Compare commits
No commits in common. '7006665d823f9ca3efcbd929e933fe0cc956adc0' and '99a334d93d11745ac241bf5d869ebc3c4c61b95a' have entirely different histories.
7006665d82
...
99a334d93d
@ -1,35 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_allocate_mnnz_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_allocate_mnnz(m, n, a, nz)
|
|
||||||
implicit none
|
|
||||||
integer(psb_ipk_), intent(in) :: m, n
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
integer(psb_ipk_), intent(in), optional :: nz
|
|
||||||
integer(psb_ipk_) :: info
|
|
||||||
integer(psb_ipk_) :: err_act, nz_
|
|
||||||
character(len=20) :: name='allocate_mnz'
|
|
||||||
logical, parameter :: debug=.false.
|
|
||||||
|
|
||||||
call psb_erractionsave(err_act)
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
call a%psb_d_csr_sparse_mat%allocate(m, n, nz)
|
|
||||||
|
|
||||||
if (.not.allocated(a%val)) then
|
|
||||||
allocate(a%val(nz))
|
|
||||||
allocate(a%ja(nz))
|
|
||||||
allocate(a%irp(m+1))
|
|
||||||
end if
|
|
||||||
|
|
||||||
call a%set_dev()
|
|
||||||
if (info /= 0) goto 9999
|
|
||||||
|
|
||||||
call psb_erractionrestore(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_allocate_mnnz
|
|
||||||
end submodule psb_d_oacc_csr_allocate_mnnz_impl
|
|
||||||
@ -1,26 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_cp_from_coo_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_cp_from_coo(a, b, info)
|
|
||||||
implicit none
|
|
||||||
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_coo_sparse_mat), intent(in) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
call a%psb_d_csr_sparse_mat%cp_from_coo(b, info)
|
|
||||||
if (info /= 0) goto 9999
|
|
||||||
|
|
||||||
call a%set_dev()
|
|
||||||
if (info /= 0) goto 9999
|
|
||||||
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 continue
|
|
||||||
info = psb_err_alloc_dealloc_
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_cp_from_coo
|
|
||||||
end submodule psb_d_oacc_csr_cp_from_coo_impl
|
|
||||||
@ -1,24 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_cp_from_fmt_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_cp_from_fmt(a, b, info)
|
|
||||||
implicit none
|
|
||||||
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_base_sparse_mat), intent(in) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
select type(b)
|
|
||||||
type is (psb_d_coo_sparse_mat)
|
|
||||||
call a%cp_from_coo(b, info)
|
|
||||||
class default
|
|
||||||
call a%psb_d_csr_sparse_mat%cp_from_fmt(b, info)
|
|
||||||
if (info /= 0) return
|
|
||||||
|
|
||||||
!$acc update device(a%val, a%ja, a%irp)
|
|
||||||
end select
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_cp_from_fmt
|
|
||||||
end submodule psb_d_oacc_csr_cp_from_fmt_impl
|
|
||||||
@ -1,86 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_csmm_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_csmm(alpha, a, x, beta, y, info, trans)
|
|
||||||
implicit none
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta
|
|
||||||
real(psb_dpk_), intent(in) :: x(:,:)
|
|
||||||
real(psb_dpk_), intent(inout) :: y(:,:)
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: trans
|
|
||||||
|
|
||||||
character :: trans_
|
|
||||||
integer(psb_ipk_) :: i, j, m, n,k, nxy
|
|
||||||
logical :: tra
|
|
||||||
integer(psb_ipk_) :: err_act
|
|
||||||
character(len=20) :: name = 'd_oacc_csmm'
|
|
||||||
logical, parameter :: debug = .false.
|
|
||||||
|
|
||||||
info = psb_success_
|
|
||||||
call psb_erractionsave(err_act)
|
|
||||||
|
|
||||||
if (present(trans)) then
|
|
||||||
trans_ = trans
|
|
||||||
else
|
|
||||||
trans_ = 'N'
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (.not.a%is_asb()) then
|
|
||||||
info = psb_err_invalid_mat_state_
|
|
||||||
call psb_errpush(info, name)
|
|
||||||
goto 9999
|
|
||||||
endif
|
|
||||||
tra = (psb_toupper(trans_) == 'T') .or. (psb_toupper(trans_) == 'C')
|
|
||||||
|
|
||||||
if (tra) then
|
|
||||||
m = a%get_ncols()
|
|
||||||
n = a%get_nrows()
|
|
||||||
else
|
|
||||||
n = a%get_ncols()
|
|
||||||
m = a%get_nrows()
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (size(x,1) < n) then
|
|
||||||
info = 36
|
|
||||||
call psb_errpush(info, name, i_err = (/3 * ione, n, izero, izero, izero/))
|
|
||||||
goto 9999
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (size(y,1) < m) then
|
|
||||||
info = 36
|
|
||||||
call psb_errpush(info, name, i_err = (/5 * ione, m, izero, izero, izero/))
|
|
||||||
goto 9999
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (tra) then
|
|
||||||
call a%psb_d_csr_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
|
||||||
else
|
|
||||||
nxy = min(size(x,2), size(y,2))
|
|
||||||
|
|
||||||
!$acc parallel loop collapse(2) present(a, x, y)
|
|
||||||
do j = 1, nxy
|
|
||||||
do i = 1, m
|
|
||||||
y(i,j) = beta * y(i,j)
|
|
||||||
end do
|
|
||||||
end do
|
|
||||||
|
|
||||||
!$acc parallel loop collapse(2) present(a, x, y)
|
|
||||||
do j = 1, nxy
|
|
||||||
do i = 1, n
|
|
||||||
do k = a%irp(i), a%irp(i+1) - 1
|
|
||||||
y(a%ja(k), j) = y(a%ja(k), j) + alpha * a%val(k) * x(i, j)
|
|
||||||
end do
|
|
||||||
end do
|
|
||||||
end do
|
|
||||||
endif
|
|
||||||
|
|
||||||
call psb_erractionrestore(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_csmm
|
|
||||||
end submodule psb_d_oacc_csr_csmm_impl
|
|
||||||
|
|
||||||
@ -1,81 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_csmv_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_csmv(alpha, a, x, beta, y, info, trans)
|
|
||||||
implicit none
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta
|
|
||||||
real(psb_dpk_), intent(in) :: x(:)
|
|
||||||
real(psb_dpk_), intent(inout) :: y(:)
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: trans
|
|
||||||
|
|
||||||
character :: trans_
|
|
||||||
integer(psb_ipk_) :: i, j, m, n
|
|
||||||
logical :: tra
|
|
||||||
integer(psb_ipk_) :: err_act
|
|
||||||
character(len=20) :: name = 'd_oacc_csmv'
|
|
||||||
logical, parameter :: debug = .false.
