Remove obsolete files
sfilippone
5 months ago
parent
f783478df3
commit
6236f3489c
@ -1,86 +0,0 @@
|
||||
submodule (psb_c_oacc_csr_mat_mod) psb_c_oacc_csr_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_c_oacc_csr_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_c_oacc_csr_sparse_mat), intent(in) :: a
|
||||
complex(psb_spk_), intent(in) :: alpha, beta
|
||||
complex(psb_spk_), intent(in) :: x(:,:)
|
||||
complex(psb_spk_), 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 = 'c_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_c_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_c_oacc_csr_csmm
|
||||
end submodule psb_c_oacc_csr_csmm_impl
|
||||
|
||||
@ -1,81 +0,0 @@
|
||||
submodule (psb_c_oacc_csr_mat_mod) psb_c_oacc_csr_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_c_oacc_csr_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_c_oacc_csr_sparse_mat), intent(in) :: a
|
||||
complex(psb_spk_), intent(in) :: alpha, beta
|
||||
complex(psb_spk_), intent(in) :: x(:)
|
||||
complex(psb_spk_), 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 = 'c_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_c_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_c_oacc_csr_csmv
|
||||
end submodule psb_c_oacc_csr_csmv_impl
|
||||
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_c_oacc_ell_mat_mod) psb_c_oacc_ell_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_c_oacc_ell_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_c_oacc_ell_sparse_mat), intent(in) :: a
|
||||
complex(psb_spk_), intent(in) :: alpha, beta
|
||||
complex(psb_spk_), intent(in) :: x(:,:)
|
||||
complex(psb_spk_), 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, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'c_oacc_ell_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_c_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, n
|
||||
do k = 1, nzt
|
||||
y(i, j) = y(i, j) + alpha * a%val(i, k) * x(a%ja(i, k), 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_c_oacc_ell_csmm
|
||||
end submodule psb_c_oacc_ell_csmm_impl
|
||||
@ -1,82 +0,0 @@
|
||||
submodule (psb_c_oacc_ell_mat_mod) psb_c_oacc_ell_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_c_oacc_ell_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_c_oacc_ell_sparse_mat), intent(in) :: a
|
||||
complex(psb_spk_), intent(in) :: alpha, beta
|
||||
complex(psb_spk_), intent(in) :: x(:)
|
||||
complex(psb_spk_), intent(inout) :: y(:)
|
||||
integer(psb_ipk_), intent(out) :: info
|
||||
character, optional, intent(in) :: trans
|
||||
|
||||
character :: trans_
|
||||
integer(psb_ipk_) :: i, j, m, n, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'c_oacc_ell_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_c_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
do j = 1, nzt
|
||||
y(i) = y(i) + alpha * a%val(i, j) * x(a%ja(i, j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_c_oacc_ell_csmv
|
||||
end submodule psb_c_oacc_ell_csmv_impl
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_c_oacc_hll_mat_mod) psb_c_oacc_hll_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_c_oacc_hll_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_c_oacc_hll_sparse_mat), intent(in) :: a
|
||||
complex(psb_spk_), intent(in) :: alpha, beta
|
||||
complex(psb_spk_), intent(in) :: x(:,:)
|
||||
complex(psb_spk_), 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, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'c_oacc_hll_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_c_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nhacks = (a%get_nrows() + a%hksz - 1) / a%hksz
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do k = 1, nhacks
|
||||
do i = a%hkoffs(k), a%hkoffs(k + 1) - 1
|
||||
y(a%irn(i), j) = y(a%irn(i), j) + alpha * a%val(i) * x(a%ja(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_c_oacc_hll_csmm
|
||||
end submodule psb_c_oacc_hll_csmm_impl
|
||||
@ -1,84 +0,0 @@
|
||||
submodule (psb_c_oacc_hll_mat_mod) psb_c_oacc_hll_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_c_oacc_hll_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_c_oacc_hll_sparse_mat), intent(in) :: a
|
||||
complex(psb_spk_), intent(in) :: alpha, beta
|
||||
complex(psb_spk_), intent(in) :: x(:)
|
||||
complex(psb_spk_), intent(inout) :: y(:)
|
||||
integer(psb_ipk_), intent(out) :: info
|
||||
character, optional, intent(in) :: trans
|
||||
|
||||
character :: trans_
|
||||
integer(psb_ipk_) :: i, j, m, n, hksz, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'c_oacc_hll_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_c_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
hksz = a%hksz
|
||||
nhacks = (a%get_nrows() + hksz - 1) / hksz
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
! This loop nest cannot be run with collapse, since
|
||||
! the inner loop extent varies.
