Streamline new smoothed implementation, first step

mlprec/impl/aggregator/mld_d_spmm_bld_inner.f90

@@ -35,6 +35,186 @@
 ! File: mld_daggrmat_nosmth_bld.F90
 !
 !
+subroutine mld_d_new_spmm_bld_inner(a_csr,desc_a,nlaggr,parms,ac,&
+     & coo_prol,desc_cprol,coo_restr,info)
+  use psb_base_mod
+  use mld_d_inner_mod
+  use mld_d_base_aggregator_mod, mld_protect_name => mld_d_new_spmm_bld_inner
+  implicit none
+
+  ! Arguments
+  type(psb_ld_csr_sparse_mat), intent(inout) :: a_csr
+  type(psb_desc_type), intent(in)            :: desc_a
+  integer(psb_lpk_), intent(inout)           :: nlaggr(:)
+  type(mld_dml_parms), intent(inout)         :: parms 
+  type(psb_ld_coo_sparse_mat), intent(inout) :: coo_prol, coo_restr
+  type(psb_desc_type), intent(inout)         :: desc_cprol
+  type(psb_ldspmat_type), intent(out)        :: ac
+  integer(psb_ipk_), intent(out)             :: info
+
+  ! Local variables
+  integer(psb_ipk_)  :: err_act
+  integer(psb_ipk_)  :: ictxt,np,me, icomm, ndx, minfo
+  character(len=40)  :: name
+  integer(psb_ipk_)  :: ierr(5)
+  type(psb_ld_coo_sparse_mat) :: ac_coo, tmpcoo
+  type(psb_ld_csr_sparse_mat) :: acsr3, csr_prol, ac_csr, csr_restr
+  integer(psb_ipk_) :: debug_level, debug_unit, naggr
+  integer(psb_lpk_) :: nrow, nglob, ncol, ntaggr, nzl, ip, &
+       &  nzt, naggrm1, naggrp1, i, k
+  integer(psb_lpk_) ::  nrsave, ncsave, nzsave, nza, nrpsave, ncpsave, nzpsave
+  logical, parameter :: do_timings=.true., oldstyle=.false., debug=.true.  
+  integer(psb_ipk_), save :: idx_spspmm=-1
+
+  name='mld_spmm_bld_inner'
+  if(psb_get_errstatus().ne.0) return 
+  info=psb_success_
+  call psb_erractionsave(err_act)
+
+
+  ictxt = desc_a%get_context()
+  icomm = desc_a%get_mpic()
+  call psb_info(ictxt, me, np)
+  debug_unit  = psb_get_debug_unit()
+  debug_level = psb_get_debug_level()
+  nglob = desc_a%get_global_rows()
+  nrow  = desc_a%get_local_rows()
+  ncol  = desc_a%get_local_cols()
+
+  if ((do_timings).and.(idx_spspmm==-1)) &
+       & idx_spspmm = psb_get_timer_idx("SPMM_BLD: par_spspmm")
+
+  naggr   = nlaggr(me+1)
+  ntaggr  = sum(nlaggr)
+  naggrm1 = sum(nlaggr(1:me))
+  naggrp1 = sum(nlaggr(1:me+1)) 
+  !write(0,*)me,' ',name,' input sizes',nlaggr(:),':',naggr
+  nrpsave = coo_prol%get_nrows()
+  ncpsave = coo_prol%get_ncols()
+  nzpsave = coo_prol%get_nzeros()    
+  write(0,*)me,' ',name,' input coo_prol ',nrpsave,ncpsave,nzpsave
+
+  !
+  ! Here COO_PROL should be with GLOBAL indices on the cols
+  ! and LOCAL indices on the rows. 
+  !
+  if (debug) write(0,*)  me,' ',trim(name),' Size check on entry New: ',&
+       & coo_prol%get_fmt(),coo_prol%get_nrows(),coo_prol%get_ncols(),coo_prol%get_nzeros(),&
+       & nrow,ntaggr,naggr
+
+  call coo_prol%cp_to_fmt(csr_prol,info)
+
+  if (debug) write(0,*) me,trim(name),' Product AxPROL ',&
+       & a_csr%get_nrows(),a_csr%get_ncols(), csr_prol%get_nrows(), &
+       & desc_a%get_local_rows(),desc_a%get_local_cols(),&
+       & desc_cprol%get_local_rows(),desc_a%get_local_cols()
+  if (debug) flush(0)
+
+  if (do_timings) call psb_tic(idx_spspmm)
+  call psb_par_spspmm(a_csr,desc_a,csr_prol,acsr3,desc_cprol,info)
+  if (do_timings) call psb_toc(idx_spspmm)  
+
+  if (debug) write(0,*) me,trim(name),' Done AxPROL ',&
+       & acsr3%get_nrows(),acsr3%get_ncols(), acsr3%get_nzeros(),&
+       & desc_cprol%get_local_rows(),desc_cprol%get_local_cols()
+
+  !
+  ! Ok first product done.
+  !
+  ! Remember that RESTR must be built from PROL after halo extension,
+  ! which is done above in psb_par_spspmm
+  if (debug) write(0,*)me,' ',name,' No inp_restr, transposing prol ',&
+       & csr_prol%get_nrows(),csr_prol%get_ncols(),csr_prol%get_nzeros()
+  call csr_prol%cp_to_fmt(coo_restr,info)
+!!$      write(0,*)me,' ',name,' new into transposition ',coo_restr%get_nrows(),&
+!!$           & coo_restr%get_ncols(),coo_restr%get_nzeros()
+  call coo_restr%transp()
+  nzl = coo_restr%get_nzeros()
