fdurastante/psblas3
1
0
Fork 0
mirror of https://github.com/sfilippone/psblas3.git synced 2026-10-24 23:01:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

merged dev-openmp into omp-walther

Browse Source
This commit is contained in:
wlthr
2023-08-26 10:55:09 +02:00
parent 1af76c067c f4f37924e2
commit 979a3da95f
37 changed files with 4938 additions and 793 deletions
Download Patch File Download Diff File Expand all files Collapse all files
base/serial/impl/psb_c_coo_impl.F90
+603 -9
View File
@@ -2864,7 +2864,9 @@ subroutine psb_c_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
if (nz == 0) return
if (a%is_bld()) then
! Structure here is peculiar, because this function can be called
! either within a parallel region, or outside.
! Hence the call to set_nzeros done here.
!$omp critical
nza = a%get_nzeros()
isza = a%get_size()
@@ -2879,9 +2881,6 @@ subroutine psb_c_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
nzaold = nza
nza = nza + nz
call a%set_nzeros(nza)
#if defined(OPENMP)
!write(0,*) 'From thread ',omp_get_thread_num(),nzaold,nz,nza,a%get_nzeros()
#endif
end if
!$omp end critical
if (info /= 0) goto 9999
@@ -2960,8 +2959,7 @@ contains
end if
end do
!$OMP END PARALLEL DO
!nza = nza + nz
nza = nza + nz
#else
do i=1, nz
ir = ia(i)
@@ -3489,6 +3487,602 @@ subroutine psb_c_coo_mv_from(a,b)
end subroutine psb_c_coo_mv_from
!
! COO implementation of tril/triu
!
subroutine psb_c_coo_tril(a,l,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,u)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_c_base_mat_mod, psb_protect_name => psb_c_coo_tril
implicit none
class(psb_c_coo_sparse_mat), intent(in) :: a
class(psb_c_coo_sparse_mat), intent(out) :: l
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_c_coo_sparse_mat), optional, intent(out) :: u
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='tril'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call l%set_nzeros(nzlin)
call u%set_nzeros(nzuin)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
nzin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzin = nzin + 1
l%ia(nzin) = i
l%ja(nzin) = ja(k)
l%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call l%set_nzeros(nzin)
end if
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_c_coo_tril
subroutine psb_c_coo_triu(a,u,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,l)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_c_base_mat_mod, psb_protect_name => psb_c_coo_triu
implicit none
class(psb_c_coo_sparse_mat), intent(in) :: a
class(psb_c_coo_sparse_mat), intent(out) :: u
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_c_coo_sparse_mat), optional, intent(out) :: l
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='triu'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call u%set_nzeros(nzuin)
call l%set_nzeros(nzlin)
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <=1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
else
nzin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
nzin = nzin + 1
u%ia(nzin) = i
u%ja(nzin) = ja(k)
u%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call u%set_nzeros(nzin)
end if
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_c_coo_triu
subroutine psb_c_fix_coo(a,info,idir)
use psb_const_mod
@@ -4563,7 +5157,7 @@ function psb_lc_coo_maxval(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, nnz
res = max(res,abs(a%val(i)))
end do
@@ -4630,7 +5224,7 @@ function psb_lc_coo_csnmi(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, m
res = max(res,abs(vt(i)))
end do
@@ -4680,7 +5274,7 @@ function psb_lc_coo_csnm1(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, n
res = max(res,abs(vt(i)))
end do
base/serial/impl/psb_c_csc_impl.F90
+43 -6
View File
@@ -2329,9 +2329,28 @@ subroutine psb_c_cp_csc_to_fmt(a,b,info)
b%psb_c_base_sparse_mat = a%psb_c_base_sparse_mat
nc = a%get_ncols()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,b%icp,info)
call psb_realloc(nz,b%ia,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