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, nhacks
|
||||
do j = a%hkoffs(i), a%hkoffs(i + 1) - 1
|
||||
y(a%irn(j)) = y(a%irn(j)) + alpha * a%val(j) * x(a%ja(j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_c_oacc_hll_csmv
|
||||
end submodule psb_c_oacc_hll_csmv_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
|
||||
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_d_oacc_ell_mat_mod) psb_d_oacc_ell_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_d_oacc_ell_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_d_oacc_ell_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, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'd_oacc_ell_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_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, n
|
||||
do k = 1, nzt
|
||||
y(i, j) = y(i, j) + alpha * a%val(i, k) * x(a%ja(i, k), 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_ell_csmm
|
||||
end submodule psb_d_oacc_ell_csmm_impl
|
||||
@ -1,82 +0,0 @@
|
||||
submodule (psb_d_oacc_ell_mat_mod) psb_d_oacc_ell_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_d_oacc_ell_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_d_oacc_ell_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, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'd_oacc_ell_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_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
do j = 1, nzt
|
||||
y(i) = y(i) + alpha * a%val(i, j) * x(a%ja(i, j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_d_oacc_ell_csmv
|
||||
end submodule psb_d_oacc_ell_csmv_impl
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_d_oacc_hll_mat_mod) psb_d_oacc_hll_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_d_oacc_hll_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_d_oacc_hll_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, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'd_oacc_hll_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_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nhacks = (a%get_nrows() + a%hksz - 1) / a%hksz
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do k = 1, nhacks
|
||||
do i = a%hkoffs(k), a%hkoffs(k + 1) - 1
|
||||
y(a%irn(i), j) = y(a%irn(i), j) + alpha * a%val(i) * x(a%ja(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_hll_csmm
|
||||
end submodule psb_d_oacc_hll_csmm_impl
|
||||
@ -1,84 +0,0 @@
|
||||
submodule (psb_d_oacc_hll_mat_mod) psb_d_oacc_hll_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_d_oacc_hll_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_d_oacc_hll_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, hksz, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'd_oacc_hll_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_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
hksz = a%hksz
|
||||
nhacks = (a%get_nrows() + hksz - 1) / hksz
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
! This loop nest cannot be run with collapse, since
|
||||
! the inner loop extent varies.
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, nhacks
|
||||
do j = a%hkoffs(i), a%hkoffs(i + 1) - 1
|
||||
y(a%irn(j)) = y(a%irn(j)) + alpha * a%val(j) * x(a%ja(j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_d_oacc_hll_csmv
|
||||
end submodule psb_d_oacc_hll_csmv_impl
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_s_oacc_csr_mat_mod) psb_s_oacc_csr_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_s_oacc_csr_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_s_oacc_csr_sparse_mat), intent(in) :: a
|
||||
real(psb_spk_), intent(in) :: alpha, beta
|
||||
real(psb_spk_), intent(in) :: x(:,:)
|
||||
real(psb_spk_), 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 = 's_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_s_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_s_oacc_csr_csmm
|
||||
end submodule psb_s_oacc_csr_csmm_impl
|
||||
|
||||
@ -1,81 +0,0 @@
|
||||
submodule (psb_s_oacc_csr_mat_mod) psb_s_oacc_csr_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_s_oacc_csr_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_s_oacc_csr_sparse_mat), intent(in) :: a
|
||||
real(psb_spk_), intent(in) :: alpha, beta
|
||||
real(psb_spk_), intent(in) :: x(:)
|
||||
real(psb_spk_), 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 = 's_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_s_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_s_oacc_csr_csmv
|
||||
end submodule psb_s_oacc_csr_csmv_impl
|
||||
