OpenMP target extension of the tridiagonal solver operators.
The host tdsops_t is built as usual and its coefficient arrays are
copied once into device arrays, as cuda_tdsops_t does, so the kernels
read them in place instead of mapping them on every launch.
The device arrays are not freed: there is no finaliser, because the
intrinsic assignment of a structure constructor would finalise the
temporary and release storage the copy still points at. Operators are
built once at start-up and live for the whole run, so construct them in
place with omptgt_tdsops_init rather than by assignment.
| Type | Visibility | Attributes | Name | Initial | |||
|---|---|---|---|---|---|---|---|
| real(kind=dp), | public, | allocatable, dimension(:) | :: | dist_fw |
fw/bw phase back subs. the auxiliary factors |
||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | dist_bw |
fw/bw phase back subs. the auxiliary factors |
||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | dist_sa |
fw/bw phase back subs. the auxiliary factors |
||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | dist_sc |
fw/bw phase back subs. the auxiliary factors |
||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | dist_af |
fw/bw phase back subs. the auxiliary factors |
||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | thom_f | |||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | thom_s | |||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | thom_w | |||
| real(kind=dp), | public, | allocatable, dimension(:) | :: | thom_p | |||
| real(kind=dp), | public, | allocatable | :: | stretch(:) | |||
| real(kind=dp), | public, | allocatable | :: | stretch_correct(:) | |||
| real(kind=dp), | public, | allocatable | :: | coeffs(:) | |||
| real(kind=dp), | public, | allocatable | :: | coeffs_s(:,:) | |||
| real(kind=dp), | public, | allocatable | :: | coeffs_e(:,:) | |||
| real(kind=dp), | public | :: | alpha |
Compact scheme coeffs |
|||
| real(kind=dp), | public | :: | a |
Compact scheme coeffs |
|||
| real(kind=dp), | public | :: | b |
Compact scheme coeffs |
|||
| real(kind=dp), | public | :: | c | = | 0._dp |
Compact scheme coeffs |
|
| real(kind=dp), | public | :: | d | = | 0._dp |
Compact scheme coeffs |
|
| real(kind=dp), | public | :: | beta | = | 0._dp |
LHS 2nd off-diagonal (pentadiag LHS only) |
|
| real(kind=dp), | public | :: | beta_lhs_s | = | 0._dp |
j=1 backward-sub beta (0/2β Neumann, β otherwise) |
|
| logical, | public | :: | periodic | ||||
| logical, | public | :: | pentadiag | = | .false. |
.true. for pentadiagonal LHS schemes Solve with the serial Thomas algorithm instead of the distributed one. The distributed solver truncates a neighbour's influence at the subdomain edge, which is invisible for the well-conditioned derivative operators but not for the low-pass filter, whose system is only marginally diagonally dominant as alpha approaches 0.5. Thomas is serial along the line, so this is only valid when the direction is not decomposed. |
|
| logical, | public | :: | prefer_thomas | = | .false. | ||
| integer, | public | :: | n_tds |
Tridiagonal system size |
|||
| integer, | public | :: | n_rhs |
Right-hand-side builder size |
|||
| integer, | public | :: | move | = | 0 |
move between vertices and cell centres |
|
| integer, | public | :: | n_halo |
number of halo points |
|||
| real(kind=dp), | public, | pointer, contiguous | :: | dist_fw_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | dist_bw_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | dist_sa_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | dist_sc_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | dist_af_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | thom_f_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | thom_s_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | thom_w_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | thom_p_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | stretch_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | stretch_correct_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | coeffs_dev(:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | coeffs_s_dev(:,:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | coeffs_e_dev(:,:) | => | null() | |
| real(kind=dp), | public, | pointer, contiguous | :: | penta_z_dev(:,:) | => | null() |
Columns of A_np^{-1}*[e_1|e_2|e_{n-1}|e_n] for the cyclic pentadiagonal solve. They depend on the operator alone, so they are built once here where the CUDA kernel rebuilds them in every block. |
| procedure, public :: lowpass | |
| procedure, public :: stagder_1st | |
| procedure, public :: interpl_mid | |
| procedure, public :: deriv_2nd | |
| procedure, public :: deriv_1st | |
| procedure, public :: preprocess_thom | |
| procedure, public :: preprocess_penta_dist | |
| procedure, public :: preprocess_dist |
Solves A_np*z_k = e_{p_k} for p = 1, 2, n-1, n with the non-periodic pentadiagonal LU factors, as thread 1 of der_penta_periodic does in the CUDA backend.
| Type | Intent | Optional | Attributes | Name | ||
|---|---|---|---|---|---|---|
| class(tdsops_t), | intent(in) | :: | tdsops |
Initialises tdsops in place. See tdsops_t for the arguments.
| Type | Intent | Optional | Attributes | Name | ||
|---|---|---|---|---|---|---|
| type(omptgt_tdsops_t), | intent(inout) | :: | tdsops | |||
| integer, | intent(in) | :: | n_tds | |||
| real(kind=dp), | intent(in) | :: | delta | |||
| character(len=*), | intent(in) | :: | operation | |||
| character(len=*), | intent(in) | :: | scheme | |||
| integer, | intent(in) | :: | bc_start | |||
| integer, | intent(in) | :: | bc_end | |||
| real(kind=dp), | intent(in), | optional | :: | stretch(:) | ||
| real(kind=dp), | intent(in), | optional | :: | stretch_correct(:) | ||
| integer, | intent(in), | optional | :: | n_halo | ||
| character(len=*), | intent(in), | optional | :: | from_to | ||
| logical, | intent(in), | optional | :: | sym | ||
| real(kind=dp), | intent(in), | optional | :: | c_nu | ||
| real(kind=dp), | intent(in), | optional | :: | nu0_nu | ||
| real(kind=dp), | intent(in), | optional | :: | filter_alpha |