incomplete LU factorization + integer(psb_ipk_) :: invfill ! Inverse fill-in for INVK + real(psb_dpk_) :: thr ! threshold for ILUT factorization + + ! AMG post-smoother; ignored by 1-lev preconditioner + character(len=16) :: smther2 ! post-smoother type: BJAC, AS + integer(psb_ipk_) :: jsweeps2 ! post-smoother sweeps + integer(psb_ipk_) :: novr2 ! number of overlap layers + character(len=16) :: restr2 ! restriction over application of AS + character(len=16) :: prol2 ! prolongation over application of AS + character(len=16) :: solve2 ! local subsolver type: ILU, MILU, ILUT, + ! UMF, MUMPS, SLU, FWGS, BWGS, JAC + character(len=16) :: variant2 ! AINV variant: LLK, etc + integer(psb_ipk_) :: fill2 ! fill-in for incomplete LU factorization + integer(psb_ipk_) :: invfill2 ! Inverse fill-in for INVK + real(psb_dpk_) :: thr2 ! threshold for ILUT factorization + + ! coarsest-level solver + character(len=16) :: cmat ! coarsest matrix layout: REPL, DIST + character(len=16) :: csolve ! coarsest-lev solver: BJAC, SLUDIST (distr. + ! mat.); UMF, MUMPS, SLU, ILU, ILUT, MILU + ! (repl. mat.) + character(len=16) :: csbsolve ! coarsest-lev local subsolver: ILU, ILUT, + ! MILU, UMF, MUMPS, SLU + integer(psb_ipk_) :: cfill ! fill-in for incomplete LU factorization + real(psb_dpk_) :: cthres ! threshold for ILUT factorization + integer(psb_ipk_) :: cjswp ! sweeps for GS or JAC coarsest-lev subsolver + + end type precdata + type(precdata) :: p_choice + + ! other variables + integer(psb_ipk_) :: info, i, k + character(len=20) :: name,ch_err + + info=psb_success_ + + + call psb_init(ctxt) + call psb_info(ctxt,iam,np) + ! + ! BEWARE: if you have NGPUS per node, the default is to + ! attach to mod(IAM,NGPUS) + ! + call psb_gpu_init(ictxt) + + if (iam < 0) then + ! This should not happen, but just in case + call psb_exit(ctxt) + stop + endif + if(psb_get_errstatus() /= 0) goto 9999 + name='amg_d_pde3d' + call psb_set_errverbosity(itwo) + ! + ! Hello world + ! + if (iam == psb_root_) then + write(*,*) 'Welcome to AMG4PSBLAS version: ',amg_version_string_ + write(*,*) 'This is the ',trim(name),' sample program' + end if + write(*,*) 'Process ',iam,' running on device: ', psb_cuda_getDevice(),' out of', psb_cuda_getDeviceCount() + write(*,*) 'Process ',iam,' device ', psb_cuda_getDevice(),' is a: ', trim(psb_gpu_DeviceName()) + + + ! + ! get parameters + ! + call get_parms(ctxt,afmt,idim,s_choice,p_choice,pdecoeff) + + ! + ! allocate and fill in the coefficient matrix, rhs and initial guess + ! + + call psb_barrier(ctxt) + t1 = psb_wtime() + select case(psb_toupper(trim(pdecoeff))) + case("CONST") + call amg_gen_pde3d(ctxt,idim,a,b,x,desc_a,afmt,& + & a1,a2,a3,b1,b2,b3,c,g,info) + case("EXP") + call amg_gen_pde3d(ctxt,idim,a,b,x,desc_a,afmt,& + & a1_exp,a2_exp,a3_exp,b1_exp,b2_exp,b3_exp,c_exp,g_exp,info) + case("GAUSS") + call amg_gen_pde3d(ctxt,idim,a,b,x,desc_a,afmt,& + & a1_gauss,a2_gauss,a3_gauss,b1_gauss,b2_gauss,b3_gauss,c_gauss,g_gauss,info) + case default + info=psb_err_from_subroutine_ + ch_err='amg_gen_pdecoeff' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end select + + + call psb_barrier(ctxt) + t2 = psb_wtime() - t1 + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='amg_gen_pde3d' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + + if (iam == psb_root_) & + & write(psb_out_unit,'("PDE Coefficients : ",a)')pdecoeff + if (iam == psb_root_) & + & write(psb_out_unit,'("Overall matrix creation time : ",es12.5)')t2 + if (iam == psb_root_) & + & write(psb_out_unit,'(" ")') + ! + ! initialize the preconditioner + ! + call prec%init(ctxt,p_choice%ptype,info) + select case(trim(psb_toupper(p_choice%ptype))) + case ('NONE','NOPREC') + ! Do nothing, keep defaults + + case ('JACOBI','L1-JACOBI','GS','FWGS','FBGS') + ! 