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_s_oacc_ell_mat_mod) psb_s_oacc_ell_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_s_oacc_ell_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_s_oacc_ell_sparse_mat), intent(in) :: a
|
||||
real(psb_spk_), intent(in) :: alpha, beta
|
||||
real(psb_spk_), intent(in) :: x(:,:)
|
||||
real(psb_spk_), 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, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 's_oacc_ell_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_s_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, n
|
||||
do k = 1, nzt
|
||||
y(i, j) = y(i, j) + alpha * a%val(i, k) * x(a%ja(i, k), 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_s_oacc_ell_csmm
|
||||
end submodule psb_s_oacc_ell_csmm_impl
|
||||
@ -1,82 +0,0 @@
|
||||
submodule (psb_s_oacc_ell_mat_mod) psb_s_oacc_ell_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_s_oacc_ell_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_s_oacc_ell_sparse_mat), intent(in) :: a
|
||||
real(psb_spk_), intent(in) :: alpha, beta
|
||||
real(psb_spk_), intent(in) :: x(:)
|
||||
real(psb_spk_), intent(inout) :: y(:)
|
||||
integer(psb_ipk_), intent(out) :: info
|
||||
character, optional, intent(in) :: trans
|
||||
|
||||
character :: trans_
|
||||
integer(psb_ipk_) :: i, j, m, n, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 's_oacc_ell_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_s_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
do j = 1, nzt
|
||||
y(i) = y(i) + alpha * a%val(i, j) * x(a%ja(i, j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_s_oacc_ell_csmv
|
||||
end submodule psb_s_oacc_ell_csmv_impl
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_s_oacc_hll_mat_mod) psb_s_oacc_hll_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_s_oacc_hll_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_s_oacc_hll_sparse_mat), intent(in) :: a
|
||||
real(psb_spk_), intent(in) :: alpha, beta
|
||||
real(psb_spk_), intent(in) :: x(:,:)
|
||||
real(psb_spk_), 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, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 's_oacc_hll_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_s_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nhacks = (a%get_nrows() + a%hksz - 1) / a%hksz
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do k = 1, nhacks
|
||||
do i = a%hkoffs(k), a%hkoffs(k + 1) - 1
|
||||
y(a%irn(i), j) = y(a%irn(i), j) + alpha * a%val(i) * x(a%ja(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_s_oacc_hll_csmm
|
||||
end submodule psb_s_oacc_hll_csmm_impl
|
||||
@ -1,84 +0,0 @@
|
||||
submodule (psb_s_oacc_hll_mat_mod) psb_s_oacc_hll_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_s_oacc_hll_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_s_oacc_hll_sparse_mat), intent(in) :: a
|
||||
real(psb_spk_), intent(in) :: alpha, beta
|
||||
real(psb_spk_), intent(in) :: x(:)
|
||||
real(psb_spk_), intent(inout) :: y(:)
|
||||
integer(psb_ipk_), intent(out) :: info
|
||||
character, optional, intent(in) :: trans
|
||||
|
||||
character :: trans_
|
||||
integer(psb_ipk_) :: i, j, m, n, hksz, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 's_oacc_hll_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_s_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
hksz = a%hksz
|
||||
nhacks = (a%get_nrows() + hksz - 1) / hksz
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
! This loop nest cannot be run with collapse, since
|
||||
! the inner loop extent varies.
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, nhacks
|
||||
do j = a%hkoffs(i), a%hkoffs(i + 1) - 1
|
||||
y(a%irn(j)) = y(a%irn(j)) + alpha * a%val(j) * x(a%ja(j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_s_oacc_hll_csmv
|
||||
end submodule psb_s_oacc_hll_csmv_impl
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_z_oacc_csr_mat_mod) psb_z_oacc_csr_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_z_oacc_csr_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_z_oacc_csr_sparse_mat), intent(in) :: a
|
||||
complex(psb_dpk_), intent(in) :: alpha, beta
|
||||
complex(psb_dpk_), intent(in) :: x(:,:)
|
||||
complex(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 = 'z_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_z_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_z_oacc_csr_csmm
|
||||
end submodule psb_z_oacc_csr_csmm_impl
|
||||
|
||||
@ -1,81 +0,0 @@
|
||||
submodule (psb_z_oacc_csr_mat_mod) psb_z_oacc_csr_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_z_oacc_csr_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_z_oacc_csr_sparse_mat), intent(in) :: a