|
|
||||||
|
|
||||||
call psb_erractionsave(err_act)
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
if (present(trans)) then
|
|
||||||
trans_ = trans
|
|
||||||
else
|
|
||||||
trans_ = 'N'
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (.not.a%is_asb()) then
|
|
||||||
info = psb_err_invalid_mat_state_
|
|
||||||
call psb_errpush(info, name)
|
|
||||||
goto 9999
|
|
||||||
endif
|
|
||||||
|
|
||||||
tra = (psb_toupper(trans_) == 'T') .or. (psb_toupper(trans_) == 'C')
|
|
||||||
|
|
||||||
if (tra) then
|
|
||||||
m = a%get_ncols()
|
|
||||||
n = a%get_nrows()
|
|
||||||
else
|
|
||||||
n = a%get_ncols()
|
|
||||||
m = a%get_nrows()
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (size(x,1) < n) then
|
|
||||||
info = 36
|
|
||||||
call psb_errpush(info, name, i_err = (/3 * ione, n, izero, izero, izero/))
|
|
||||||
goto 9999
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (size(y,1) < m) then
|
|
||||||
info = 36
|
|
||||||
call psb_errpush(info, name, i_err = (/5 * ione, m, izero, izero, izero/))
|
|
||||||
goto 9999
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (tra) then
|
|
||||||
call a%psb_d_csr_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
|
||||||
else
|
|
||||||
!$acc parallel loop present(a, x, y)
|
|
||||||
do i = 1, m
|
|
||||||
y(i) = beta * y(i)
|
|
||||||
end do
|
|
||||||
|
|
||||||
!$acc parallel loop present(a, x, y)
|
|
||||||
do i = 1, n
|
|
||||||
do j = a%irp(i), a%irp(i+1) - 1
|
|
||||||
y(a%ja(j)) = y(a%ja(j)) + alpha * a%val(j) * x(i)
|
|
||||||
end do
|
|
||||||
end do
|
|
||||||
endif
|
|
||||||
|
|
||||||
call psb_erractionrestore(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_csmv
|
|
||||||
end submodule psb_d_oacc_csr_csmv_impl
|
|
||||||
|
|
||||||
@ -1,83 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_inner_vect_sv_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_inner_vect_sv(alpha, a, x, beta, y, info, trans)
|
|
||||||
implicit none
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta
|
|
||||||
class(psb_d_base_vect_type), intent(inout) :: x, y
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: trans
|
|
||||||
|
|
||||||
real(psb_dpk_), allocatable :: rx(:), ry(:)
|
|
||||||
logical :: tra
|
|
||||||
character :: trans_
|
|
||||||
integer(psb_ipk_) :: err_act
|
|
||||||
character(len=20) :: name = 'd_oacc_csr_inner_vect_sv'
|
|
||||||
logical, parameter :: debug = .false.
|
|
||||||
integer(psb_ipk_) :: i
|
|
||||||
|
|
||||||
call psb_get_erraction(err_act)
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
if (present(trans)) then
|
|
||||||
trans_ = trans
|
|
||||||
else
|
|
||||||
trans_ = 'N'
|
|
||||||
end if
|
|
||||||
|
|
||||||
if (.not.a%is_asb()) then
|
|
||||||
info = psb_err_invalid_mat_state_
|
|
||||||
call psb_errpush(info, name)
|
|
||||||
goto 9999
|
|
||||||
endif
|
|
||||||
|
|
||||||
tra = (psb_toupper(trans_) == 'T') .or. (psb_toupper(trans_) == 'C')
|
|
||||||
|
|
||||||
if (tra .or. (beta /= dzero)) then
|
|
||||||
call x%sync()
|
|
||||||
call y%sync()
|
|
||||||
call a%psb_d_csr_sparse_mat%inner_spsm(alpha, x, beta, y, info, trans)
|
|
||||||
call y%set_host()
|
|
||||||
else
|
|
||||||
select type (xx => x)
|
|
||||||
type is (psb_d_vect_oacc)
|
|
||||||
select type(yy => y)
|
|
||||||
type is (psb_d_vect_oacc)
|
|
||||||
if (xx%is_host()) call xx%sync()
|
|
||||||
if (beta /= dzero) then
|
|
||||||
if (yy%is_host()) call yy%sync()
|
|
||||||
end if
|
|
||||||
!$acc parallel loop present(a, xx, yy)
|
|
||||||
do i = 1, size(a%val)
|
|
||||||
yy%v(i) = alpha * a%val(i) * xx%v(a%ja(i)) + beta * yy%v(i)
|
|
||||||
end do
|
|
||||||
call yy%set_dev()
|
|
||||||
class default
|
|
||||||
rx = xx%get_vect()
|
|
||||||
ry = y%get_vect()
|
|
||||||
call a%psb_d_csr_sparse_mat%inner_spsm(alpha, rx, beta, ry, info)
|
|
||||||
call y%bld(ry)
|
|
||||||
end select
|
|
||||||
class default
|
|
||||||
rx = x%get_vect()
|
|
||||||
ry = y%get_vect()
|
|
||||||
call a%psb_d_csr_sparse_mat%inner_spsm(alpha, rx, beta, ry, info)
|
|
||||||
call y%bld(ry)
|
|
||||||
end select
|
|
||||||
endif
|
|
||||||
|
|
||||||
if (info /= psb_success_) then
|
|
||||||
info = psb_err_from_subroutine_
|
|
||||||
call psb_errpush(info, name, a_err = 'csrg_vect_sv')
|
|
||||||
goto 9999
|
|
||||||
endif
|
|
||||||
|
|
||||||
call psb_erractionrestore(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
return
|
|
||||||
end subroutine psb_d_oacc_csr_inner_vect_sv
|
|
||||||
end submodule psb_d_oacc_csr_inner_vect_sv_impl
|
|
||||||
|
|
||||||
@ -1,35 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_mold_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_mold(a, b, info)
|
|
||||||
implicit none
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
class(psb_d_base_sparse_mat), intent(inout), allocatable :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
integer(psb_ipk_) :: err_act
|
|
||||||
character(len=20) :: name='csr_mold'
|
|
||||||
logical, parameter :: debug=.false.
