You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
psblas3/base/tools/psb_dcdovr.f90

690 lines
21 KiB
Fortran

!!$
!!$ Parallel Sparse BLAS v2.0
!!$ (C) Copyright 2006 Salvatore Filippone University of Rome Tor Vergata
!!$ Alfredo Buttari University of Rome Tor Vergata
!!$
!!$ 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 PSBLAS 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 PSBLAS 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: psb_cdovr.f90
!
! Subroutine: psb_cdovr
! This routine takes a matrix A with its descriptor, and builds the
! auxiliary descriptor corresponding to the number of overlap levels
! specified on input. It really is just a size estimation/allocation
! front end for <psb_cdovrbld>.
!
! Parameters:
! a - type(<psb_dspmat_type>). The input sparse matrix.
! desc_a - type(<psb_desc_type>). The input communication descriptor.
! norv - integer. The number of overlap levels.
! desc_ov - type(<psb_desc_type>). The auxiliary output communication
! descriptor.
! info - integer. Eventually returns an error code.
!
Subroutine psb_dcdovr(a,desc_a,novr,desc_ov,info)
use psb_serial_mod
use psb_descriptor_type
Use psb_prec_type
Use psb_prec_mod
use psb_error_mod
use psb_penv_mod
use psb_tools_mod
use psb_realloc_mod
use psi_mod
use mpi
Implicit None
! .. Array Arguments ..
integer, intent(in) :: novr
Type(psb_dspmat_type), Intent(in) :: a
Type(psb_desc_type), Intent(in) :: desc_a
Type(psb_desc_type), Intent(inout) :: desc_ov
integer, intent(out) :: info
integer icomm, err_act
! .. Local Scalars ..
Integer :: i, j, k, np, me,m,nnzero,&
& ictxt, lovr, lworks,lworkr, n_row,n_col, int_err(5),&
& index_dim,elem_dim, l_tmp_ovr_idx,l_tmp_halo, nztot,nhalo
Integer :: counter,counter_h, counter_o, counter_e,idx,gidx,proc,n_elem_recv,&
& n_elem_send,tot_recv,tot_elem,&
& counter_t,n_elem,i_ovr,jj,proc_id,isz, mglob, glx, &
& idxr, idxs, lx, iszr, iszs, nxch, nsnd, nrcv,lidx
type(psb_dspmat_type) :: blk
Integer, allocatable :: tmp_halo(:),tmp_ovr_idx(:)
Integer,allocatable :: halo(:),works(:),workr(:),t_halo_in(:),&
& t_halo_out(:),temp(:),maskr(:)
Integer,allocatable :: brvindx(:),rvsz(:), bsdindx(:),sdsz(:)
Logical,Parameter :: debug=.false.
character(len=20) :: name, ch_err
name='psb_cdovr'
info = 0
call psb_erractionsave(err_act)
ictxt=psb_cd_get_context(desc_a)
Call psb_info(ictxt, me, np)
If(debug) Write(0,*)'in psb_cdovr',novr
m = psb_cd_get_local_rows(desc_a)
nnzero = Size(a%aspk)
n_row = psb_cd_get_local_rows(desc_a)
n_col = psb_cd_get_local_cols(desc_a)
nhalo = n_col-m
If(debug) Write(0,*)'IN CDOVR1',novr ,m,nnzero,n_col
if (novr<0) then
info=10
int_err(1)=1
int_err(2)=novr
call psb_errpush(info,name,i_err=int_err)
goto 9999
endif
if (debug) write(0,*) 'Calling desccpy'
call psb_cdcpy(desc_a,desc_ov,info)
if (info /= 0) then
info=4010
ch_err='psb_cdcpy'
call psb_errpush(info,name,a_err=ch_err)
goto 9999
end if
if (debug) write(0,*) 'From desccpy'
if (novr==0) then
!
! Just copy the input.
!
return
endif
call psb_get_mpicomm(ictxt,icomm )
If(debug)then
Write(0,*)'BEGIN cdovr',me,nhalo
call psb_barrier(ictxt)
endif
!
! Ok, since we are only estimating, do it as follows:
! LOVR= (NNZ/NROW)*N_HALO*NOVR This assumes that the local average
! nonzeros per row is the same as the global.