|
||||
complex(psb_dpk_), intent(in) :: alpha, beta
|
||||
complex(psb_dpk_), intent(in) :: x(:)
|
||||
complex(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 = 'z_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_z_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_z_oacc_csr_csmv
|
||||
end submodule psb_z_oacc_csr_csmv_impl
|
||||
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_z_oacc_ell_mat_mod) psb_z_oacc_ell_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_z_oacc_ell_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_z_oacc_ell_sparse_mat), intent(in) :: a
|
||||
complex(psb_dpk_), intent(in) :: alpha, beta
|
||||
complex(psb_dpk_), intent(in) :: x(:,:)
|
||||
complex(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, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'z_oacc_ell_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_z_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, n
|
||||
do k = 1, nzt
|
||||
y(i, j) = y(i, j) + alpha * a%val(i, k) * x(a%ja(i, k), 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_z_oacc_ell_csmm
|
||||
end submodule psb_z_oacc_ell_csmm_impl
|
||||
@ -1,82 +0,0 @@
|
||||
submodule (psb_z_oacc_ell_mat_mod) psb_z_oacc_ell_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_z_oacc_ell_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_z_oacc_ell_sparse_mat), intent(in) :: a
|
||||
complex(psb_dpk_), intent(in) :: alpha, beta
|
||||
complex(psb_dpk_), intent(in) :: x(:)
|
||||
complex(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, nzt
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'z_oacc_ell_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_z_ell_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nzt = a%nzt
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
do j = 1, nzt
|
||||
y(i) = y(i) + alpha * a%val(i, j) * x(a%ja(i, j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_z_oacc_ell_csmv
|
||||
end submodule psb_z_oacc_ell_csmv_impl
|
||||
@ -1,86 +0,0 @@
|
||||
submodule (psb_z_oacc_hll_mat_mod) psb_z_oacc_hll_csmm_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_z_oacc_hll_csmm(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_z_oacc_hll_sparse_mat), intent(in) :: a
|
||||
complex(psb_dpk_), intent(in) :: alpha, beta
|
||||
complex(psb_dpk_), intent(in) :: x(:,:)
|
||||
complex(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, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'z_oacc_hll_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_z_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
nxy = min(size(x,2), size(y,2))
|
||||
nhacks = (a%get_nrows() + a%hksz - 1) / a%hksz
|
||||
|
||||
!$acc parallel loop collapse(2) present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do i = 1, m
|
||||
y(i,j) = beta * y(i,j)
|
||||
end do
|
||||
end do
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do j = 1, nxy
|
||||
do k = 1, nhacks
|
||||
do i = a%hkoffs(k), a%hkoffs(k + 1) - 1
|
||||
y(a%irn(i), j) = y(a%irn(i), j) + alpha * a%val(i) * x(a%ja(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_z_oacc_hll_csmm
|
||||
end submodule psb_z_oacc_hll_csmm_impl
|
||||
@ -1,84 +0,0 @@
|
||||
submodule (psb_z_oacc_hll_mat_mod) psb_z_oacc_hll_csmv_impl
|
||||
use psb_base_mod
|
||||
contains
|
||||
module subroutine psb_z_oacc_hll_csmv(alpha, a, x, beta, y, info, trans)
|
||||
implicit none
|
||||
class(psb_z_oacc_hll_sparse_mat), intent(in) :: a
|
||||
complex(psb_dpk_), intent(in) :: alpha, beta
|
||||
complex(psb_dpk_), intent(in) :: x(:)
|
||||
complex(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, hksz, nhacks
|
||||
logical :: tra
|
||||
integer(psb_ipk_) :: err_act
|
||||
character(len=20) :: name = 'z_oacc_hll_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_z_hll_sparse_mat%spmm(alpha, x, beta, y, info, trans)
|
||||
else
|
||||
hksz = a%hksz
|
||||
nhacks = (a%get_nrows() + hksz - 1) / hksz
|
||||
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, m
|
||||
y(i) = beta * y(i)
|
||||
end do
|
||||
! This loop nest cannot be run with collapse, since
|
||||
! the inner loop extent varies.
|
||||
!$acc parallel loop present(a, x, y)
|
||||
do i = 1, nhacks
|
||||
do j = a%hkoffs(i), a%hkoffs(i + 1) - 1
|
||||
y(a%irn(j)) = y(a%irn(j)) + alpha * a%val(j) * x(a%ja(j))
|
||||
end do
|
||||
end do
|
||||
endif
|
||||
|
||||
call psb_erractionrestore(err_act)
|
||||
return
|
||||
|
||||
9999 call psb_error_handler(err_act)
|
||||
return
|
||||
|
||||
end subroutine psb_z_oacc_hll_csmv
|
||||
end submodule psb_z_oacc_hll_csmv_impl
|
||||
Loading…
Reference in New Issue