|
|
||||||
|
|
||||||
call psb_get_erraction(err_act)
|
|
||||||
|
|
||||||
info = 0
|
|
||||||
if (allocated(b)) then
|
|
||||||
call b%free()
|
|
||||||
deallocate(b, stat=info)
|
|
||||||
end if
|
|
||||||
if (info == 0) allocate(psb_d_oacc_csr_sparse_mat :: b, stat=info)
|
|
||||||
|
|
||||||
if (info /= psb_success_) then
|
|
||||||
info = psb_err_alloc_dealloc_
|
|
||||||
call psb_errpush(info, name)
|
|
||||||
goto 9999
|
|
||||||
end if
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_mold
|
|
||||||
end submodule psb_d_oacc_csr_mold_impl
|
|
||||||
|
|
||||||
@ -1,25 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_mv_from_coo_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_mv_from_coo(a, b, info)
|
|
||||||
implicit none
|
|
||||||
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_coo_sparse_mat), intent(inout) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
call a%psb_d_csr_sparse_mat%mv_from_coo(b, info)
|
|
||||||
if (info /= 0) goto 9999
|
|
||||||
|
|
||||||
!$acc update device(a%val, a%ja, a%irp)
|
|
||||||
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 continue
|
|
||||||
info = psb_err_alloc_dealloc_
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_mv_from_coo
|
|
||||||
end submodule psb_d_oacc_csr_mv_from_coo_impl
|
|
||||||
@ -1,24 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_mv_from_fmt_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_mv_from_fmt(a, b, info)
|
|
||||||
implicit none
|
|
||||||
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_base_sparse_mat), intent(inout) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
select type(b)
|
|
||||||
type is (psb_d_coo_sparse_mat)
|
|
||||||
call a%mv_from_coo(b, info)
|
|
||||||
class default
|
|
||||||
call a%psb_d_csr_sparse_mat%mv_from_fmt(b, info)
|
|
||||||
if (info /= 0) return
|
|
||||||
|
|
||||||
!$acc update device(a%val, a%ja, a%irp)
|
|
||||||
end select
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_mv_from_fmt
|
|
||||||
end submodule psb_d_oacc_csr_mv_from_fmt_impl
|
|
||||||
@ -1,28 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_reallocate_nz_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_reallocate_nz(nz, a)
|
|
||||||
implicit none
|
|
||||||
integer(psb_ipk_), intent(in) :: nz
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
integer(psb_ipk_) :: info
|
|
||||||
integer(psb_ipk_) :: err_act
|
|
||||||
character(len=20) :: name='d_oacc_csr_reallocate_nz'
|
|
||||||
logical, parameter :: debug=.false.
|
|
||||||
|
|
||||||
call psb_erractionsave(err_act)
|
|
||||||
info = psb_success_
|
|
||||||
|
|
||||||
call a%psb_d_csr_sparse_mat%reallocate(nz)
|
|
||||||
|
|
||||||
call a%set_dev()
|
|
||||||
if (info /= 0) goto 9999
|
|
||||||
|
|
||||||
call psb_erractionrestore(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_reallocate_nz
|
|
||||||
end submodule psb_d_oacc_csr_reallocate_nz_impl
|
|
||||||
@ -1,53 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_scal_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_scal(d, a, info, side)
|
|
||||||
implicit none
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: d(:)
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, intent(in), optional :: side
|
|
||||||
|
|
||||||
integer(psb_ipk_) :: err_act
|
|
||||||
character(len=20) :: name='scal'
|
|
||||||
logical, parameter :: debug=.false.
|
|
||||||
integer(psb_ipk_) :: i, j
|
|
||||||
|
|
||||||
info = psb_success_
|
|
||||||
call psb_erractionsave(err_act)
|
|
||||||
|
|
||||||
if (a%is_host()) call a%sync()
|
|
||||||
|
|
||||||
if (present(side)) then
|
|
||||||
if (side == 'L') then
|
|
||||||
!$acc parallel loop present(a, d)
|
|
||||||
do i = 1, a%get_nrows()
|
|
||||||
do j = a%irp(i), a%irp(i+1) - 1
|
|
||||||
a%val(j) = a%val(j) * d(i)
|
|
||||||
end do
|
|
||||||
end do
|
|
||||||
else if (side == 'R') then
|
|
||||||
!$acc parallel loop present(a, d)
|
|
||||||
do i = 1, a%get_ncols()
|
|
||||||
do j = a%irp(i), a%irp(i+1) - 1
|
|
||||||
a%val(j) = a%val(j) * d(a%ja(j))
|
|
||||||
end do
|
|
||||||
end do
|
|
||||||
end if
|
|
||||||
else
|
|
||||||
!$acc parallel loop present(a, d)
|
|
||||||
do i = 1, size(a%val)
|
|
||||||
a%val(i) = a%val(i) * d(i)
|
|
||||||
end do
|
|
||||||
end if
|
|
||||||
|
|
||||||
call a%set_dev()
|
|
||||||
|
|
||||||
call psb_erractionrestore(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_scal
|
|
||||||
end submodule psb_d_oacc_csr_scal_impl
|
|
||||||
@ -1,34 +0,0 @@
|
|||||||
submodule (psb_d_oacc_csr_mat_mod) psb_d_oacc_csr_scals_impl
|
|
||||||
use psb_base_mod
|
|
||||||
contains
|
|
||||||
module subroutine psb_d_oacc_csr_scals(d, a, info)
|
|
||||||
implicit none
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: d
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
|
|
||||||
integer(psb_ipk_) :: err_act
|
|
||||||
character(len=20) :: name='scal'
|
|
||||||
logical, parameter :: debug=.false.