!
nztot = psb_sp_get_nnzeros(a)
if (nztot>0) then
lovr = ((nztot+m-1)/m)*nhalo*novr
lworks = ((nztot+m-1)/m)*nhalo
lworkr = ((nztot+m-1)/m)*nhalo
else
info=-1
call psb_errpush(info,name)
goto 9999
endif
If(debug)Write(0,*)'ovr_est done',me,novr,lovr
index_dim = size(desc_a%halo_index)
elem_dim = size(desc_a%halo_index)
l_tmp_ovr_idx = novr*(3*Max(2*index_dim,1)+1)
l_tmp_halo = novr*(3*Size(desc_a%halo_index))
call psb_cd_set_bld(desc_ov,info)
desc_ov%matrix_data(psb_ovl_state_)=psb_cd_ovl_bld_
If(debug) then
Write(0,*)'Start cdovrbld',me,lworks,lworkr
call psb_barrier(ictxt)
endif
Allocate(brvindx(np+1),rvsz(np),sdsz(np),bsdindx(np+1),stat=info)
if (info /= 0) then
call psb_errpush(4010,name,a_err='Allocate')
goto 9999
end if
Allocate(works(lworks),workr(lworkr),t_halo_in(l_tmp_halo),&
& t_halo_out(l_tmp_halo), temp(lworkr),stat=info)
if (info /= 0) then
call psb_errpush(4010,name,a_err='Allocate')
goto 9999
end if
call psb_sp_all(blk,max(lworks,lworkr),info)
if (info /= 0) then
info=4010
ch_err='psb_sp_all'
call psb_errpush(info,name,a_err=ch_err)
goto 9999
end if
blk%fida='COO'
Allocate(tmp_ovr_idx(l_tmp_ovr_idx),tmp_halo(l_tmp_halo),&
& halo(size(desc_a%halo_index)),stat=info)
if (info /= 0) then
call psb_errpush(4010,name,a_err='Allocate')
goto 9999
end if
halo(:) = desc_a%halo_index(:)
desc_ov%ovrlap_elem(:) = -1
tmp_ovr_idx(:) = -1
tmp_halo(:) = -1
counter_e = 1
tot_recv = 0
counter_h = 1
counter_o = 1
! Init overlap with desc_a%ovrlap (if any)
counter = 1
Do While (desc_a%ovrlap_index(counter) /= -1)
proc = desc_a%ovrlap_index(counter+psb_proc_id_)
n_elem_recv = desc_a%ovrlap_index(counter+psb_n_elem_recv_)
n_elem_send = desc_a%ovrlap_index(counter+n_elem_recv+psb_n_elem_send_)
Do j=0,n_elem_recv-1
idx = desc_a%ovrlap_index(counter+psb_elem_recv_+j)
If(idx > Size(desc_ov%loc_to_glob)) then
info=-3
call psb_errpush(info,name)
goto 9999
endif
gidx = desc_ov%loc_to_glob(idx)
call psb_check_size((counter_o+3),tmp_ovr_idx,info,pad=-1)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
tmp_ovr_idx(counter_o)=proc
tmp_ovr_idx(counter_o+1)=1
tmp_ovr_idx(counter_o+2)=gidx
tmp_ovr_idx(counter_o+3)=-1
counter_o=counter_o+3
end Do
counter=counter+n_elem_recv+n_elem_send+2
end Do
!
! A picture is in order to understand what goes on here.
! I is the internal part; H is halo, R row, C column. The final
! matrix with N levels of overlap looks like this
!
! I | Hc1 | 0 | 0 |
! -------|-----|-----|-----|
! Hr1 | Hd1 | Hc2 | 0 |
! -------|-----|-----|-----|
! 0 | Hr2 | Hd2 | Hc2 |
! -------|-----|-----|-----|
! 0 | 0 | Hr3 | Hd3 | Hc3
!
! At the start we already have I and Hc1, so we know the row
! indices that will make up Hr1, and also who owns them. As we
! actually get those rows, we receive the column indices in Hc2;
! these define the row indices for Hr2, and so on. When we have
! reached the desired level HrN, we may ignore HcN.
!
!
Do i_ovr = 1, novr
if (debug) write(0,*) me,'Running on overlap level ',i_ovr,' of ',novr
!
