|
|
|
|
@ -1,10 +1,10 @@
|
|
|
|
|
!!$
|
|
|
|
|
!!$
|
|
|
|
|
!!$ MLD2P4 version 1.0
|
|
|
|
|
!!$ MLD2P4 version 1.1
|
|
|
|
|
!!$ MultiLevel Domain Decomposition Parallel Preconditioners Package
|
|
|
|
|
!!$ based on PSBLAS (Parallel Sparse BLAS version 2.2)
|
|
|
|
|
!!$ based on PSBLAS (Parallel Sparse BLAS version 2.3.1)
|
|
|
|
|
!!$
|
|
|
|
|
!!$ (C) Copyright 2008
|
|
|
|
|
!!$ (C) Copyright 2008,2009
|
|
|
|
|
!!$
|
|
|
|
|
!!$ Salvatore Filippone University of Rome Tor Vergata
|
|
|
|
|
!!$ Alfredo Buttari University of Rome Tor Vergata
|
|
|
|
|
@ -42,7 +42,7 @@
|
|
|
|
|
! Version: complex
|
|
|
|
|
!
|
|
|
|
|
! This routine builds a coarse-level matrix A_C from a fine-level matrix A
|
|
|
|
|
! by using a Galerkin approach, i.e.
|
|
|
|
|
! by using the Galerkin approach, i.e.
|
|
|
|
|
!
|
|
|
|
|
! A_C = P_C^T A P_C,
|
|
|
|
|
!
|
|
|
|
|
@ -59,7 +59,7 @@
|
|
|
|
|
! of A, and omega is a suitable smoothing parameter. An estimate of the spectral
|
|
|
|
|
! radius of D^(-1)A, to be used in the computation of omega, is provided,
|
|
|
|
|
! according to the value of p%iprcparm(mld_aggr_omega_alg_), specified by the user
|
|
|
|
|
! through mld_dprecinit and mld_dprecset.
|
|
|
|
|
! through mld_cprecinit and mld_cprecset.
|
|
|
|
|
!
|
|
|
|
|
! This routine can also build A_C according to a "bizarre" aggregation algorithm,
|
|
|
|
|
! using a "naive" prolongator proposed by the authors of MLD2P4. However, this
|
|
|
|
|
@ -68,7 +68,7 @@
|
|
|
|
|
!
|
|
|
|
|
! The coarse-level matrix A_C is distributed among the parallel processes or
|
|
|
|
|
! replicated on each of them, according to the value of p%iprcparm(mld_coarse_mat_),
|
|
|
|
|
! specified by the user through mld_dprecinit and mld_dprecset.
|
|
|
|
|
! specified by the user through mld_cprecinit and mld_cprecset.
|
|
|
|
|
!
|
|
|
|
|
! For more details see
|
|
|
|
|
! M. Brezina and P. Vanek, A black-box iterative solver based on a
|
|
|
|
|
@ -84,9 +84,17 @@
|
|
|
|
|
! desc_a - type(psb_desc_type), input.
|
|
|
|
|
! The communication descriptor of the fine-level matrix.
|
|
|
|
|
! p - type(mld_conelev_type), input/output.
|
|
|
|
|
! The one-level preconditioner data structure containing the local
|
|
|
|
|
! part of the base preconditioner to be built as well as the
|
|
|
|
|
! aggregate matrices.
|
|
|
|
|
! The 'one-level' data structure that will contain the local
|
|
|
|
|
! part of the matrix to be built as well as the information
|
|
|
|
|
! concerning the prolongator and its transpose.
|
|
|
|
|
! ilaggr - integer, dimension(:), allocatable.
|
|
|
|
|
! The mapping between the row indices of the coarse-level
|
|
|
|
|
! matrix and the row indices of the fine-level matrix.
|
|
|
|
|
! ilaggr(i)=j means that node i in the adjacency graph
|
|
|
|
|
! of the fine-level matrix is mapped onto node j in the
|
|
|
|
|
! adjacency graph of the coarse-level matrix.
|
|
|
|
|
! nlaggr - integer, dimension(:), allocatable.
|
|
|
|
|
! nlaggr(i) contains the aggregates held by process i.
|
|
|
|
|
! info - integer, output.
|
|
|
|
|
! Error code.
|
|
|
|
|
!