|
|
||||||
integer(psb_ipk_) :: i
|
|
||||||
|
|
||||||
info = psb_success_
|
|
||||||
call psb_erractionsave(err_act)
|
|
||||||
|
|
||||||
if (a%is_host()) call a%sync()
|
|
||||||
|
|
||||||
!$acc parallel loop present(a)
|
|
||||||
do i = 1, size(a%val)
|
|
||||||
a%val(i) = a%val(i) * d
|
|
||||||
end do
|
|
||||||
|
|
||||||
call a%set_dev()
|
|
||||||
|
|
||||||
call psb_erractionrestore(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
9999 call psb_error_handler(err_act)
|
|
||||||
return
|
|
||||||
|
|
||||||
end subroutine psb_d_oacc_csr_scals
|
|
||||||
end submodule psb_d_oacc_csr_scals_impl
|
|
||||||
@ -1,34 +1,31 @@
|
|||||||
submodule (psb_d_oacc_vect_mod) psb_d_oacc_mlt_v_impl
|
|
||||||
use psb_string_mod
|
|
||||||
contains
|
|
||||||
|
|
||||||
module subroutine psb_d_oacc_mlt_v(x, y, info)
|
subroutine d_oacc_mlt_v(x, y, info)
|
||||||
|
use psb_d_oacc_vect_mod, psb_protect_name => d_oacc_mlt_v
|
||||||
|
|
||||||
implicit none
|
implicit none
|
||||||
class(psb_d_base_vect_type), intent(inout) :: x
|
class(psb_d_base_vect_type), intent(inout) :: x
|
||||||
class(psb_d_vect_oacc), intent(inout) :: y
|
class(psb_d_vect_oacc), intent(inout) :: y
|
||||||
integer(psb_ipk_), intent(out) :: info
|
integer(psb_ipk_), intent(out) :: info
|
||||||
|
|
||||||
integer(psb_ipk_) :: i, n
|
integer(psb_ipk_) :: i, n
|
||||||
|
|
||||||
info = 0
|
info = 0
|
||||||
n = min(x%get_nrows(), y%get_nrows())
|
n = min(x%get_nrows(), y%get_nrows())
|
||||||
select type(xx => x)
|
select type(xx => x)
|
||||||
class is (psb_d_vect_oacc)
|
class is (psb_d_vect_oacc)
|
||||||
if (y%is_host()) call y%sync()
|
if (y%is_host()) call y%sync()
|
||||||
if (xx%is_host()) call xx%sync()
|
if (xx%is_host()) call xx%sync()
|
||||||
!$acc parallel loop
|
!$acc parallel loop
|
||||||
do i = 1, n
|
do i = 1, n
|
||||||
y%v(i) = y%v(i) * xx%v(i)
|
y%v(i) = y%v(i) * xx%v(i)
|
||||||
end do
|
end do
|
||||||
call y%set_dev()
|
call y%set_dev()
|
||||||
class default
|
class default
|
||||||
if (xx%is_dev()) call xx%sync()
|
if (xx%is_dev()) call xx%sync()
|
||||||
if (y%is_dev()) call y%sync()
|
if (y%is_dev()) call y%sync()
|
||||||
do i = 1, n
|
do i = 1, n
|
||||||
y%v(i) = y%v(i) * xx%v(i)
|
y%v(i) = y%v(i) * xx%v(i)
|
||||||
end do
|
end do
|
||||||
call y%set_host()
|
call y%set_host()
|
||||||
end select
|
end select
|
||||||
end subroutine psb_d_oacc_mlt_v
|
end subroutine d_oacc_mlt_v
|
||||||
end submodule psb_d_oacc_mlt_v_impl
|
|
||||||
|
|||||||
@ -1,55 +1,53 @@
|
|||||||
submodule (psb_d_oacc_vect_mod) d_oacc_mlt_v_2_impl
|
subroutine d_oacc_mlt_v_2(alpha, x, y, beta, z, info, conjgx, conjgy)
|
||||||
|
use psb_d_oacc_vect_mod, psb_protect_name => d_oacc_mlt_v_2
|
||||||
use psb_string_mod
|
use psb_string_mod
|
||||||
contains
|
implicit none
|
||||||
module subroutine d_oacc_mlt_v_2(alpha, x, y, beta, z, info, conjgx, conjgy)
|
real(psb_dpk_), intent(in) :: alpha, beta
|
||||||
implicit none
|
class(psb_d_base_vect_type), intent(inout) :: x
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta
|
class(psb_d_base_vect_type), intent(inout) :: y
|
||||||
class(psb_d_base_vect_type), intent(inout) :: x
|
class(psb_d_vect_oacc), intent(inout) :: z
|
||||||
class(psb_d_base_vect_type), intent(inout) :: y
|
integer(psb_ipk_), intent(out) :: info
|
||||||
class(psb_d_vect_oacc), intent(inout) :: z
|
character(len=1), intent(in), optional :: conjgx, conjgy
|
||||||
integer(psb_ipk_), intent(out) :: info
|
integer(psb_ipk_) :: i, n
|
||||||
character(len=1), intent(in), optional :: conjgx, conjgy
|
logical :: conjgx_, conjgy_
|
||||||
integer(psb_ipk_) :: i, n
|
|
||||||
logical :: conjgx_, conjgy_
|
|
||||||
|
|
||||||
conjgx_ = .false.
|
conjgx_ = .false.
|
||||||
conjgy_ = .false.
|
conjgy_ = .false.