! At this point, halo contains a valid halo corresponding to the
! matrix enlarged with the elements in the frontier for I_OVR-1.
! At the start, this is just the halo for A; the rows for indices in
! the first halo will contain column indices defining the second halo
! level and so on.
!
bsdindx(:) = 0
sdsz(:) = 0
brvindx(:) = 0
rvsz(:) = 0
idxr = 0
idxs = 0
counter = 1
counter_t = 1
Do While (halo(counter) /= -1)
tot_elem=0
proc=halo(counter+psb_proc_id_)
n_elem_recv=halo(counter+psb_n_elem_recv_)
n_elem_send=halo(counter+n_elem_recv+psb_n_elem_send_)
If ((counter+n_elem_recv+n_elem_send) > Size(halo)) then
info = -1
call psb_errpush(info,name)
goto 9999
end If
tot_recv=tot_recv+n_elem_recv
if (debug) write(0,*) me,' CDOVRBLD tot_recv:',proc,n_elem_recv,tot_recv
!
!
! The format of the halo vector exists in two forms: 1. Temporary
! 2. Assembled. In this loop we are using the (assembled) halo_in and
! copying it into (temporary) halo_out; this is because tmp_halo will
! be enlarged with the new column indices received, and will reassemble
! everything for the next iteration.
!
!
! add recv elements in halo_index into ovrlap_index
!
Do j=0,n_elem_recv-1
If((counter+psb_elem_recv_+j)>Size(halo)) then
info=-2
call psb_errpush(info,name)
goto 9999
end If
idx = halo(counter+psb_elem_recv_+j)
If(idx > Size(desc_ov%loc_to_glob)) then
info=-3
call psb_errpush(info,name)
goto 9999
endif
gidx = desc_ov%loc_to_glob(idx)
call psb_check_size((counter_o+3),tmp_ovr_idx,info,pad=-1)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
tmp_ovr_idx(counter_o)=proc
tmp_ovr_idx(counter_o+1)=1
tmp_ovr_idx(counter_o+2)=gidx
tmp_ovr_idx(counter_o+3)=-1
counter_o=counter_o+3
if (.not.psb_is_large_desc(desc_ov)) then
call psb_check_size((counter_h+3),tmp_halo,info,pad=-1)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
tmp_halo(counter_h)=proc
tmp_halo(counter_h+1)=1
tmp_halo(counter_h+2)=idx
tmp_halo(counter_h+3)=-1
counter_h=counter_h+3
end if
Enddo
if (debug) write(0,*) me,'Checktmp_o_i Loop Mid1',tmp_ovr_idx(1:10)
counter = counter+n_elem_recv
!
! add send elements in halo_index into ovrlap_index
!
Do j=0,n_elem_send-1
idx = halo(counter+psb_elem_send_+j)
gidx = desc_ov%loc_to_glob(idx)
if (idx > psb_cd_get_local_rows(Desc_a)) &
& write(0,*) me,i_ovr,'Out of local rows ',&
& idx,psb_cd_get_local_rows(Desc_a)
call psb_check_size((counter_o+3),tmp_ovr_idx,info,pad=-1)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
tmp_ovr_idx(counter_o)=proc
tmp_ovr_idx(counter_o+1)=1
tmp_ovr_idx(counter_o+2)=gidx
tmp_ovr_idx(counter_o+3)=-1
counter_o=counter_o+3
!
! Prepare to exchange the halo rows with the other proc.
!