b%icp(i)=a%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ia(j) = a%ia(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -2443,9 +2462,27 @@ subroutine psb_c_cp_csc_from_fmt(a,b,info)
a%psb_c_base_sparse_mat = b%psb_c_base_sparse_mat
nc = b%get_ncols()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,a%icp,info)
call psb_realloc(nz,a%ia,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
a%icp(i)=b%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ia(j)=b%ia(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default
base/serial/impl/psb_c_csr_impl.F90
+345 -8
View File
@@ -2289,7 +2289,155 @@ subroutine psb_c_csr_tril(a,l,info,&
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
@@ -2359,7 +2507,7 @@ subroutine psb_c_csr_tril(a,l,info,&
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -2443,6 +2591,158 @@ subroutine psb_c_csr_triu(a,u,info,&
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
@@ -2511,7 +2811,7 @@ subroutine psb_c_csr_triu(a,u,info,&
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -3190,9 +3490,28 @@ subroutine psb_c_cp_csr_to_fmt(a,b,info)
b%psb_c_base_sparse_mat = a%psb_c_base_sparse_mat
nr = a%get_nrows()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,b%irp,info)
call psb_realloc(nz,b%ja,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
b%irp(i)=a%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ja(j) = a%ja(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -3276,9 +3595,27 @@ subroutine psb_c_cp_csr_from_fmt(a,b,info)
a%psb_c_base_sparse_mat = b%psb_c_base_sparse_mat
nr = b%get_nrows()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,a%irp,info)
call psb_realloc(nz,a%ja,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
a%irp(i)=b%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ja(j)=b%ja(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default
base/serial/impl/psb_d_coo_impl.F90
+603 -9
View File
@@ -2864,7 +2864,9 @@ subroutine psb_d_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
if (nz == 0) return
if (a%is_bld()) then
! Structure here is peculiar, because this function can be called
! either within a parallel region, or outside.
! Hence the call to set_nzeros done here.
!$omp critical
nza = a%get_nzeros()
isza = a%get_size()
@@ -2879,9 +2881,6 @@ subroutine psb_d_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
nzaold = nza
nza = nza + nz
call a%set_nzeros(nza)
#if defined(OPENMP)
!write(0,*) 'From thread ',omp_get_thread_num(),nzaold,nz,nza,a%get_nzeros()
#endif
end if
!$omp end critical
if (info /= 0) goto 9999
@@ -2960,8 +2959,7 @@ contains
end if
end do
!$OMP END PARALLEL DO
!nza = nza + nz
nza = nza + nz
#else
do i=1, nz
ir = ia(i)
@@ -3489,6 +3487,602 @@ subroutine psb_d_coo_mv_from(a,b)
end subroutine psb_d_coo_mv_from
!
! COO implementation of tril/triu
!
subroutine psb_d_coo_tril(a,l,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,u)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_d_base_mat_mod, psb_protect_name => psb_d_coo_tril
implicit none
class(psb_d_coo_sparse_mat), intent(in) :: a
class(psb_d_coo_sparse_mat), intent(out) :: l
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_d_coo_sparse_mat), optional, intent(out) :: u
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='tril'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call l%set_nzeros(nzlin)
call u%set_nzeros(nzuin)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
nzin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzin = nzin + 1
l%ia(nzin) = i
l%ja(nzin) = ja(k)
l%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call l%set_nzeros(nzin)
end if
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_d_coo_tril
subroutine psb_d_coo_triu(a,u,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,l)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_d_base_mat_mod, psb_protect_name => psb_d_coo_triu
implicit none
class(psb_d_coo_sparse_mat), intent(in) :: a
class(psb_d_coo_sparse_mat), intent(out) :: u
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_d_coo_sparse_mat), optional, intent(out) :: l
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='triu'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call u%set_nzeros(nzuin)