|
|
|
|
|
@ -116,10 +124,10 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
& naggr, nzl,naggrm1,naggrp1, i, j, k, jd, icolF
|
|
|
|
|
integer ::ictxt,np,me, err_act, icomm
|
|
|
|
|
character(len=20) :: name
|
|
|
|
|
type(psb_cspmat_type) :: am1,am2, AF
|
|
|
|
|
type(psb_cspmat_type) :: am1,am2, af
|
|
|
|
|
type(psb_cspmat_type) :: am3,am4
|
|
|
|
|
complex(psb_spk_), allocatable :: adiag(:)
|
|
|
|
|
logical :: ml_global_nmb
|
|
|
|
|
logical :: ml_global_nmb, filter_mat
|
|
|
|
|
integer :: debug_level, debug_unit
|
|
|
|
|
integer, parameter :: ncmax=16
|
|
|
|
|
real(psb_spk_) :: omega, anorm, tmp, dg, theta
|
|
|
|
|
@ -168,6 +176,8 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
& ( (p%iprcparm(mld_aggr_kind_) == mld_biz_prol_).and.&
|
|
|
|
|
& (p%iprcparm(mld_coarse_mat_) == mld_repl_mat_)) )
|
|
|
|
|
|
|
|
|
|
filter_mat = (p%iprcparm(mld_aggr_filter_) == mld_filter_mat_)
|
|
|
|
|
|
|
|
|
|
if (ml_global_nmb) then
|
|
|
|
|
ilaggr(1:nrow) = ilaggr(1:nrow) + naggrm1
|
|
|
|
|
call psb_halo(ilaggr,desc_a,info)
|
|
|
|
|
@ -190,7 +200,7 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
! Get diagonal D
|
|
|
|
|
! Get the diagonal D
|
|
|
|
|
call psb_sp_getdiag(a,adiag,info)
|
|
|
|
|
if (info == 0) &
|
|
|
|
|
& call psb_halo(adiag,desc_a,info)
|
|
|
|
|
@ -243,10 +253,9 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
& write(debug_unit,*) me,' ',trim(name),&
|
|
|
|
|
& ' Initial copies done.'
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
!********************************+
|
|
|
|
|
! building the filtered matrix af from A
|
|
|
|
|
if (filter_mat) then
|
|
|
|
|
!
|
|
|
|
|
! Build the filtered matrix Af from A
|
|
|
|
|
!
|
|
|
|
|
call psb_spcnv(a,af,info,afmt='csr',dupl=psb_dupl_add_)
|
|
|
|
|
|
|
|
|
|
@ -262,12 +271,12 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
|
|
|
|
|
enddo
|
|
|
|
|
if (jd == -1) then
|
|
|
|
|
write(0,*) 'Wrong input: we need the diagonal!!!!'
|
|
|
|
|
write(0,*) 'Wrong input: we need the diagonal!!!!', i
|
|
|
|
|
else
|
|
|
|
|
af%aspk(jd)=af%aspk(jd)-tmp
|
|
|
|
|
end if
|
|
|
|
|
enddo
|
|
|
|
|
! Ora eliminiamo i termini che sono stato azzerati
|
|
|
|
|
! Take out zeroed terms
|
|
|
|
|
call psb_spcnv(af,info,afmt='coo')
|
|
|
|
|
k = 0
|
|
|
|
|
do j=1,psb_sp_get_nnzeros(af)
|
|
|
|
|
@ -281,7 +290,7 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
!!$ write(debug_unit,*) me,' ',trim(name),' Non zeros from filtered matrix:',k,af%m,af%k
|
|
|
|
|
call psb_sp_setifld(k,psb_nnz_,af,info)
|
|
|
|
|
call psb_spcnv(af,info,afmt='csr')
|
|
|
|
|
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
do i=1,size(adiag)
|
|
|
|
|
@ -291,10 +300,13 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
adiag(i) = cone
|
|
|
|
|
end if
|
|
|
|
|
end do
|
|
|
|
|
call psb_sp_scal(af,adiag,info)
|
|
|
|
|
|
|
|
|
|
if (filter_mat) call psb_sp_scal(af,adiag,info)
|
|
|
|
|
|
|
|
|
|
call psb_sp_scal(am3,adiag,info)
|
|
|
|
|
if (info /= 0) goto 9999
|
|
|
|
|
!*******************************************
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if (p%iprcparm(mld_aggr_omega_alg_) == mld_eig_est_) then
|
|
|
|
|
|
|
|
|
|
if (p%iprcparm(mld_aggr_eig_) == mld_max_norm_) then
|
|
|
|
|
@ -302,7 +314,7 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
if (p%iprcparm(mld_aggr_kind_) == mld_biz_prol_) then
|
|
|
|
|
|
|
|
|
|
!