1-level sweeps from "outer_sweeps" + call prec%set('smoother_sweeps', p_choice%jsweeps, info) + + case ('BJAC') + call prec%set('smoother_sweeps', p_choice%jsweeps, info) + call prec%set('sub_solve', p_choice%solve, info) + call prec%set('sub_fillin', p_choice%fill, info) + call prec%set('sub_iluthrs', p_choice%thr, info) + + case('AS') + call prec%set('smoother_sweeps', p_choice%jsweeps, info) + call prec%set('sub_ovr', p_choice%novr, info) + call prec%set('sub_restr', p_choice%restr, info) + call prec%set('sub_prol', p_choice%prol, info) + call prec%set('sub_solve', p_choice%solve, info) + call prec%set('sub_fillin', p_choice%fill, info) + call prec%set('sub_iluthrs', p_choice%thr, info) + + case ('ML') + ! multilevel preconditioner + + call prec%set('ml_cycle', p_choice%mlcycle, info) + call prec%set('outer_sweeps', p_choice%outer_sweeps,info) + if (p_choice%csizepp>0)& + & call prec%set('min_coarse_size_per_process', p_choice%csizepp, info) + if (p_choice%mncrratio>1)& + & call prec%set('min_cr_ratio', p_choice%mncrratio, info) + if (p_choice%maxlevs>0)& + & call prec%set('max_levs', p_choice%maxlevs, info) + if (p_choice%athres >= dzero) & + & call prec%set('aggr_thresh', p_choice%athres, info) + if (p_choice%thrvsz>0) then + do k=1,min(p_choice%thrvsz,size(prec%precv)-1) + call prec%set('aggr_thresh', p_choice%athresv(k), info,ilev=(k+1)) + end do + end if + + call prec%set('aggr_prol', p_choice%aggr_prol, info) + call prec%set('par_aggr_alg', p_choice%par_aggr_alg, info) + call prec%set('aggr_ord', p_choice%aggr_ord, info) + call prec%set('aggr_filter', p_choice%aggr_filter,info) + + + call prec%set('smoother_type', p_choice%smther, info) + call prec%set('smoother_sweeps', p_choice%jsweeps, info) + + select case (psb_toupper(p_choice%smther)) + case ('GS','BWGS','FBGS','JACOBI','L1-JACOBI','L1-FBGS') + ! do nothing + case default + call prec%set('sub_ovr', p_choice%novr, info) + call prec%set('sub_restr', p_choice%restr, info) + call prec%set('sub_prol', p_choice%prol, info) + select case(trim(psb_toupper(p_choice%solve))) + case('INVK') + call prec%set('sub_solve', p_choice%solve, info) + case('INVT') + call prec%set('sub_solve', p_choice%solve, info) + case('AINV') + call prec%set('sub_solve', p_choice%solve, info) + call prec%set('ainv_alg', p_choice%variant, info) + case default + call prec%set('sub_solve', p_choice%solve, info) + end select + + call prec%set('sub_fillin', p_choice%fill, info) + call prec%set('inv_fillin', p_choice%invfill, info) + call prec%set('sub_iluthrs', p_choice%thr, info) + end select + + if (psb_toupper(p_choice%smther2) /= 'NONE') then + call prec%set('smoother_type', p_choice%smther2, info,pos='post') + call prec%set('smoother_sweeps', p_choice%jsweeps2, info,pos='post') + select case (psb_toupper(p_choice%smther2)) + case ('GS','BWGS','FBGS','JACOBI','L1-JACOBI','L1-FBGS') + ! do nothing + case default + call prec%set('sub_ovr', p_choice%novr2, info,pos='post') + call prec%set('sub_restr', p_choice%restr2, info,pos='post') + call prec%set('sub_prol', p_choice%prol2, info,pos='post') + select