call l%set_nzeros(nzlin)
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <=1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
else
nzin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
nzin = nzin + 1
u%ia(nzin) = i
u%ja(nzin) = ja(k)
u%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call u%set_nzeros(nzin)
end if
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_d_coo_triu
subroutine psb_d_fix_coo(a,info,idir)
use psb_const_mod
@@ -4563,7 +5157,7 @@ function psb_ld_coo_maxval(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, nnz
res = max(res,abs(a%val(i)))
end do
@@ -4630,7 +5224,7 @@ function psb_ld_coo_csnmi(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, m
res = max(res,abs(vt(i)))
end do
@@ -4680,7 +5274,7 @@ function psb_ld_coo_csnm1(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, n
res = max(res,abs(vt(i)))
end do
base/serial/impl/psb_d_csc_impl.F90
+43 -6
View File
@@ -2329,9 +2329,28 @@ subroutine psb_d_cp_csc_to_fmt(a,b,info)
b%psb_d_base_sparse_mat = a%psb_d_base_sparse_mat
nc = a%get_ncols()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,b%icp,info)
call psb_realloc(nz,b%ia,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
b%icp(i)=a%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ia(j) = a%ia(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -2443,9 +2462,27 @@ subroutine psb_d_cp_csc_from_fmt(a,b,info)
a%psb_d_base_sparse_mat = b%psb_d_base_sparse_mat
nc = b%get_ncols()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,a%icp,info)
call psb_realloc(nz,a%ia,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
a%icp(i)=b%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ia(j)=b%ia(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default
base/serial/impl/psb_d_csr_impl.F90
+345 -8
View File
@@ -2289,7 +2289,155 @@ subroutine psb_d_csr_tril(a,l,info,&
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
@@ -2359,7 +2507,7 @@ subroutine psb_d_csr_tril(a,l,info,&
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -2443,6 +2591,158 @@ subroutine psb_d_csr_triu(a,u,info,&
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
@@ -2511,7 +2811,7 @@ subroutine psb_d_csr_triu(a,u,info,&
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -3190,9 +3490,28 @@ subroutine psb_d_cp_csr_to_fmt(a,b,info)
b%psb_d_base_sparse_mat = a%psb_d_base_sparse_mat
nr = a%get_nrows()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,b%irp,info)
call psb_realloc(nz,b%ja,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
b%irp(i)=a%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ja(j) = a%ja(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -3276,9 +3595,27 @@ subroutine psb_d_cp_csr_from_fmt(a,b,info)
a%psb_d_base_sparse_mat = b%psb_d_base_sparse_mat
nr = b%get_nrows()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,a%irp,info)
call psb_realloc(nz,a%ja,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
a%irp(i)=b%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ja(j)=b%ja(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default
base/serial/impl/psb_s_coo_impl.F90
+603 -9
View File
@@ -2864,7 +2864,9 @@ subroutine psb_s_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
if (nz == 0) return
if (a%is_bld()) then
! Structure here is peculiar, because this function can be called
! either within a parallel region, or outside.
! Hence the call to set_nzeros done here.
!$omp critical
nza = a%get_nzeros()
isza = a%get_size()
@@ -2879,9 +2881,6 @@ subroutine psb_s_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
nzaold = nza
nza = nza + nz
call a%set_nzeros(nza)
#if defined(OPENMP)
!write(0,*) 'From thread ',omp_get_thread_num(),nzaold,nz,nza,a%get_nzeros()
#endif
end if
!$omp end critical
if (info /= 0) goto 9999
@@ -2960,8 +2959,7 @@ contains
end if
end do
!$OMP END PARALLEL DO
!nza = nza + nz
nza = nza + nz
#else
do i=1, nz
ir = ia(i)
@@ -3489,6 +3487,602 @@ subroutine psb_s_coo_mv_from(a,b)
end subroutine psb_s_coo_mv_from
!
! COO implementation of tril/triu
!