+  call desc_cprol%l2gip(coo_restr%ia(1:nzl),info)
+  i=0
+  !
+  ! Now we have to fix this.  The only rows of the restrictor that are correct 
+  ! are those corresponding to "local" aggregates, i.e. indices in ilaggr(:)
+  !
+  do k=1, nzl
+    if ((naggrm1 < coo_restr%ia(k)) .and.(coo_restr%ia(k) <= naggrp1)) then
+      i = i+1
+      coo_restr%val(i) = coo_restr%val(k)
+      coo_restr%ia(i)  = coo_restr%ia(k)
+      coo_restr%ja(i)  = coo_restr%ja(k)
+    end if
+  end do
+  call coo_restr%set_nzeros(i)
+  call coo_restr%fix(info)
+  call coo_restr%cp_to_coo(tmpcoo,info)
+!!$      write(0,*)me,' ',name,' after transposition ',coo_restr%get_nrows(),coo_restr%get_ncols(),coo_restr%get_nzeros()
+
+  if (info /= psb_success_) then 
+    call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
+    goto 9999
+  end if
+  if (debug_level >= psb_debug_outer_) &
+       & write(debug_unit,*) me,' ',trim(name),&
+       & 'starting sphalo/ rwxtd'
+  nzl    = tmpcoo%get_nzeros()    
+  call psb_glob_to_loc(tmpcoo%ia(1:nzl),desc_cprol,info,iact='I',owned=.true.)
+  call tmpcoo%clean_negidx(info)
+  nzl  = tmpcoo%get_nzeros()
+  call tmpcoo%set_nrows(desc_cprol%get_local_rows())
+  call tmpcoo%set_ncols(desc_a%get_local_cols())
+!!$    write(0,*)me,' ',name,' after G2L on rows ',tmpcoo%get_nrows(),tmpcoo%get_ncols(),tmpcoo%get_nzeros()      
+  call csr_restr%mv_from_coo(tmpcoo,info)
+
+  if (debug) write(0,*) me,trim(name),' Product RESTRxAP ',&
+       & csr_restr%get_nrows(),csr_restr%get_ncols(), &
+       & desc_cprol%get_local_rows(),desc_a%get_local_cols(),&
+       & acsr3%get_nrows(),acsr3%get_ncols()
+  if (do_timings) call psb_tic(idx_spspmm)      
+  call psb_par_spspmm(csr_restr,desc_a,acsr3,ac_csr,desc_cprol,info)
+  if (do_timings) call psb_toc(idx_spspmm)      
+  call csr_restr%free()
+  call ac_csr%mv_to_coo(ac_coo,info)
+  nza    = ac_coo%get_nzeros()
+  if (debug) write(0,*) me,trim(name),' Fixing ac ',&
+       & ac_coo%get_nrows(),ac_coo%get_ncols(), nza
+  call ac_coo%fix(info)
+  call desc_cprol%indxmap%l2gip(ac_coo%ia(1:nza),info)
+  call desc_cprol%indxmap%l2gip(ac_coo%ja(1:nza),info)
+  call ac_coo%set_nrows(ntaggr)
+  call ac_coo%set_ncols(ntaggr)
+  if (debug) write(0,*)  me,' ',trim(name),' Before mv_from',psb_get_errstatus()
+  if (info == 0) call ac%mv_from(ac_coo)
+  if (debug) write(0,*)  me,' ',trim(name),' After  mv_from',psb_get_errstatus()
+  if (debug) write(0,*)  me,' ',trim(name),' ',ac%get_fmt(),ac%get_nrows(),ac%get_ncols(),ac%get_nzeros(),naggr,ntaggr
+  ! write(0,*)  me,' ',trim(name),' Final AC newstyle ',ac%get_fmt(),ac%get_nrows(),ac%get_ncols(),ac%get_nzeros()
+
+  nza = coo_prol%get_nzeros()
+  call desc_cprol%indxmap%l2gip(coo_prol%ja(1:nza),info)
+  
+  if (debug) then
+    write(0,*) me,' ',trim(name),' Checkpoint at exit'
+    call psb_barrier(ictxt)
+    write(0,*) me,' ',trim(name),' Checkpoint through'
+  end if
+
+  if (info /= psb_success_) then
+    call psb_errpush(psb_err_internal_error_,name,a_err='Build ac = coo_restr x am3')
+    goto 9999
+  end if
+
+
+  if (debug_level >= psb_debug_outer_) &
+       & write(debug_unit,*) me,' ',trim(name),&
+       & 'Done smooth_aggregate '
+
+  call psb_erractionrestore(err_act)
+  return
+
+9999 call psb_error_handler(err_act)
+
+  return
+  
+end subroutine mld_d_new_spmm_bld_inner
 subroutine mld_d_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,ac,&
      & op_prol,op_restr,info)
   use psb_base_mod
@@ -63,7 +243,7 @@ subroutine mld_d_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,ac,&
   integer(psb_lpk_) :: nrow, nglob, ncol, ntaggr, nzl, ip, &
        &  nzt, naggrm1, naggrp1, i, k
   integer(psb_lpk_) ::  nrsave, ncsave, nzsave, nza, nrpsave, ncpsave, nzpsave
-  logical, parameter :: do_timings=.true., oldstyle=.false., debug=.false.  
+  logical, parameter :: do_timings=.true., oldstyle=.false., debug=.true.  
   integer(psb_ipk_), save :: idx_spspmm=-1
 