case(trim(psb_toupper(p_choice%solve2))) + case('INVK') + call prec%set('sub_solve', p_choice%solve, info) + case('INVT') + call prec%set('sub_solve', p_choice%solve, info) + case('AINV') + call prec%set('sub_solve', p_choice%solve, info) + call prec%set('ainv_alg', p_choice%variant, info) + case default + call prec%set('sub_solve', p_choice%solve2, info, pos='post') + end select + + call prec%set('sub_fillin', p_choice%fill2, info,pos='post') + call prec%set('inv_fillin', p_choice%invfill2, info,pos='post') + call prec%set('sub_iluthrs', p_choice%thr2, info,pos='post') + end select + end if + + call prec%set('coarse_solve', p_choice%csolve, info) + if (psb_toupper(p_choice%csolve) == 'BJAC') & + & call prec%set('coarse_subsolve', p_choice%csbsolve, info) + call prec%set('coarse_mat', p_choice%cmat, info) + call prec%set('coarse_fillin', p_choice%cfill, info) + call prec%set('coarse_iluthrs', p_choice%cthres, info) + call prec%set('coarse_sweeps', p_choice%cjswp, info) + + end select + + ! build the preconditioner + call psb_barrier(ctxt) + t1 = psb_wtime() + call prec%hierarchy_build(a,desc_a,info) + thier = psb_wtime()-t1 + if (info /= psb_success_) then + call psb_errpush(psb_err_from_subroutine_,name,a_err='amg_hierarchy_bld') + goto 9999 + end if + call psb_barrier(ctxt) + t1 = psb_wtime() + call prec%smoothers_build(a,desc_a,info, amold=agmold, vmold=vgmold, imold=igmold) + tprec = psb_wtime()-t1 + if (info /= psb_success_) then + call psb_errpush(psb_err_from_subroutine_,name,a_err='amg_smoothers_bld') + goto 9999 + end if + + call psb_amx(ctxt, thier) + call psb_amx(ctxt, tprec) + + if(iam == psb_root_) then + write(psb_out_unit,'(" ")') + write(psb_out_unit,'("Preconditioner: ",a)') trim(p_choice%descr) + write(psb_out_unit,'("Preconditioner time: ",es12.5)')thier+tprec + write(psb_out_unit,'(" ")') + end if + call desc_a%cnv(mold=igmold) + call a%cscnv(info,mold=agmold) + call psb_geasb(x,desc_a,info,mold=vgmold) + call psb_geasb(b,desc_a,info,mold=vgmold) + + ! + ! iterative method parameters + ! + call psb_barrier(ctxt) + call prec%allocate_wrk(info) + t1 = psb_wtime() + call psb_krylov(s_choice%kmethd,a,prec,b,x,s_choice%eps,& + & desc_a,info,itmax=s_choice%itmax,iter=iter,err=err,itrace=s_choice%itrace,& + & istop=s_choice%istopc,irst=s_choice%irst) + call prec%deallocate_wrk(info) + call psb_barrier(ctxt) + tslv = psb_wtime() - t1 + + call psb_amx(ctxt,tslv) + + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='solver routine' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + + call psb_barrier(ctxt) + tslv = psb_wtime() - t1 + call psb_amx(ctxt,tslv) + + ! compute residual norms + call psb_geall(r,desc_a,info) + call r%zero() + call psb_geasb(r,desc_a,info) + call psb_geaxpby(done,b,dzero,r,desc_a,info) + call psb_spmm(-done,a,x,done,r,desc_a,info) + resmx = psb_genrm2(r,desc_a,info) + resmxp = psb_geamax(r,desc_a,info) + + amatsize = a%sizeof() + descsize = desc_a%sizeof() + precsize = prec%sizeof() + system_size = desc_a%get_global_rows() + call psb_sum(ctxt,amatsize) + call psb_sum(ctxt,descsize) + call psb_sum(ctxt,precsize) + call prec%descr(iout=psb_out_unit) + if (iam == psb_root_) then + write(psb_out_unit,'("Computed solution on ",i8," processors")') np + write(psb_out_unit,'("Linear system size : ",i12)') system_size + write(psb_out_unit,'("PDE Coefficients : ",a)') trim(pdecoeff) + write(psb_out_unit,'("Krylov method : ",a)') trim(s_choice%kmethd) + write(psb_out_unit,'("Preconditioner : ",a)') trim(p_choice%descr) + write(psb_out_unit,'("Iterations to convergence : ",i12)') iter + write(psb_out_unit,'("Relative error estimate on exit : ",es12.5)') err + write(psb_out_unit,'("Number of levels in hierarchy : ",i12)') prec%get_nlevs() + write(psb_out_unit,'("Time to build hierarchy : ",es12.5)') thier + write(psb_out_unit,'("Time to build smoothers : ",es12.5)') tprec + write(psb_out_unit,'("Total time for preconditioner : ",es12.5)') tprec+thier + write(psb_out_unit,'("Time to solve system : ",es12.5)') tslv + write(psb_out_unit,'("Time per iteration : ",es12.5)') tslv/iter + write(psb_out_unit,'("Total time : ",es12.5)') tslv+tprec+thier + write(psb_out_unit,'("Residual 2-norm : ",es12.5)') resmx + write(psb_out_unit,'("Residual inf-norm : ",es12.5)') resmxp + write(psb_out_unit,'("Total memory occupation for A : ",i12)') amatsize + write(psb_out_unit,'("Total memory occupation for DESC_A : ",i12)') descsize + write(psb_out_unit,'("Total memory occupation for PREC : ",i12)') precsize + write(psb_out_unit,'("Storage format for A : ",a )') a%get_fmt() + write(psb_out_unit,'("Storage format for DESC_A : ",a )') desc_a%get_fmt() + + end if + + ! + ! cleanup storage and exit + ! + call psb_gefree(b,desc_a,info) + call psb_gefree(x,desc_a,info) + call psb_spfree(a,desc_a,info) + call prec%free(info) + call psb_cdfree(desc_a,info) + if(info /= psb_success_) then + info=psb_err_from_subroutine_ + ch_err='free routine' + call psb_errpush(info,name,a_err=ch_err) + goto 9999 + end if + call psb_gpu_exit() + call psb_exit(ctxt) + stop + +9999 continue + call psb_error(ctxt) + +contains + ! + ! get iteration parameters from standard input + ! + ! + ! get iteration parameters from standard input + ! + subroutine get_parms(ctxt,afmt,idim,solve,prec,pdecoeff) + + implicit none + + type(psb_ctxt_type) :: ctxt + integer(psb_ipk_) :: idim + character(len=*) :: afmt + type(solverdata) :: solve + type(precdata) :: prec + character(len=*) :: pdecoeff + integer(psb_ipk_) :: iam, nm, np, inp_unit + character(len=1024) :: filename + + call psb_info(ctxt,iam,np) + + if (iam == psb_root_) then + if (command_argument_count()>0) then + call get_command_argument(1,filename) + inp_unit = 30 + open(inp_unit,file=filename,action='read',iostat=info) + if (info /= 0) then + write(psb_err_unit,*) 'Could not open file ',filename,' for input' + call psb_abort(ctxt) + stop + else + write(psb_err_unit,*) 'Opened file ',trim(filename),' for input' + end if + else + inp_unit=psb_inp_unit + end if + ! read input data + ! + call read_data(afmt,inp_unit) ! matrix storage format + call read_data(idim,inp_unit) ! Discretization grid size + call read_data(pdecoeff,inp_unit) ! PDE Coefficients + ! Krylov solver data + call read_data(solve%kmethd,inp_unit) ! Krylov solver + call read_data(solve%istopc,inp_unit) ! stopping criterion + call read_data(solve%itmax,inp_unit) ! max num iterations + call read_data(solve%itrace,inp_unit) ! tracing + call read_data(solve%irst,inp_unit) ! restart + call read_data(solve%eps,inp_unit) ! tolerance + ! preconditioner type + call read_data(prec%descr,inp_unit) ! verbose description of the prec + call read_data(prec%ptype,inp_unit) ! preconditioner type + ! First smoother / 1-lev preconditioner + call read_data(prec%smther,inp_unit) ! smoother type + call read_data(prec%jsweeps,inp_unit) ! (pre-)smoother / 1-lev