subroutine psb_s_coo_tril(a,l,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,u)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_s_base_mat_mod, psb_protect_name => psb_s_coo_tril
implicit none
class(psb_s_coo_sparse_mat), intent(in) :: a
class(psb_s_coo_sparse_mat), intent(out) :: l
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_s_coo_sparse_mat), optional, intent(out) :: u
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='tril'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call l%set_nzeros(nzlin)
call u%set_nzeros(nzuin)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
nzin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzin = nzin + 1
l%ia(nzin) = i
l%ja(nzin) = ja(k)
l%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call l%set_nzeros(nzin)
end if
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_s_coo_tril
subroutine psb_s_coo_triu(a,u,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,l)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_s_base_mat_mod, psb_protect_name => psb_s_coo_triu
implicit none
class(psb_s_coo_sparse_mat), intent(in) :: a
class(psb_s_coo_sparse_mat), intent(out) :: u
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_s_coo_sparse_mat), optional, intent(out) :: l
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='triu'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call u%set_nzeros(nzuin)
call l%set_nzeros(nzlin)
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <=1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
else
nzin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
nzin = nzin + 1
u%ia(nzin) = i
u%ja(nzin) = ja(k)
u%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call u%set_nzeros(nzin)
end if
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_s_coo_triu
subroutine psb_s_fix_coo(a,info,idir)
use psb_const_mod
@@ -4563,7 +5157,7 @@ function psb_ls_coo_maxval(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, nnz
res = max(res,abs(a%val(i)))
end do
@@ -4630,7 +5224,7 @@ function psb_ls_coo_csnmi(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, m
res = max(res,abs(vt(i)))
end do
@@ -4680,7 +5274,7 @@ function psb_ls_coo_csnm1(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, n
res = max(res,abs(vt(i)))
end do
base/serial/impl/psb_s_csc_impl.F90
+43 -6
View File
@@ -2329,9 +2329,28 @@ subroutine psb_s_cp_csc_to_fmt(a,b,info)
b%psb_s_base_sparse_mat = a%psb_s_base_sparse_mat
nc = a%get_ncols()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,b%icp,info)
call psb_realloc(nz,b%ia,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
b%icp(i)=a%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ia(j) = a%ia(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -2443,9 +2462,27 @@ subroutine psb_s_cp_csc_from_fmt(a,b,info)
a%psb_s_base_sparse_mat = b%psb_s_base_sparse_mat
nc = b%get_ncols()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,a%icp,info)
call psb_realloc(nz,a%ia,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
a%icp(i)=b%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ia(j)=b%ia(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default
base/serial/impl/psb_s_csr_impl.F90
+345 -8
View File
@@ -2289,7 +2289,155 @@ subroutine psb_s_csr_tril(a,l,info,&
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
@@ -2359,7 +2507,7 @@ subroutine psb_s_csr_tril(a,l,info,&
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -2443,6 +2591,158 @@ subroutine psb_s_csr_triu(a,u,info,&
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
@@ -2511,7 +2811,7 @@ subroutine psb_s_csr_triu(a,u,info,&
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -3190,9 +3490,28 @@ subroutine psb_s_cp_csr_to_fmt(a,b,info)
b%psb_s_base_sparse_mat = a%psb_s_base_sparse_mat
nr = a%get_nrows()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,b%irp,info)
call psb_realloc(nz,b%ja,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
b%irp(i)=a%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ja(j) = a%ja(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -3276,9 +3595,27 @@ subroutine psb_s_cp_csr_from_fmt(a,b,info)
a%psb_s_base_sparse_mat = b%psb_s_base_sparse_mat
nr = b%get_nrows()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,a%irp,info)
call psb_realloc(nz,a%ja,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
a%irp(i)=b%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ja(j)=b%ja(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default
base/serial/impl/psb_z_coo_impl.F90
+603 -9
View File
@@ -2864,7 +2864,9 @@ subroutine psb_z_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
if (nz == 0) return
if (a%is_bld()) then
! Structure here is peculiar, because this function can be called
! either within a parallel region, or outside.
! Hence the call to set_nzeros done here.
!$omp critical
nza = a%get_nzeros()
isza = a%get_size()
@@ -2879,9 +2881,6 @@ subroutine psb_z_coo_csput_a(nz,ia,ja,val,a,imin,imax,jmin,jmax,info)
nzaold = nza
nza = nza + nz
call a%set_nzeros(nza)
#if defined(OPENMP)
!write(0,*) 'From thread ',omp_get_thread_num(),nzaold,nz,nza,a%get_nzeros()
#endif
end if
!$omp end critical
if (info /= 0) goto 9999
@@ -2960,8 +2959,7 @@ contains
end if
end do
!$OMP END PARALLEL DO
!nza = nza + nz
nza = nza + nz
#else
do i=1, nz
ir = ia(i)
@@ -3489,6 +3487,602 @@ subroutine psb_z_coo_mv_from(a,b)
end subroutine psb_z_coo_mv_from
!