   name='mld_spmm_bld_inner'
@@ -92,7 +272,7 @@ subroutine mld_d_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,ac,&
   nrpsave = op_prol%get_nrows()
   ncpsave = op_prol%get_ncols()
   nzpsave = op_prol%get_nzeros()    
-  !write(0,*)me,' ',name,' input op_prol ',nrpsave,ncpsave,nzpsave
+  write(0,*)me,' ',name,' input op_prol ',nrpsave,ncpsave,nzpsave
 
   !
   ! Here OP_PROL should be with GLOBAL indices on the cols
@@ -219,3 +399,4 @@ subroutine mld_d_spmm_bld_inner(a_csr,desc_a,ilaggr,nlaggr,parms,ac,&
   return
   
 end subroutine mld_d_spmm_bld_inner
+

mlprec/impl/aggregator/mld_daggrmat_smth_bld.f90

@@ -126,7 +126,7 @@ subroutine mld_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,ac,op_prol,op_rest
   integer(psb_ipk_) :: ictxt, np, me
   character(len=20) :: name
   type(psb_desc_type) :: tmp_desc
-  type(psb_ld_coo_sparse_mat) :: tmpcoo
+  type(psb_ld_coo_sparse_mat) :: coo_prol, coo_restr, tmpcoo
   type(psb_ld_csr_sparse_mat) :: acsr1,  acsrf, csr_prol, acsr
   real(psb_dpk_), allocatable :: adiag(:)
   real(psb_dpk_), allocatable :: arwsum(:)
@@ -176,7 +176,7 @@ subroutine mld_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,ac,op_prol,op_rest
   if (info == psb_success_) &
        & call psb_halo(adiag,desc_a,info)
   if (info == psb_success_) call a%cp_to(acsr)
-  call op_prol%cp_to(tmpcoo)
+  call op_prol%mv_to(coo_prol)
 
   if(info /= psb_success_) then
     call psb_errpush(psb_err_from_subroutine_,name,a_err='sp_getdiag')
@@ -257,10 +257,10 @@ subroutine mld_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,ac,op_prol,op_rest
 