|
||||||
if (present(conjgx)) conjgx_ = (psb_toupper(conjgx) == 'C')
|
if (present(conjgx)) conjgx_ = (psb_toupper(conjgx) == 'C')
|
||||||
if (present(conjgy)) conjgy_ = (psb_toupper(conjgy) == 'C')
|
if (present(conjgy)) conjgy_ = (psb_toupper(conjgy) == 'C')
|
||||||
|
|
||||||
n = min(x%get_nrows(), y%get_nrows(), z%get_nrows())
|
n = min(x%get_nrows(), y%get_nrows(), z%get_nrows())
|
||||||
|
|
||||||
info = 0
|
info = 0
|
||||||
select type(xx => x)
|
select type(xx => x)
|
||||||
|
class is (psb_d_vect_oacc)
|
||||||
|
select type (yy => y)
|
||||||
class is (psb_d_vect_oacc)
|
class is (psb_d_vect_oacc)
|
||||||
select type (yy => y)
|
if (xx%is_host()) call xx%sync()
|
||||||
class is (psb_d_vect_oacc)
|
if (yy%is_host()) call yy%sync()
|
||||||
if (xx%is_host()) call xx%sync()
|
if ((beta /= dzero) .and. (z%is_host())) call z%sync()
|
||||||
if (yy%is_host()) call yy%sync()
|
!$acc parallel loop
|
||||||
if ((beta /= dzero) .and. (z%is_host())) call z%sync()
|
do i = 1, n
|
||||||
!$acc parallel loop
|
z%v(i) = alpha * xx%v(i) * yy%v(i) + beta * z%v(i)
|
||||||
do i = 1, n
|
end do
|
||||||
z%v(i) = alpha * xx%v(i) * yy%v(i) + beta * z%v(i)
|
call z%set_dev()
|
||||||
end do
|
|
||||||
call z%set_dev()
|
|
||||||
class default
|
|
||||||
if (xx%is_dev()) call xx%sync()
|
|
||||||
if (yy%is_dev()) call yy%sync()
|
|
||||||
if ((beta /= dzero) .and. (z%is_dev())) call z%sync()
|
|
||||||
do i = 1, n
|
|
||||||
z%v(i) = alpha * xx%v(i) * yy%v(i) + beta * z%v(i)
|
|
||||||
end do
|
|
||||||
call z%set_host()
|
|
||||||
end select
|
|
||||||
class default
|
class default
|
||||||
if (x%is_dev()) call x%sync()
|
if (xx%is_dev()) call xx%sync()
|
||||||
if (y%is_dev()) call y%sync()
|
if (yy%is_dev()) call yy%sync()
|
||||||
if ((beta /= dzero) .and. (z%is_dev())) call z%sync()
|
if ((beta /= dzero) .and. (z%is_dev())) call z%sync()
|
||||||
do i = 1, n
|
do i = 1, n
|
||||||
z%v(i) = alpha * x%v(i) * y%v(i) + beta * z%v(i)
|
z%v(i) = alpha * xx%v(i) * yy%v(i) + beta * z%v(i)
|
||||||
end do
|
end do
|
||||||
call z%set_host()
|
call z%set_host()
|
||||||
end select
|
end select
|
||||||
end subroutine d_oacc_mlt_v_2
|
class default
|
||||||
end submodule d_oacc_mlt_v_2_impl
|
if (x%is_dev()) call x%sync()
|
||||||
|
if (y%is_dev()) call y%sync()
|
||||||
|
if ((beta /= dzero) .and. (z%is_dev())) call z%sync()
|
||||||
|
do i = 1, n
|
||||||
|
z%v(i) = alpha * x%v(i) * y%v(i) + beta * z%v(i)
|
||||||
|
end do
|
||||||
|
call z%set_host()
|
||||||
|
end select
|
||||||
|
end subroutine d_oacc_mlt_v_2
|
||||||
|
|
||||||
|
|||||||
@ -1,346 +1,185 @@
|
|||||||
module psb_d_oacc_csr_mat_mod
|
module psb_d_oacc_csr_mat_mod
|
||||||
|
|
||||||
use iso_c_binding
|
use iso_c_binding
|
||||||
use psb_d_mat_mod
|
use psb_d_mat_mod
|
||||||
use psb_d_oacc_vect_mod
|
use psb_d_oacc_vect_mod
|
||||||
!use oaccsparse_mod
|
!use oaccsparse_mod
|
||||||
|
|
||||||
integer(psb_ipk_), parameter, private :: is_host = -1
|
integer(psb_ipk_), parameter, private :: is_host = -1
|
||||||
integer(psb_ipk_), parameter, private :: is_sync = 0
|
integer(psb_ipk_), parameter, private :: is_sync = 0
|
||||||
integer(psb_ipk_), parameter, private :: is_dev = 1
|
integer(psb_ipk_), parameter, private :: is_dev = 1
|
||||||
|
|
||||||
type, extends(psb_d_csr_sparse_mat) :: psb_d_oacc_csr_sparse_mat
|
type, extends(psb_d_csr_sparse_mat) :: psb_d_oacc_csr_sparse_mat
|
||||||
integer(psb_ipk_) :: devstate = is_host
|
integer(psb_ipk_) :: devstate = is_host
|
||||||
contains
|
contains
|
||||||
procedure, nopass :: get_fmt => d_oacc_csr_get_fmt
|
procedure, pass(a) :: all => d_oacc_csr_all
|
||||||
procedure, pass(a) :: sizeof => d_oacc_csr_sizeof
|
procedure, pass(a) :: is_host => d_oacc_csr_is_host
|
||||||
procedure, pass(a) :: vect_mv => psb_d_oacc_csr_vect_mv
|
procedure, pass(a) :: is_sync => d_oacc_csr_is_sync
|
||||||
procedure, pass(a) :: in_vect_sv => psb_d_oacc_csr_inner_vect_sv
|
procedure, pass(a) :: is_dev => d_oacc_csr_is_dev
|
||||||
procedure, pass(a) :: csmm => psb_d_oacc_csr_csmm
|
procedure, pass(a) :: set_host => d_oacc_csr_set_host
|
||||||
procedure, pass(a) :: csmv => psb_d_oacc_csr_csmv
|
procedure, pass(a) :: set_sync => d_oacc_csr_set_sync
|
||||||
procedure, pass(a) :: scals => psb_d_oacc_csr_scals
|
procedure, pass(a) :: set_dev => d_oacc_csr_set_dev
|
||||||
procedure, pass(a) :: scalv => psb_d_oacc_csr_scal
|
procedure, pass(a) :: sync_space => d_oacc_csr_sync_space
|
||||||
procedure, pass(a) :: reallocate_nz => psb_d_oacc_csr_reallocate_nz
|
procedure, pass(a) :: sync => d_oacc_csr_sync
|
||||||
procedure, pass(a) :: allocate_mnnz => psb_d_oacc_csr_allocate_mnnz
|
procedure, pass(a) :: vect_mv => psb_d_oacc_csr_vect_mv
|
||||||
procedure, pass(a) :: cp_from_coo => psb_d_oacc_csr_cp_from_coo
|
end type psb_d_oacc_csr_sparse_mat
|
||||||
procedure, pass(a) :: cp_from_fmt => psb_d_oacc_csr_cp_from_fmt
|
|
||||||
procedure, pass(a) :: mv_from_coo => psb_d_oacc_csr_mv_from_coo