! COO implementation of tril/triu
!
subroutine psb_z_coo_tril(a,l,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,u)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_z_base_mat_mod, psb_protect_name => psb_z_coo_tril
implicit none
class(psb_z_coo_sparse_mat), intent(in) :: a
class(psb_z_coo_sparse_mat), intent(out) :: l
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_z_coo_sparse_mat), optional, intent(out) :: u
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='tril'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call l%set_nzeros(nzlin)
call u%set_nzeros(nzuin)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
nzin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
nzin = nzin + 1
l%ia(nzin) = i
l%ja(nzin) = ja(k)
l%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call l%set_nzeros(nzin)
end if
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_z_coo_tril
subroutine psb_z_coo_triu(a,u,info,&
& diag,imin,imax,jmin,jmax,rscale,cscale,l)
! Output is always in COO format
use psb_error_mod
use psb_const_mod
use psb_z_base_mat_mod, psb_protect_name => psb_z_coo_triu
implicit none
class(psb_z_coo_sparse_mat), intent(in) :: a
class(psb_z_coo_sparse_mat), intent(out) :: u
integer(psb_ipk_),intent(out) :: info
integer(psb_ipk_), intent(in), optional :: diag,imin,imax,jmin,jmax
logical, intent(in), optional :: rscale,cscale
class(psb_z_coo_sparse_mat), optional, intent(out) :: l
integer(psb_ipk_) :: err_act, nzin, nzout, i, j, k
integer(psb_ipk_) :: imin_, imax_, jmin_, jmax_, mb,nb, diag_, nzlin, nzuin, nz
character(len=20) :: name='triu'
logical :: rscale_, cscale_
logical, parameter :: debug=.false.
call psb_erractionsave(err_act)
info = psb_success_
if (present(diag)) then
diag_ = diag
else
diag_ = 0
end if
if (present(imin)) then
imin_ = imin
else
imin_ = 1
end if
if (present(imax)) then
imax_ = imax
else
imax_ = a%get_nrows()
end if
if (present(jmin)) then
jmin_ = jmin
else
jmin_ = 1
end if
if (present(jmax)) then
jmax_ = jmax
else
jmax_ = a%get_ncols()
end if
if (present(rscale)) then
rscale_ = rscale
else
rscale_ = .true.
end if
if (present(cscale)) then
cscale_ = cscale
else
cscale_ = .true.
end if
if (rscale_) then
mb = imax_ - imin_ +1
else
mb = imax_
endif
if (cscale_) then
nb = jmax_ - jmin_ +1
else
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = a%ia(k)
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do k=1,nz
i = a%ia(k)
j = a%ja(k)
if ((i>=imin_).and.(i<=imax_).and.(jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = a%ia(k)
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop1: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
nzlin = nzlin + 1
l%ia(nzlin) = i
l%ja(nzlin) = ja(k)
l%val(nzlin) = val(k)
else
nzuin = nzuin + 1
u%ia(nzuin) = i
u%ja(nzuin) = ja(k)
u%val(nzuin) = val(k)
end if
end if
end do loop1
end associate
call u%set_nzeros(nzuin)
call l%set_nzeros(nzlin)
call l%fix(info)
nzout = l%get_nzeros()
if (rscale_) &
& l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
if (cscale_) &
& l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
if ((diag_ <=1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
else
nzin = u%get_nzeros() ! At this point it should be 0
associate(val =>a%val, ja => a%ja, ia=>a%ia)
loop2: do k=1,nz
i = ia(k)
j = ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
nzin = nzin + 1
u%ia(nzin) = i
u%ja(nzin) = ja(k)
u%val(nzin) = val(k)
end if
end if
end do loop2
end associate
call u%set_nzeros(nzin)
end if
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) &
& u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
if (cscale_) &
& u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine psb_z_coo_triu
subroutine psb_z_fix_coo(a,info,idir)
use psb_const_mod
@@ -4563,7 +5157,7 @@ function psb_lz_coo_maxval(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, nnz
res = max(res,abs(a%val(i)))
end do
@@ -4630,7 +5224,7 @@ function psb_lz_coo_csnmi(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, m
res = max(res,abs(vt(i)))
end do
@@ -4680,7 +5274,7 @@ function psb_lz_coo_csnm1(a) result(res)
#if defined(OPENMP)
block
integer(psb_ipk_) :: i
!$omp parallel do private(i)
!$omp parallel do private(i) reduction(max:res)