   inaggr = naggr
   call psb_cdall(ictxt,tmp_desc,info,nl=inaggr)
-  nzlp = tmpcoo%get_nzeros()
+  nzlp = coo_prol%get_nzeros()
-  call tmp_desc%indxmap%g2lip_ins(tmpcoo%ja(1:nzlp),info) 
+  call tmp_desc%indxmap%g2lip_ins(coo_prol%ja(1:nzlp),info) 
-  call tmpcoo%set_ncols(tmp_desc%get_local_cols())
+  call coo_prol%set_ncols(tmp_desc%get_local_cols())
-  call tmpcoo%mv_to_fmt(csr_prol,info)  
+  call coo_prol%mv_to_fmt(csr_prol,info)  
   !
   ! Build the smoothed prolongator using either A or Af
   !    acsr1 = (I-w*D*A) Prol      acsr1 = (I-w*D*Af) Prol 
@@ -280,6 +280,9 @@ subroutine mld_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,ac,op_prol,op_rest
        & 'Done SPSPMM 1'
 
   nzl = acsr1%get_nzeros()
+  write(0,*) acsr1%get_nrows(),acsr1%get_ncols(),acsr1%get_nzeros()
+  if (.false.) then
+
     call tmp_desc%l2gip(acsr1%ja(1:nzl),info)
     call acsr1%set_dupl(psb_dupl_add_)
     call acsr1%set_ncols(ntaggr)
@@ -287,7 +290,20 @@ subroutine mld_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,ac,op_prol,op_rest
 
     call mld_spmm_bld_inner(acsr,desc_a,ilaggr,nlaggr,parms,ac,&
          & op_prol,op_restr,info)
+    call op_prol%print(0)
+  else
+
+    !call tmp_desc%l2gip(acsr1%ja(1:nzl),info)
 
+    call acsr1%mv_to_coo(coo_prol,info)
+    
+    call mld_spmm_bld_inner(acsr,desc_a,nlaggr,parms,ac,&
+         & coo_prol,tmp_desc,coo_restr,info)
+    
+    call op_prol%mv_from(coo_prol)
+    call op_prol%print(0)
+    call op_restr%mv_from(coo_restr)
+  end if
 
   if (debug_level >= psb_debug_outer_) &
        & write(debug_unit,*) me,' ',trim(name),&

mlprec/mld_d_base_aggregator_mod.f90

@@ -44,6 +44,7 @@ module mld_d_base_aggregator_mod
   use mld_base_prec_type, only : mld_dml_parms, mld_daggr_data 
   use psb_base_mod, only : psb_dspmat_type, psb_ldspmat_type, psb_d_vect_type, &
        & psb_d_base_vect_type, psb_dlinmap_type, psb_dpk_, psb_ld_csr_sparse_mat, &
+       & psb_ld_coo_sparse_mat, & 
        & psb_ipk_, psb_epk_, psb_lpk_, psb_desc_type, psb_i_base_vect_type, &
        & psb_erractionsave, psb_error_handler, psb_success_, psb_toupper
   !
@@ -124,8 +125,6 @@ module mld_d_base_aggregator_mod
       import :: psb_ld_csr_sparse_mat, psb_ldspmat_type, psb_desc_type, &
            & mld_dml_parms, psb_dpk_, psb_ipk_, psb_lpk_
       implicit none
-      
-      ! Arguments
       type(psb_ld_csr_sparse_mat), intent(inout) :: a_csr
       type(psb_desc_type), intent(in)            :: desc_a
       integer(psb_lpk_), intent(inout)           :: ilaggr(:), nlaggr(:)
@@ -134,6 +133,20 @@ module mld_d_base_aggregator_mod
       type(psb_ldspmat_type), intent(out)        :: ac
       integer(psb_ipk_), intent(out)             :: info
     end subroutine mld_d_spmm_bld_inner
+    subroutine mld_d_new_spmm_bld_inner(a_csr,desc_a,nlaggr,parms,ac,&
+         & coo_prol,desc_cprol,coo_restr,info)
+      import :: psb_ld_csr_sparse_mat, psb_ldspmat_type, psb_desc_type, &
+           & psb_ld_coo_sparse_mat, mld_dml_parms, psb_dpk_, psb_ipk_, psb_lpk_
+      implicit none
+      type(psb_ld_csr_sparse_mat), intent(inout) :: a_csr
+      type(psb_desc_type), intent(in)            :: desc_a
+      integer(psb_lpk_), intent(inout)           :: nlaggr(:)
+      type(mld_dml_parms), intent(inout)         :: parms 
+      type(psb_ld_coo_sparse_mat), intent(inout) :: coo_prol, coo_restr
+      type(psb_desc_type), intent(inout)         :: desc_cprol
+      type(psb_ldspmat_type), intent(out)        :: ac
+      integer(psb_ipk_), intent(out)             :: info
+    end subroutine mld_d_new_spmm_bld_inner
   end interface mld_spmm_bld_inner
   
 contains