|
interface
|
||||||
procedure, pass(a) :: mv_from_fmt => psb_d_oacc_csr_mv_from_fmt
|
subroutine psb_d_oacc_csr_vect_mv(alpha, a, x, beta, y, info, trans)
|
||||||
procedure, pass(a) :: free => d_oacc_csr_free
|
import :: psb_d_oacc_csr_sparse_mat, psb_dpk_, psb_d_base_vect_type, psb_ipk_
|
||||||
procedure, pass(a) :: mold => psb_d_oacc_csr_mold
|
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
||||||
procedure, pass(a) :: all => d_oacc_csr_all
|
real(psb_dpk_), intent(in) :: alpha, beta
|
||||||
procedure, pass(a) :: is_host => d_oacc_csr_is_host
|
class(psb_d_base_vect_type), intent(inout) :: x, y
|
||||||
procedure, pass(a) :: is_sync => d_oacc_csr_is_sync
|
integer(psb_ipk_), intent(out) :: info
|
||||||
procedure, pass(a) :: is_dev => d_oacc_csr_is_dev
|
character, optional, intent(in) :: trans
|
||||||
procedure, pass(a) :: set_host => d_oacc_csr_set_host
|
end subroutine psb_d_oacc_csr_vect_mv
|
||||||
procedure, pass(a) :: set_sync => d_oacc_csr_set_sync
|
end interface
|
||||||
procedure, pass(a) :: set_dev => d_oacc_csr_set_dev
|
|
||||||
procedure, pass(a) :: sync_space => d_oacc_csr_sync_space
|
|
||||||
procedure, pass(a) :: sync => d_oacc_csr_sync
|
|
||||||
end type psb_d_oacc_csr_sparse_mat
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_mold(a,b,info)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
class(psb_d_base_sparse_mat), intent(inout), allocatable :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
end subroutine psb_d_oacc_csr_mold
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_cp_from_fmt(a,b,info)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_base_sparse_mat), intent(in) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
end subroutine psb_d_oacc_csr_cp_from_fmt
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_mv_from_coo(a,b,info)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_coo_sparse_mat), intent(inout) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
end subroutine psb_d_oacc_csr_mv_from_coo
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_mv_from_fmt(a,b,info)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_base_sparse_mat), intent(inout) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
end subroutine psb_d_oacc_csr_mv_from_fmt
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_vect_mv(alpha, a, x, beta, y, info, trans)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta
|
|
||||||
class(psb_d_base_vect_type), intent(inout) :: x, y
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: trans
|
|
||||||
end subroutine psb_d_oacc_csr_vect_mv
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_inner_vect_sv(alpha, a, x, beta, y, info, trans)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta
|
|
||||||
class(psb_d_base_vect_type), intent(inout) :: x,y
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: trans
|
|
||||||
end subroutine psb_d_oacc_csr_inner_vect_sv
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_csmm(alpha, a, x, beta, y, info, trans)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta, x(:,:)
|
|
||||||
real(psb_dpk_), intent(inout) :: y(:,:)
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: trans
|
|
||||||
end subroutine psb_d_oacc_csr_csmm
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_csmv(alpha, a, x, beta, y, info, trans)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: alpha, beta, x(:)
|
|
||||||
real(psb_dpk_), intent(inout) :: y(:)
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: trans
|
|
||||||
end subroutine psb_d_oacc_csr_csmv
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_scals(d, a, info)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: d
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
end subroutine psb_d_oacc_csr_scals
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_scal(d,a,info,side)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
real(psb_dpk_), intent(in) :: d(:)
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
character, optional, intent(in) :: side
|
|
||||||
end subroutine psb_d_oacc_csr_scal
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_reallocate_nz(nz,a)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
integer(psb_ipk_), intent(in) :: nz
|
|
||||||
end subroutine psb_d_oacc_csr_reallocate_nz
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_allocate_mnnz(m,n,a,nz)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
integer(psb_ipk_), intent(in) :: m,n
|
|
||||||
integer(psb_ipk_), intent(in), optional :: nz
|
|
||||||
end subroutine psb_d_oacc_csr_allocate_mnnz
|
|
||||||
end interface
|
|
||||||
|
|
||||||
interface
|
|
||||||
module subroutine psb_d_oacc_csr_cp_from_coo(a,b,info)
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
|
||||||
class(psb_d_coo_sparse_mat), intent(in) :: b
|
|
||||||
integer(psb_ipk_), intent(out) :: info
|
|
||||||
end subroutine psb_d_oacc_csr_cp_from_coo