do i=1, n
res = max(res,abs(vt(i)))
end do
base/serial/impl/psb_z_csc_impl.F90
+43 -6
View File
@@ -2329,9 +2329,28 @@ subroutine psb_z_cp_csc_to_fmt(a,b,info)
b%psb_z_base_sparse_mat = a%psb_z_base_sparse_mat
nc = a%get_ncols()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%icp(1:nc+1), b%icp , info)
if (info == 0) call psb_safe_cpy( a%ia(1:nz), b%ia , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,b%icp,info)
call psb_realloc(nz,b%ia,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
b%icp(i)=a%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ia(j) = a%ia(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -2443,9 +2462,27 @@ subroutine psb_z_cp_csc_from_fmt(a,b,info)
a%psb_z_base_sparse_mat = b%psb_z_base_sparse_mat
nc = b%get_ncols()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%icp(1:nc+1), a%icp , info)
if (info == 0) call psb_safe_cpy( b%ia(1:nz), a%ia , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz), a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nc+1,a%icp,info)
call psb_realloc(nz,a%ia,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nc+1
a%icp(i)=b%icp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ia(j)=b%ia(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default
base/serial/impl/psb_z_csr_impl.F90
+345 -8
View File
@@ -2289,7 +2289,155 @@ subroutine psb_z_csr_tril(a,l,info,&
nb = jmax_
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,lrws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(u)) then
nzlin = l%get_nzeros() ! At this point it should be 0
call u%allocate(mb,nb,nz)
nzuin = u%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >=-1).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_lower(.false.)
end if
else
lnz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws) reduction(+: lnz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,lrws,info)
!$omp parallel do private(i,j,k,lpnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<=diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call l%set_nzeros(lnz)
call l%fix(info)
end if
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <= 0).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
end block
#else
nz = a%get_nzeros()
call l%allocate(mb,nb,nz)
@@ -2359,7 +2507,7 @@ subroutine psb_z_csr_tril(a,l,info,&
call l%set_triangle(.true.)
call l%set_lower(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -2443,6 +2591,158 @@ subroutine psb_z_csr_triu(a,u,info,&
endif
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: lrws(:),urws(:)
integer(psb_ipk_) :: lpnt, upnt, lnz, unz
call psb_realloc(mb,urws,info)
!$omp workshare
urws(:) = 0
!$omp end workshare
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
!write(0,*) 'Invocation of COO%TRIL', present(u),nz
if (present(l)) then
nzuin = u%get_nzeros() ! At this point it should be 0
call l%allocate(mb,nb,nz)
nzlin = l%get_nzeros() ! At this point it should be 0
if (info == 0) call psb_realloc(mb,urws,info)
!$omp workshare
lrws(:) = 0
!$omp end workshare
!write(0,*) 'omp version of COO%TRIL/TRIU'
lnz = 0
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,lrws,urws) reduction(+: lnz,unz)
loop1: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic update
lrws(i-imin_+1) = lrws(i-imin_+1) +1
!$omp end atomic
lnz = lnz + 1
else
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop1
!$omp end parallel do
call psi_exscan(mb,lrws,info)
call psi_exscan(mb,urws,info)
!write(0,*) lrws(:), urws(:)
!$omp parallel do private(i,j,k,lpnt,upnt) shared(imin_,imax_,a)
loop2: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)<diag_) then
!$omp atomic capture
lrws(i-imin_+1) = lrws(i-imin_+1) +1
lpnt = lrws(i-imin_+1)
!$omp end atomic
l%ia(lpnt) = i
l%ja(lpnt) = a%ja(k)
l%val(lpnt) = a%val(k)
else
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop2
!$omp end parallel do
!write(0,*) 'End of copyout',lnz,unz
call l%set_nzeros(lnz)
call l%fix(info)
call u%set_nzeros(unz)
call u%fix(info)
nzout = l%get_nzeros()
if (rscale_) then
!$omp workshare
l%ia(1:nzout) = l%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
l%ja(1:nzout) = l%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ <=-1).and.(imin_ == jmin_)) then
call l%set_triangle(.true.)