prec sweeps + call read_data(prec%novr,inp_unit) ! number of overlap layers + call read_data(prec%restr,inp_unit) ! restriction over application of AS + call read_data(prec%prol,inp_unit) ! prolongation over application of AS + call read_data(prec%solve,inp_unit) ! local subsolver + call read_data(prec%variant,inp_unit) ! AINV variant + call read_data(prec%fill,inp_unit) ! fill-in for incomplete LU + call read_data(prec%invfill,inp_unit) !Inverse fill-in for INVK + call read_data(prec%thr,inp_unit) ! threshold for ILUT + ! Second smoother/ AMG post-smoother (if NONE ignored in main) + call read_data(prec%smther2,inp_unit) ! smoother type + call read_data(prec%jsweeps2,inp_unit) ! (post-)smoother sweeps + call read_data(prec%novr2,inp_unit) ! number of overlap layers + call read_data(prec%restr2,inp_unit) ! restriction over application of AS + call read_data(prec%prol2,inp_unit) ! prolongation over application of AS + call read_data(prec%solve2,inp_unit) ! local subsolver + call read_data(prec%variant2,inp_unit) ! AINV variant + call read_data(prec%fill2,inp_unit) ! fill-in for incomplete LU + call read_data(prec%invfill2,inp_unit) !Inverse fill-in for INVK + call read_data(prec%thr2,inp_unit) ! threshold for ILUT + ! general AMG data + call read_data(prec%mlcycle,inp_unit) ! AMG cycle type + call read_data(prec%outer_sweeps,inp_unit) ! number of 1lev/outer sweeps + call read_data(prec%maxlevs,inp_unit) ! max number of levels in AMG prec + call read_data(prec%csizepp,inp_unit) ! min size coarsest mat + ! aggregation + call read_data(prec%aggr_prol,inp_unit) ! aggregation type + call read_data(prec%par_aggr_alg,inp_unit) ! parallel aggregation alg + call read_data(prec%aggr_ord,inp_unit) ! ordering for aggregation + call read_data(prec%aggr_filter,inp_unit) ! filtering + call read_data(prec%mncrratio,inp_unit) ! minimum aggregation ratio + call read_data(prec%thrvsz,inp_unit) ! size of aggr thresh vector + if (prec%thrvsz > 0) then + call psb_realloc(prec%thrvsz,prec%athresv,info) + call read_data(prec%athresv,inp_unit) ! aggr thresh vector + else + read(inp_unit,*) ! dummy read to skip a record + end if + call read_data(prec%athres,inp_unit) ! smoothed aggr thresh + ! coasest-level solver + call read_data(prec%csolve,inp_unit) ! coarsest-lev solver + call read_data(prec%csbsolve,inp_unit) ! coarsest-lev subsolver + call read_data(prec%cmat,inp_unit) ! coarsest mat layout + call read_data(prec%cfill,inp_unit) ! fill-in for incompl LU + call read_data(prec%cthres,inp_unit) ! Threshold for ILUT + call read_data(prec%cjswp,inp_unit) ! sweeps for GS/JAC subsolver + if (inp_unit /= psb_inp_unit) then + close(inp_unit) + end if + end if + + call psb_bcast(ctxt,afmt) + call psb_bcast(ctxt,idim) + call psb_bcast(ctxt,pdecoeff) + + call psb_bcast(ctxt,solve%kmethd) + call psb_bcast(ctxt,solve%istopc) + call psb_bcast(ctxt,solve%itmax) + call psb_bcast(ctxt,solve%itrace) + call psb_bcast(ctxt,solve%irst) + call psb_bcast(ctxt,solve%eps) + + call psb_bcast(ctxt,prec%descr) + call psb_bcast(ctxt,prec%ptype) + + ! broadcast first (pre-)smoother / 1-lev prec data + call psb_bcast(ctxt,prec%smther) + call psb_bcast(ctxt,prec%jsweeps) + call psb_bcast(ctxt,prec%novr) + call psb_bcast(ctxt,prec%restr) + call psb_bcast(ctxt,prec%prol) + call psb_bcast(ctxt,prec%solve) + call psb_bcast(ctxt,prec%variant) + call psb_bcast(ctxt,prec%fill) + call psb_bcast(ctxt,prec%invfill) + call