If (i_ovr < (novr)) Then
n_elem = psb_sp_get_nnz_row(idx,a)
call psb_check_size((idxs+tot_elem+n_elem),works,info)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
If((n_elem) > size(blk%ia2)) Then
isz = max((3*size(blk%ia2))/2,(n_elem))
if (debug) write(0,*) me,'Realloc blk',isz
call psb_sp_reall(blk,isz,info)
if (info /= 0) then
info=4010
ch_err='psb_sp_reall'
call psb_errpush(info,name,a_err=ch_err)
goto 9999
end if
End If
call psb_sp_getblk(idx,a,blk,info)
if (info /= 0) then
info=4010
ch_err='psb_sp_getblk'
call psb_errpush(info,name,a_err=ch_err)
goto 9999
end if
!!$ write(0,*) me,'Iteration: ',j,i_ovr
Do jj=1,n_elem
works(idxs+tot_elem+jj)=desc_ov%loc_to_glob(blk%ia2(jj))
End Do
tot_elem=tot_elem+n_elem
End If
Enddo
if (i_ovr < novr) then
if (tot_elem > 1) then
call imsr(tot_elem,works(idxs+1))
lx = works(idxs+1)
i = 1
j = 1
do
j = j + 1
if (j > tot_elem) exit
if (works(idxs+j) /= lx) then
i = i + 1
works(idxs+i) = works(idxs+j)
lx = works(idxs+i)
end if
end do
tot_elem = i
endif
if (debug) write(0,*) me,'Checktmp_o_i Loop Mid2',tmp_ovr_idx(1:10)
sdsz(proc+1) = tot_elem
idxs = idxs + tot_elem
end if
counter = counter+n_elem_send+3
if (debug) write(0,*) me,'Checktmp_o_i Loop End',tmp_ovr_idx(1:10)
Enddo
if (debug) write(0,*)me,'End phase 1 CDOVRBLD', m, n_col, tot_recv
if (i_ovr < novr) then
!
! Exchange data requests with everybody else: so far we have
! accumulated RECV requests, we have an all-to-all to build
! matchings SENDs.
!
call mpi_alltoall(sdsz,1,mpi_integer,rvsz,1,mpi_integer,icomm,info)
if (info /= 0) then
info=4010
ch_err='mpi_alltoall'
call psb_errpush(info,name,a_err=ch_err)
goto 9999
end if
idxs = 0
idxr = 0
counter = 1
Do
proc=halo(counter)
if (proc == -1) exit
n_elem_recv = halo(counter+psb_n_elem_recv_)
counter = counter+n_elem_recv
n_elem_send = halo(counter+psb_n_elem_send_)
bsdindx(proc+1) = idxs
idxs = idxs + sdsz(proc+1)
brvindx(proc+1) = idxr
idxr = idxr + rvsz(proc+1)
counter = counter+n_elem_send+3
Enddo
iszr=sum(rvsz)
if (max(iszr,1) > lworkr) then
call psb_realloc(max(iszr,1),workr,info)
if (info /= 0) then
info=4010
ch_err='psb_realloc'
call psb_errpush(info,name,a_err=ch_err)
goto 9999
end if
lworkr=max(iszr,1)
end if
call mpi_alltoallv(works,sdsz,bsdindx,mpi_integer,&
& workr,rvsz,brvindx,mpi_integer,icomm,info)
if (info /= 0) then
info=4010
ch_err='mpi_alltoallv'
call psb_errpush(info,name,a_err=ch_err)
goto 9999
end if
if (debug) write(0,*) 'ISZR :',iszr
if (psb_is_large_desc(desc_a)) then
call psb_check_size(iszr,maskr,info)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
call psi_idx_cnv(iszr,workr,maskr,desc_ov,info)
iszs = count(maskr(1:iszr)<=0)
if (iszs > size(works)) call psb_realloc(iszs,works,info)
j = 0
do i=1,iszr
if (maskr(i) < 0) then
j=j+1
works(j) = workr(i)
end if
end do
!
! fnd_owner on desc_a because we want the procs who
! owned the rows from the beginning!
!
call psi_fnd_owner(iszs,works,temp,desc_a,info)
n_col=psb_cd_get_local_cols(desc_ov)
do i=1,iszs
idx = works(i)
n_col = psb_cd_get_local_cols(desc_ov)
call psi_idx_ins_cnv(idx,lidx,desc_ov,info)
if (psb_cd_get_local_cols(desc_ov) > n_col ) then
!
! This is a new index. Assigning a local index as
! we receive them guarantees that all indices for HALO(I)
! will be less than those for HALO(J) whenever I<J
!
proc_id = temp(i)
call psb_check_size((counter_t+3),t_halo_in,info,pad=-1)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
t_halo_in(counter_t)=proc_id
t_halo_in(counter_t+1)=1
t_halo_in(counter_t+2)=lidx
counter_t=counter_t+3
if (.false.) write(0,*) me,' CDOVRBLD: Added t_halo_in ',&
&proc_id,lidx,idx
endif
end Do
else
Do i=1,iszr
idx=workr(i)
if (idx <1) then
write(0,*) me,'Error in CDOVRBLD level',i_ovr,' : ',idx,i,iszr
else If (desc_ov%glob_to_loc(idx) < -np) Then
!