call l%set_lower(.false.)
end if
else
unz = 0
!$omp parallel do private(i,j,k) shared(imin_,imax_,a,urws) reduction(+: unz)
loop3: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic update
urws(i-imin_+1) = urws(i-imin_+1) +1
!$omp end atomic
unz = unz + 1
end if
end if
end do
end do loop3
!$omp end parallel do
call psi_exscan(mb,urws,info)
!$omp parallel do private(i,j,k,upnt) shared(imin_,imax_,a)
loop4: do i=imin_,imax_
do k = a%irp(i),a%irp(i+1)-1
j = a%ja(k)
if ((jmin_<=j).and.(j<=jmax_)) then
if ((j-i)>=diag_) then
!$omp atomic capture
urws(i-imin_+1) = urws(i-imin_+1) +1
upnt = urws(i-imin_+1)
!$omp end atomic
u%ia(upnt) = i
u%ja(upnt) = a%ja(k)
u%val(upnt) = a%val(k)
end if
end if
end do
end do loop4
!$omp end parallel do
call u%set_nzeros(unz)
call u%fix(info)
end if
nzout = u%get_nzeros()
if (rscale_) then
!$omp workshare
u%ia(1:nzout) = u%ia(1:nzout) - imin_ + 1
!$omp end workshare
end if
if (cscale_) then
!$omp workshare
u%ja(1:nzout) = u%ja(1:nzout) - jmin_ + 1
!$omp end workshare
end if
if ((diag_ >= 0).and.(imin_ == jmin_)) then
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
end block
#else
nz = a%get_nzeros()
call u%allocate(mb,nb,nz)
@@ -2511,7 +2811,7 @@ subroutine psb_z_csr_triu(a,u,info,&
call u%set_triangle(.true.)
call u%set_upper(.true.)
end if
#endif
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
@@ -3190,9 +3490,28 @@ subroutine psb_z_cp_csr_to_fmt(a,b,info)
b%psb_z_base_sparse_mat = a%psb_z_base_sparse_mat
nr = a%get_nrows()
nz = a%get_nzeros()
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( a%irp(1:nr+1), b%irp , info)
if (info == 0) call psb_safe_cpy( a%ja(1:nz), b%ja , info)
if (info == 0) call psb_safe_cpy( a%val(1:nz), b%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,b%irp,info)
call psb_realloc(nz,b%ja,info)
call psb_realloc(nz,b%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
b%irp(i)=a%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
b%ja(j) = a%ja(j)
b%val(j) = a%val(j)
end do
!$omp end parallel do
end if
call b%set_host()
class default
@@ -3276,9 +3595,27 @@ subroutine psb_z_cp_csr_from_fmt(a,b,info)
a%psb_z_base_sparse_mat = b%psb_z_base_sparse_mat
nr = b%get_nrows()
nz = b%get_nzeros()
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
if (.false.) then
if (info == 0) call psb_safe_cpy( b%irp(1:nr+1), a%irp , info)
if (info == 0) call psb_safe_cpy( b%ja(1:nz) , a%ja , info)
if (info == 0) call psb_safe_cpy( b%val(1:nz) , a%val , info)
else
! Despite the implementation in safe_cpy, it seems better this way
call psb_realloc(nr+1,a%irp,info)
call psb_realloc(nz,a%ja,info)
call psb_realloc(nz,a%val,info)
!$omp parallel do private(i) schedule(static)
do i=1,nr+1
a%irp(i)=b%irp(i)
end do
!$omp end parallel do
!$omp parallel do private(j) schedule(static)
do j=1,nz
a%ja(j)=b%ja(j)
a%val(j)=b%val(j)
end do
!$omp end parallel do
end if
call a%set_host()
class default