|
|
||||||
end interface
|
|
||||||
|
|
||||||
contains
|
contains
|
||||||
|
|
||||||
|
subroutine d_oacc_csr_all(m, n, nz, a, info)
|
||||||
subroutine d_oacc_csr_free(a)
|
implicit none
|
||||||
use psb_base_mod
|
integer(psb_ipk_), intent(in) :: m, n, nz
|
||||||
implicit none
|
class(psb_d_oacc_csr_sparse_mat), intent(out) :: a
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
integer(psb_ipk_), intent(out) :: info
|
||||||
integer(psb_ipk_) :: info
|
|
||||||
|
info = 0
|
||||||
if (allocated(a%val)) then
|
if (allocated(a%val)) then
|
||||||
!$acc exit data delete(a%val)
|
!$acc exit data delete(a%val) finalize
|
||||||
end if
|
deallocate(a%val, stat=info)
|
||||||
if (allocated(a%ja)) then
|
end if
|
||||||
!$acc exit data delete(a%ja)
|
if (allocated(a%ja)) then
|
||||||
end if
|
!$acc exit data delete(a%ja) finalize
|
||||||
if (allocated(a%irp)) then
|
deallocate(a%ja, stat=info)
|
||||||
!$acc exit data delete(a%irp)
|
end if
|
||||||
end if
|
if (allocated(a%irp)) then
|
||||||
|
!$acc exit data delete(a%irp) finalize
|
||||||
call a%psb_d_csr_sparse_mat%free()
|
deallocate(a%irp, stat=info)
|
||||||
|
end if
|
||||||
return
|
|
||||||
end subroutine d_oacc_csr_free
|
call a%set_nrows(m)
|
||||||
|
call a%set_ncols(n)
|
||||||
|
|
||||||
|
allocate(a%val(nz),stat=info)
|
||||||
function d_oacc_csr_sizeof(a) result(res)
|
allocate(a%ja(nz),stat=info)
|
||||||
implicit none
|
allocate(a%irp(m+1),stat=info)
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
if (info == 0) call a%set_host()
|
||||||
integer(psb_epk_) :: res
|
if (info == 0) call a%sync_space()
|
||||||
|
end subroutine d_oacc_csr_all
|
||||||
if (a%is_dev()) call a%sync()
|
|
||||||
|
function d_oacc_csr_is_host(a) result(res)
|
||||||
res = 8
|
implicit none
|
||||||
res = res + psb_sizeof_dp * size(a%val)
|
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
||||||
res = res + psb_sizeof_ip * size(a%irp)
|
logical :: res
|
||||||
res = res + psb_sizeof_ip * size(a%ja)
|
|
||||||
|
res = (a%devstate == is_host)
|
||||||
end function d_oacc_csr_sizeof
|
end function d_oacc_csr_is_host
|
||||||
|
|
||||||
|
function d_oacc_csr_is_sync(a) result(res)
|
||||||
function d_oacc_csr_get_fmt() result(res)
|
implicit none
|
||||||
implicit none
|
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
||||||
character(len=5) :: res
|
logical :: res
|
||||||
res = 'CSR_oacc'
|
|
||||||
end function d_oacc_csr_get_fmt
|
res = (a%devstate == is_sync)
|
||||||
|
end function d_oacc_csr_is_sync
|
||||||
subroutine d_oacc_csr_all(m, n, nz, a, info)
|
|
||||||
implicit none
|
function d_oacc_csr_is_dev(a) result(res)
|
||||||
integer(psb_ipk_), intent(in) :: m, n, nz
|
implicit none
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(out) :: a
|
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
||||||
integer(psb_ipk_), intent(out) :: info
|
logical :: res
|
||||||
|
|
||||||
info = 0
|
res = (a%devstate == is_dev)
|
||||||
if (allocated(a%val)) then
|
end function d_oacc_csr_is_dev
|
||||||
!$acc exit data delete(a%val) finalize
|
|
||||||
deallocate(a%val, stat=info)
|
subroutine d_oacc_csr_set_host(a)
|
||||||
end if
|
implicit none
|
||||||
if (allocated(a%ja)) then
|
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
||||||
!$acc exit data delete(a%ja) finalize
|
|
||||||
deallocate(a%ja, stat=info)
|
a%devstate = is_host
|
||||||
end if
|
end subroutine d_oacc_csr_set_host
|
||||||
if (allocated(a%irp)) then
|
|
||||||
!$acc exit data delete(a%irp) finalize
|
subroutine d_oacc_csr_set_sync(a)
|
||||||
deallocate(a%irp, stat=info)
|
implicit none
|
||||||
end if
|
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
||||||
|
|
||||||
call a%set_nrows(m)
|
a%devstate = is_sync
|
||||||
call a%set_ncols(n)
|
end subroutine d_oacc_csr_set_sync
|
||||||
|
|
||||||
allocate(a%val(nz),stat=info)
|
subroutine d_oacc_csr_set_dev(a)
|
||||||
allocate(a%ja(nz),stat=info)
|
implicit none
|
||||||
allocate(a%irp(m+1),stat=info)
|
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
||||||
if (info == 0) call a%set_host()
|
|
||||||
if (info == 0) call a%sync_space()
|
a%devstate = is_dev
|
||||||
end subroutine d_oacc_csr_all
|
end subroutine d_oacc_csr_set_dev
|
||||||
|
|
||||||
function d_oacc_csr_is_host(a) result(res)
|
subroutine d_oacc_csr_sync_space(a)
|
||||||
implicit none
|
implicit none
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
||||||
logical :: res
|
if (allocated(a%val)) then
|
||||||
|
call d_oacc_create_dev(a%val)
|
||||||
res = (a%devstate == is_host)
|
end if
|
||||||
end function d_oacc_csr_is_host
|
if (allocated(a%ja)) then
|
||||||
|
call i_oacc_create_dev(a%ja)
|
||||||
function d_oacc_csr_is_sync(a) result(res)
|
end if
|
||||||
implicit none
|
if (allocated(a%irp)) then
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
call i_oacc_create_dev(a%irp)
|
||||||
logical :: res
|
end if
|
||||||
|
contains
|
||||||