|
|
|
|
|
! This only works with CSR.
|
|
|
|
|
! This only works with CSR
|
|
|
|
|
!
|
|
|
|
|
if (psb_toupper(am3%fida)=='CSR') then
|
|
|
|
|
anorm = szero
|
|
|
|
|
@ -352,7 +364,10 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
! %
|
|
|
|
|
if (filter_mat) then
|
|
|
|
|
!
|
|
|
|
|
! Build the smoothed prolongator using the filtered matrix
|
|
|
|
|
!
|
|
|
|
|
if (psb_toupper(af%fida)=='CSR') then
|
|
|
|
|
do i=1,af%m
|
|
|
|
|
do j=af%ia2(i),af%ia2(i+1)-1
|
|
|
|
|
@ -374,8 +389,8 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
!
|
|
|
|
|
! Symbmm90 does the allocation for its result.
|
|
|
|
|
!
|
|
|
|
|
! am1 = (i-wDA)Ptilde
|
|
|
|
|
! Doing it this way means to consider diag(Ai)
|
|
|
|
|
! am1 = (I-w*D*Af)Ptilde
|
|
|
|
|
! Doing it this way means to consider diag(Af_i)
|
|
|
|
|
!
|
|
|
|
|
!
|
|
|
|
|
call psb_symbmm(af,am4,am1,info)
|
|
|
|
|
@ -389,6 +404,48 @@ subroutine mld_caggrmat_smth_asb(a,desc_a,ilaggr,nlaggr,p,info)
|
|
|
|
|
if (debug_level >= psb_debug_outer_) &
|
|
|
|
|
& write(debug_unit,*) me,' ',trim(name),&
|
|
|
|
|
& 'Done NUMBMM 1'
|
|
|
|
|
else
|
|
|
|
|
!
|
|
|
|
|
! Build the smoothed prolongator using the original matrix
|
|
|
|
|
!
|
|
|
|
|
if (psb_toupper(am3%fida)=='CSR') then
|
|
|
|
|
do i=1,am3%m
|
|
|
|
|
do j=am3%ia2(i),am3%ia2(i+1)-1
|
|
|
|
|
if (am3%ia1(j) == i) then
|
|
|
|
|
am3%aspk(j) = cone - omega*am3%aspk(j)
|
|
|
|
|
else
|
|
|
|
|
am3%aspk(j) = - omega*am3%aspk(j)
|
|
|
|
|
end if
|
|
|
|
|
end do
|
|
|
|
|
end do
|
|
|
|
|
else
|
|
|
|
|
call psb_errpush(4001,name,a_err='Invalid AM3 storage format')
|
|
|
|
|
goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
if (debug_level >= psb_debug_outer_) &
|
|
|
|
|
& write(debug_unit,*) me,' ',trim(name),&
|
|
|
|
|
& 'Done gather, going for SYMBMM 1'
|
|
|
|
|
!
|
|
|
|
|
! Symbmm90 does the allocation for its result.
|
|
|
|
|
!
|
|
|
|
|
! am1 = (I-w*D*A)Ptilde
|
|
|
|
|
! Doing it this way means to consider diag(A_i)
|
|
|
|
|
!
|
|
|
|
|
!
|
|
|
|
|
call psb_symbmm(am3,am4,am1,info)
|
|
|
|
|
if(info /= 0) then
|
|
|
|
|
call psb_errpush(4010,name,a_err='symbmm 1')
|
|
|
|
|
goto 9999
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
call psb_numbmm(am3,am4,am1)
|
|
|
|
|
|
|
|
|
|
if (debug_level >= psb_debug_outer_) &
|
|
|
|
|
& write(debug_unit,*) me,' ',trim(name),&
|
|
|
|
|
& 'Done NUMBMM 1'
|
|
|
|
|
|
|
|
|
|
end if
|
|
|
|
|
|
|
|
|
|
call psb_sp_free(am4,info)
|
|
|
|
|
if(info /= 0) then
|
|
|
|
|
|
Salvatore Filippone
16 years ago