psb_bcast(ctxt,prec%thr) + ! broadcast second (post-)smoother + call psb_bcast(ctxt,prec%smther2) + call psb_bcast(ctxt,prec%jsweeps2) + call psb_bcast(ctxt,prec%novr2) + call psb_bcast(ctxt,prec%restr2) + call psb_bcast(ctxt,prec%prol2) + call psb_bcast(ctxt,prec%solve2) + call psb_bcast(ctxt,prec%variant2) + call psb_bcast(ctxt,prec%fill2) + call psb_bcast(ctxt,prec%invfill2) + call psb_bcast(ctxt,prec%thr2) + + ! broadcast AMG parameters + call psb_bcast(ctxt,prec%mlcycle) + call psb_bcast(ctxt,prec%outer_sweeps) + call psb_bcast(ctxt,prec%maxlevs) + + call psb_bcast(ctxt,prec%aggr_prol) + call psb_bcast(ctxt,prec%par_aggr_alg) + call psb_bcast(ctxt,prec%aggr_ord) + call psb_bcast(ctxt,prec%aggr_filter) + call psb_bcast(ctxt,prec%mncrratio) + call psb_bcast(ctxt,prec%thrvsz) + if (prec%thrvsz > 0) then + if (iam /= psb_root_) call psb_realloc(prec%thrvsz,prec%athresv,info) + call psb_bcast(ctxt,prec%athresv) + end if + call psb_bcast(ctxt,prec%athres) + + call psb_bcast(ctxt,prec%csizepp) + call psb_bcast(ctxt,prec%cmat) + call psb_bcast(ctxt,prec%csolve) + call psb_bcast(ctxt,prec%csbsolve) + call psb_bcast(ctxt,prec%cfill) + call psb_bcast(ctxt,prec%cthres) + call psb_bcast(ctxt,prec%cjswp) + + + end subroutine get_parms + +end program amg_d_pde3d diff --git a/tests/gpu/amg_d_pde3d_base_mod.f90 b/tests/gpu/amg_d_pde3d_base_mod.f90 new file mode 100644 index 00000000..76aad5cd --- /dev/null +++ b/tests/gpu/amg_d_pde3d_base_mod.f90 @@ -0,0 +1,65 @@ +module amg_d_pde3d_base_mod + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_), save, private :: epsilon=done/80 +contains + subroutine pde_set_parm(dat) + real(psb_dpk_), intent(in) :: dat + epsilon = dat + end subroutine pde_set_parm + ! + ! functions parametrizing the differential equation + ! + function b1(x,y,z) + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_) :: b1 + real(psb_dpk_), intent(in) :: x,y,z + b1=done/sqrt(3.0_psb_dpk_) + end function b1 + function b2(x,y,z) + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_) :: b2 + real(psb_dpk_), intent(in) :: x,y,z + b2=done/sqrt(3.0_psb_dpk_) + end function b2 + function b3(x,y,z) + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_) :: b3 + real(psb_dpk_), intent(in) :: x,y,z + b3=done/sqrt(3.0_psb_dpk_) + end function b3 + function c(x,y,z) + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_) :: c + real(psb_dpk_), intent(in) :: x,y,z + c=dzero + end function c + function a1(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a1 + real(psb_dpk_), intent(in) :: x,y,z + a1=epsilon + end function a1 + function a2(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a2 + real(psb_dpk_), intent(in) :: x,y,z + a2=epsilon + end function a2 + function a3(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a3 + real(psb_dpk_), intent(in) :: x,y,z + a3=epsilon + end function a3 + function g(x,y,z) + use psb_base_mod, only : psb_dpk_, done, dzero + real(psb_dpk_) :: g + real(psb_dpk_), intent(in) :: x,y,z + g = dzero + if (x == done) then + g = done + else if (x == dzero) then + g = done + end if + end function g +end module amg_d_pde3d_base_mod diff --git a/tests/gpu/amg_d_pde3d_exp_mod.f90 b/tests/gpu/amg_d_pde3d_exp_mod.f90 new file mode 100644 index 00000000..fdbb2970 --- /dev/null +++ b/tests/gpu/amg_d_pde3d_exp_mod.f90 @@ -0,0 +1,65 @@ +module amg_d_pde3d_exp_mod + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_), save, private :: epsilon=done/160 +contains + subroutine pde_set_parm(dat) + real(psb_dpk_), intent(in) :: dat + epsilon = dat + end subroutine pde_set_parm + ! + ! functions parametrizing the differential equation + ! + function b1_exp(x,y,z) + use psb_base_mod, only : psb_dpk_, dzero + real(psb_dpk_) :: b1_exp + real(psb_dpk_), intent(in) :: x,y,z + b1_exp=dzero/sqrt(3.0_psb_dpk_) + end function b1_exp + function b2_exp(x,y,z) + use psb_base_mod, only : psb_dpk_, dzero + real(psb_dpk_) :: b2_exp + real(psb_dpk_), intent(in) :: x,y,z + b2_exp=dzero/sqrt(3.0_psb_dpk_) + end function b2_exp + function b3_exp(x,y,z) + use psb_base_mod, only : psb_dpk_, dzero + real(psb_dpk_) :: b3_exp + real(psb_dpk_), intent(in) :: x,y,z + b3_exp=dzero/sqrt(3.0_psb_dpk_) + end function b3_exp + function c_exp(x,y,z) + use psb_base_mod, only : psb_dpk_, dzero + real(psb_dpk_) :: c_exp + real(psb_dpk_), intent(in) :: x,y,z + c_exp=dzero + end function c_exp + function a1_exp(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a1_exp + real(psb_dpk_), intent(in) :: x,y,z + a1_exp=epsilon*exp(-(x+y+z)) + end function a1_exp + function a2_exp(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a2_exp + real(psb_dpk_), intent(in) :: x,y,z + a2_exp=epsilon*exp(-(x+y+z)) + end function a2_exp + function a3_exp(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a3_exp + real(psb_dpk_), intent(in) :: x,y,z + a3_exp=epsilon*exp(-(x+y+z)) + end function a3_exp + function g_exp(x,y,z) + use psb_base_mod, only : psb_dpk_, done, dzero + real(psb_dpk_) :: g_exp + real(psb_dpk_), intent(in) :: x,y,z + g_exp = dzero + if (x == done) then + g_exp = done + else if (x == dzero) then + g_exp = done + end if + end function g_exp +end module amg_d_pde3d_exp_mod diff --git a/tests/gpu/amg_d_pde3d_gauss_mod.f90 b/tests/gpu/amg_d_pde3d_gauss_mod.f90 new file mode 100644 index 00000000..3787c76a --- /dev/null +++ b/tests/gpu/amg_d_pde3d_gauss_mod.f90 @@ -0,0 +1,65 @@ +module amg_d_pde3d_gauss_mod + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_), save, private :: epsilon=done/80 +contains + subroutine pde_set_parm(dat) + real(psb_dpk_), intent(in) :: dat + epsilon = dat + end subroutine pde_set_parm + ! + ! functions parametrizing the differential equation + ! + function b1_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_) :: b1_gauss + real(psb_dpk_), intent(in) :: x,y,z + b1_gauss=done/sqrt(3.0_psb_dpk_)-2*x*exp(-(x**2+y**2+z**2)) + end function b1_gauss + function b2_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_) :: b2_gauss + real(psb_dpk_), intent(in) :: x,y,z + b2_gauss=done/sqrt(3.0_psb_dpk_)-2*y*exp(-(x**2+y**2+z**2)) + end function b2_gauss + function b3_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_, done + real(psb_dpk_) :: b3_gauss + real(psb_dpk_), intent(in) :: x,y,z + b3_gauss=done/sqrt(3.0_psb_dpk_)-2*z*exp(-(x**2+y**2+z**2)) + end function b3_gauss + function c_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_, dzero + real(psb_dpk_) :: c_gauss + real(psb_dpk_), intent(in) :: x,y,z + c=dzero + end function c_gauss + function a1_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a1_gauss + real(psb_dpk_), intent(in) :: x,y,z + a1_gauss=epsilon*exp(-(x**2+y**2+z**2)) + end function a1_gauss + function a2_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a2_gauss + real(psb_dpk_), intent(in) :: x,y,z + a2_gauss=epsilon*exp(-(x**2+y**2+z**2)) + end function a2_gauss + function