res = (a%devstate == is_sync)
|
subroutine d_oacc_create_dev(v)
|
||||||
end function d_oacc_csr_is_sync
|
implicit none
|
||||||
|
real(psb_dpk_), intent(in) :: v(:)
|
||||||
function d_oacc_csr_is_dev(a) result(res)
|
!$acc enter data copyin(v)
|
||||||
implicit none
|
end subroutine d_oacc_create_dev
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(in) :: a
|
subroutine i_oacc_create_dev(v)
|
||||||
logical :: res
|
implicit none
|
||||||
|
integer(psb_ipk_), intent(in) :: v(:)
|
||||||
res = (a%devstate == is_dev)
|
!$acc enter data copyin(v)
|
||||||
end function d_oacc_csr_is_dev
|
end subroutine i_oacc_create_dev
|
||||||
|
end subroutine d_oacc_csr_sync_space
|
||||||
subroutine d_oacc_csr_set_host(a)
|
|
||||||
implicit none
|
subroutine d_oacc_csr_sync(a)
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
implicit none
|
||||||
|
class(psb_d_oacc_csr_sparse_mat), target, intent(in) :: a
|
||||||
a%devstate = is_host
|
class(psb_d_oacc_csr_sparse_mat), pointer :: tmpa
|
||||||
end subroutine d_oacc_csr_set_host
|
integer(psb_ipk_) :: info
|
||||||
|
|
||||||
subroutine d_oacc_csr_set_sync(a)
|
tmpa => a
|
||||||
implicit none
|
if (a%is_dev()) then
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
call d_oacc_csr_to_host(a%val)
|
||||||
|
call i_oacc_csr_to_host(a%ja)
|
||||||
a%devstate = is_sync
|
call i_oacc_csr_to_host(a%irp)
|
||||||
end subroutine d_oacc_csr_set_sync
|
else if (a%is_host()) then
|
||||||
|
call d_oacc_csr_to_dev(a%val)
|
||||||
subroutine d_oacc_csr_set_dev(a)
|
call i_oacc_csr_to_dev(a%ja)
|
||||||
implicit none
|
call i_oacc_csr_to_dev(a%irp)
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
end if
|
||||||
|
call tmpa%set_sync()
|
||||||
a%devstate = is_dev
|
end subroutine d_oacc_csr_sync
|
||||||
end subroutine d_oacc_csr_set_dev
|
|
||||||
|
subroutine d_oacc_csr_to_dev(v)
|
||||||
subroutine d_oacc_csr_sync_space(a)
|
implicit none
|
||||||
implicit none
|
real(psb_dpk_), intent(in) :: v(:)
|
||||||
class(psb_d_oacc_csr_sparse_mat), intent(inout) :: a
|
!$acc update device(v)
|
||||||
if (allocated(a%val)) then
|
end subroutine d_oacc_csr_to_dev
|
||||||
call d_oacc_create_dev(a%val)
|
|
||||||
end if
|
subroutine d_oacc_csr_to_host(v)
|
||||||
if (allocated(a%ja)) then
|
implicit none
|
||||||
call i_oacc_create_dev(a%ja)
|
real(psb_dpk_), intent(in) :: v(:)
|
||||||
end if
|
!$acc update self(v)
|
||||||
if (allocated(a%irp)) then
|
end subroutine d_oacc_csr_to_host
|
||||||
call i_oacc_create_dev(a%irp)
|
|
||||||
end if
|
subroutine i_oacc_csr_to_dev(v)
|
||||||
contains
|
implicit none
|
||||||
subroutine d_oacc_create_dev(v)
|
integer(psb_ipk_), intent(in) :: v(:)
|
||||||
implicit none
|
!$acc update device(v)
|
||||||
real(psb_dpk_), intent(in) :: v(:)
|
end subroutine i_oacc_csr_to_dev
|
||||||
!$acc enter data copyin(v)
|
|
||||||
end subroutine d_oacc_create_dev
|
subroutine i_oacc_csr_to_host(v)
|
||||||
subroutine i_oacc_create_dev(v)
|
implicit none
|
||||||
implicit none
|
integer(psb_ipk_), intent(in) :: v(:)
|
||||||
integer(psb_ipk_), intent(in) :: v(:)
|
!$acc update self(v)
|
||||||
!$acc enter data copyin(v)
|
end subroutine i_oacc_csr_to_host
|
||||||
end subroutine i_oacc_create_dev
|
|
||||||
end subroutine d_oacc_csr_sync_space
|
|
||||||
|
|
||||||
subroutine d_oacc_csr_sync(a)
|
|
||||||
implicit none
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), target, intent(in) :: a
|
|
||||||
class(psb_d_oacc_csr_sparse_mat), pointer :: tmpa
|
|
||||||
integer(psb_ipk_) :: info
|
|
||||||
|
|
||||||
tmpa => a
|
|
||||||
if (a%is_dev()) then
|
|
||||||
call d_oacc_csr_to_host(a%val)
|
|
||||||
call i_oacc_csr_to_host(a%ja)
|
|
||||||
call i_oacc_csr_to_host(a%irp)
|
|
||||||
else if (a%is_host()) then
|
|
||||||
call d_oacc_csr_to_dev(a%val)
|
|
||||||
call i_oacc_csr_to_dev(a%ja)
|
|
||||||
call i_oacc_csr_to_dev(a%irp)
|
|
||||||
end if
|
|
||||||
call tmpa%set_sync()
|
|
||||||
end subroutine d_oacc_csr_sync
|
|
||||||
|
|
||||||
subroutine d_oacc_csr_to_dev(v)
|
|
||||||
implicit none
|
|
||||||
real(psb_dpk_), intent(in) :: v(:)
|
|
||||||
!$acc update device(v)
|
|
||||||
end subroutine d_oacc_csr_to_dev
|
|
||||||
|
|
||||||
subroutine d_oacc_csr_to_host(v)
|
|
||||||
implicit none
|
|
||||||
real(psb_dpk_), intent(in) :: v(:)
|
|
||||||
!$acc update self(v)
|
|
||||||
end subroutine d_oacc_csr_to_host
|
|
||||||
|
|
||||||
subroutine i_oacc_csr_to_dev(v)
|
|
||||||
implicit none
|
|
||||||
integer(psb_ipk_), intent(in) :: v(:)
|
|
||||||
!$acc update device(v)
|
|
||||||
end subroutine i_oacc_csr_to_dev
|
|
||||||
|
|
||||||
subroutine i_oacc_csr_to_host(v)
|
|
||||||
implicit none
|
|
||||||
integer(psb_ipk_), intent(in) :: v(:)
|
|
||||||
!$acc update self(v)
|
|
||||||
end subroutine i_oacc_csr_to_host
|
|
||||||
|
|
||||||
|
|
||||||
end module psb_d_oacc_csr_mat_mod
|
end module psb_d_oacc_csr_mat_mod
|
||||||
|
|
||||||
|
|||||||
Loading…
Reference in New Issue