! This is a new index. Assigning a local index as
! we receive them guarantees that all indices for HALO(I)
! will be less than those for HALO(J) whenever I<J
!
n_col=n_col+1
proc_id=-desc_ov%glob_to_loc(idx)-np-1
call psb_check_size(n_col,desc_ov%loc_to_glob,info,pad=-1)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
desc_ov%glob_to_loc(idx)=n_col
desc_ov%loc_to_glob(n_col)=idx
call psb_check_size((counter_t+3),t_halo_in,info,pad=-1)
if (info /= 0) then
info=4010
call psb_errpush(info,name,a_err='psb_check_size')
goto 9999
end if
t_halo_in(counter_t)=proc_id
t_halo_in(counter_t+1)=1
t_halo_in(counter_t+2)=n_col
counter_t=counter_t+3
if (debug) write(0,*) me,' CDOVRBLD: Added into t_halo_in from recv',&
&proc_id,n_col,idx
else if (desc_ov%glob_to_loc(idx) < 0) Then
if (debug) write(0,*) me,'Wrong input to cdovrbld??',&
&idx,desc_ov%glob_to_loc(idx)
End If
End Do
desc_ov%matrix_data(psb_n_col_)=n_col
end if
end if
!!$ desc_ov%matrix_data(psb_n_row_)=desc_ov%matrix_data(psb_n_col_)
!
! Ok, now we have a temporary halo with all the info for the
! next round. If we need to keep going, convert the halo format
! from temporary to final, so that we can work out the next iteration.
! This uses one of the convert_comm internals, i.e. we are doing
! the equivalent of a partial call to convert_comm
!
If (i_ovr < (novr)) Then
t_halo_in(counter_t)=-1
if (debug) write(0,*) me,'Checktmp_o_i 1',tmp_ovr_idx(1:10)
if (debug) write(0,*) me,'Calling Crea_Halo'
call psi_crea_index(desc_ov,t_halo_in,t_halo_out,.false.,&
& nxch,nsnd,nrcv,info)
if (debug) then
write(0,*) me,'Done Crea_Index'
call psb_barrier(ictxt)
end if
if (debug) write(0,*) me,'Checktmp_o_i 2',tmp_ovr_idx(1:10)
if (debug) write(0,*) me,'Done Crea_Halo'
call psb_transfer(t_halo_out,halo,info)
!
! At this point we have built the halo necessary for I_OVR+1.
!
End If
if (debug) write(0,*) me,'Checktmp_o_i ',tmp_ovr_idx(1:10)
End Do
desc_ov%matrix_data(psb_m_)=psb_cd_get_global_rows(desc_a)
desc_ov%matrix_data(psb_n_)=psb_cd_get_global_cols(desc_a)
tmp_halo(counter_h:)=-1
tmp_ovr_idx(counter_o:)=-1
!
! At this point we have gathered all the indices in the halo at
! N levels of overlap. Just call cnv_dsc. This is
! the same routine as gets called inside CDASB.
!
if (debug) then
write(0,*) 'psb_cdovrbld: converting indexes'
call psb_barrier(ictxt)
end if
!.... convert comunication stuctures....
! Note that we have to keep local_rows until the very end,
! because otherwise the halo build mechanism of cdasb
! will not work.
call psb_transfer(tmp_ovr_idx,desc_ov%ovrlap_index,info)
call psb_transfer(tmp_halo,desc_ov%halo_index,info)
call psb_cdasb(desc_ov,info)
desc_ov%matrix_data(psb_n_row_)=desc_ov%matrix_data(psb_n_col_)
if (debug) then
write(0,*) me,'Done CDASB'
call psb_barrier(ictxt)
end if
if (info == 0) call psb_sp_free(blk,info)
if (info /= 0) then
ch_err='sp_free'
call psb_errpush(4013,name,a_err=ch_err,i_err=(/info,0,0,0,0/))
goto 9999
end if
call psb_erractionrestore(err_act)
return
9999 continue
call psb_erractionrestore(err_act)
if (err_act == act_abort) then
call psb_error(ictxt)
return
end if
Return
End Subroutine psb_dcdovr