a3_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_ + real(psb_dpk_) :: a3_gauss + real(psb_dpk_), intent(in) :: x,y,z + a3_gauss=epsilon*exp(-(x**2+y**2+z**2)) + end function a3_gauss + function g_gauss(x,y,z) + use psb_base_mod, only : psb_dpk_, done, dzero + real(psb_dpk_) :: g_gauss + real(psb_dpk_), intent(in) :: x,y,z + g_gauss = dzero + if (x == done) then + g_gauss = done + else if (x == dzero) then + g_gauss = done + end if + end function g_gauss +end module amg_d_pde3d_gauss_mod diff --git a/tests/gpu/runs/amg_gpu_pde3d.inp b/tests/gpu/runs/amg_gpu_pde3d.inp new file mode 100644 index 00000000..a819710f --- /dev/null +++ b/tests/gpu/runs/amg_gpu_pde3d.inp @@ -0,0 +1,55 @@ +%%%%%%%%%%% General arguments % Lines starting with % are ignored. +CSR ! Storage format CSR COO JAD +0080 ! IDIM; domain size. Linear system size is IDIM**3 +CONST ! PDECOEFF: CONST, EXP, GAUSS Coefficients of the PDE +FCG ! Iterative method: BiCGSTAB BiCGSTABL BiCG CG CGS FCG GCR RGMRES +2 ! ISTOPC +00500 ! ITMAX +1 ! ITRACE +30 ! IRST (restart for RGMRES and BiCGSTABL) +1.d-6 ! EPS +%%%%%%%%%%% Main preconditioner choices %%%%%%%%%%%%%%%% +ML-VCYCLE-BJAC-D-BJAC ! Longer descriptive name for preconditioner (up to 20 chars) +ML ! Preconditioner type: NONE JACOBI GS FBGS BJAC AS ML +%%%%%%%%%%% First smoother (for all levels but coarsest) %%%%%%%%%%%%%%%% +BJAC ! Smoother type JACOBI FBGS GS BWGS BJAC AS. For 1-level, repeats previous. +1 ! Number of sweeps for smoother +0 ! Number of overlap layers for AS preconditioner +HALO ! AS restriction operator: NONE HALO +NONE ! AS prolongation operator: NONE SUM AVG +INVK ! Subdomain solver for BJAC/AS: JACOBI GS BGS ILU ILUT MILU MUMPS SLU UMF +LLK ! AINV variant +0 ! Fill level P for ILU(P) and ILU(T,P) +1 ! Inverse Fill level P for INVK +1.d-4 ! Threshold T for ILU(T,P) +%%%%%%%%%%% Second smoother, always ignored for non-ML %%%%%%%%%%%%%%%% +NONE ! Second (post) smoother, ignored if NONE +1 ! Number of sweeps for (post) smoother +0 ! Number of overlap layers for AS preconditioner +HALO ! AS restriction operator: NONE HALO +NONE ! AS prolongation operator: NONE SUM AVG +ILU ! Subdomain solver for BJAC/AS: JACOBI GS BGS ILU ILUT MILU MUMPS SLU UMF +LLK ! AINV variant +0 ! Fill level P for ILU(P) and ILU(T,P) +8 ! Inverse Fill level P for INVK +1.d-4 ! Threshold T for ILU(T,P) +%%%%%%%%%%% Multilevel parameters %%%%%%%%%%%%%%%% +VCYCLE ! Type of multilevel CYCLE: VCYCLE WCYCLE KCYCLE MULT ADD +1 ! Number of outer sweeps for ML +-3 ! Max Number of levels in a multilevel preconditioner; if <0, lib default +-3 ! Target coarse matrix size per process; if <0, lib default +SMOOTHED ! Type of aggregation: SMOOTHED UNSMOOTHED +DEC ! Parallel aggregation: DEC, SYMDEC +NATURAL ! Ordering of aggregation NATURAL DEGREE +NOFILTER ! Filtering of matrix: FILTER NOFILTER +-1.5 ! Coarsening ratio, if < 0 use library default +-2 ! Number of thresholds in vector, next line ignored if <= 0 +0.05 0.025 ! Thresholds +-0.0100d0 ! Smoothed aggregation threshold, ignored if < 0 +%%%%%%%%%%% Coarse level solver %%%%%%%%%%%%%%%% +BJAC ! Coarsest-level solver: MUMPS UMF SLU SLUDIST JACOBI GS BJAC +INVK ! Coarsest-level subsolver for BJAC: ILU ILUT MILU UMF MUMPS SLU +DIST ! Coarsest-level matrix distribution: DIST REPL +1 ! Coarsest-level fillin P for ILU(P) and ILU(T,P) +1.d-4 ! Coarsest-level threshold T for ILU(T,P) +1 ! Number of sweeps for JACOBI/GS/BJAC coarsest-level solver