Minutes of the Sankt-Augustin meeting (November 26, 2003)

(S. Valcke - 28/11/2003)

People who attended the meeting:

From NEC-CCRL, Rene Redler, Hubert Ritzdorf; from SGI Deutschland, Reiner Vogelsang; from CERFACS, Damien Declat and Sophie Valcke.

• Remark on Fortran unitsets: some models may not be able to list a priori the sets of Fortran units used. In that case, they will have to give the (hopefully finite) list of units possibly used (e.g. 1 to 100).

• The SMIOC API routines need to be secured (error handling); - Sophie.

• In TYPE grid, all variables related to the lats, lons, and z should be declared DOUBLE (not REAL). - Sophie

• In type_smioc.F90, all structures need to be reorganized in order to group the different types of variables together. - Sophie

• In type_smioc.F90, add transi_in_id and ig_orig_transi_id in type
  PSMILe_in_origin and transi_out_id and ig_dest_transi_id in type transient_out - Sophie

• It should be made clear to wp4b that the GUI must include the possiblity of importing and modifying already existing SCC and SMIOCs (corresponding to already assembled configurations). - Sophie

• It makes sense to give to the user the choice between a full 3D interpolation (e.g. nearest-neighbour, trilinear, conservative remapping) and a 2D1D interpolation (i.e. for each target point, 2D interpolations -independent or not, i.e. using different or same weights and addresses- are performed on the source upper and lower levels, and then a 1D linear interpolation is performed in the vertical).

• As long as no automatic reconstruction of the vertical levels in terms of meters is included in the coupler, the 2D1D will be the only possible interpolation for source and target grids which vertical levels are not expressed in meters; in that case, the units of the vertical axis have to be the same in the source and the target, and have to be suitable for linear interpolation (e.g. ln (p0/p)). The full 3D interpolation will be available only for grids which vertical levels are expressed in meters.

• The 3D nearest-neighbour and conservative remapping can be chosen for any grid type; the 3D trilinear can be chosen only for structured grids (reglonlatvrt, irrlonlat_regvrt, irrlonlatvrt, irrlonlat_sigmavrt, reglonlat_sigmavrt). The 2D1D may be chosen only if the grid is structured vertically, and if each vertical level is independent of the horizontal position (i.e. z(k): reglonlatvrt, irrlonlat_regvrt, unstructlonlat_regvrt).

• Regarding the interpolation, the source and target masks may be considered in the neighborhood search (i.e. in the PSMILe) or only in the interpolation per se (i.e. in the Transformer). If the mask is considered in the PSMILe, then id_mask=0, and ila_mask does not contain any infomation; if the mask is considered in the Transformer, then id_mask=1, and ila_mask does contain a mask (see PSMILe_Trs_set_epio3d_xx below).

• For the nearest-neighbour interpolation, the source mask may be considered in the neighborhood search (i.e. in the PSMILe) or in the interpolation per se (i.e. in the Transformer). The choice will be left to the user (see ``use_masked" in smioc.dtd, and ``iga_arg5" in type PSMILe_interp in type_smioc.F90).

• For other interpolations (i.e. bilinear, conservative remapping), the source mask is never considered in the neighborhood search. However, the user should be allowed to decide what happens regarding the interpolation in the Transformer when one source neighbour is masked (i.e. no value may be calculated for the target point, or non masked nearest-neighbour may be used, or ...).

• For all interpolations except the conservative remapping, the target mask is always considered in the neighborhood search (i.e. in the PSMILe), which means that no search is done for masked target points.

• Combinations: the combinations will take place either in the source PSMILe (if all fields to combine come from the same source) or in the target PSMILe (if the fields to combine come from different sources), but not in the Transformer.

• Complete notes_transformations.txt with the above remarks. - Sophie

• The routines used to transfer the EPIO and neighbours from the PSMILe to the Driver do not depend on the type of interpolation chosen by the user; they may depend on the type of the source and target partitions. If the source or target or both grids are regular in the vertical (lon(i,j) and lat (i,j)), some optimization in the neighbourhood search and EPIO definition should be implemented. The transfer of EPIO and neighbours from the PSMILe to the Transformer may also be optimised/reduced; we still have to decide and agree on particular PSMILe_Trs_set_epioxx and PSMILe_Trs_give_neighborsxx for those optimised cases. - Hubert, Damien.

• If ordering of the EPIO and neighbours information transfered as a 1D array is important, Fortran ordering is used.

• Units of latitude and longitude arrays used as argument in
  PSMILe_Trs_set_epioxx.F90 should be clarified. - Damien

• Convention for mask arrays used as argument in PSMILe_Trs_set_epioxx.F90 should be verified (1=not masked, 0=masked). - Damien

• prismtrs.inc should be merged with PSMILe.inc and appropriate modifications should be done in Driver/Transformer. - Damien

• PSMILe and Transformer treatment of user1D/2D/3D interpolation is not clear and will be specified later. - Sophie, Hubert

• Independently of the source or target lats and lons precision, the interpolation weights are always declared DOUBLE. - Damien

• The Transformer should be able to handle source and target fields declared integer, real, double (with possibly not the same precision on both sides) - Damien

• Conservative remapping: the PSMILe will transfer, for each target cell, a subdomain of source cells possibly intersecting it. The exact intersection will be calculated in the Transformer (with line integrals as in the SCRIP library). If cells of one target partition intersect source cells located in different source partitions, the source EPIO (now identifying cells and not points) of one source partition will be defined so to include all intersecting source cells.

• Transformer parallelisation: The conclusion from the discussion is that the Transformer should certainly be parallelised (on a EPIO basis) in order not to be a bottleneck on all types of platforms. However, as the Transformer does not perform any global operation and thanks to its design (based on independent EPIO treatment), this should not be a difficult task.

• API for PSMILe-Transformer interaction routines

  Note: Discussion is still taking place on the API definition - see my mail 27/11/2003.

- PSMILe_Trs_set_epio3d_real and PSMILe_Trs_set_epio3d_double

! Routine called by source process to define source EPIO or target EPIO
! INPUT ARGUMENTS
INTEGER, INTENT (In)                             :: il_epio_type
! = PRISM_EPIOS = 2820 if called to define a source EPIO or 
! = PRISM_EPIOT = 2821 if called to define a target EPIO (see prismtrs.inc)
INTEGER, INTENT (In)                             :: il_corresp_comp_id
! Global Id of corresponding component
INTEGER, INTENT (In)                             :: il_corresp_proc_id
! Global Id of corresponding process
INTEGER, INTENT (In)                             :: il_size
! Total EPIO size
REAL, DIMENSION(il_size), INTENT (In)            :: rla_lat
! Units of rla_lat ?
! For PSMILe_Trs_set_epio3d_double, dla_lat will be declared DOUBLE
REAL, DIMENSION(il_size), INTENT (In)            :: rla_lon
! Units of rla_lon ?
! For PSMILe_Trs_set_epio3d_double, dla_lon will be declared DOUBLE
REAL, DIMENSION(il_size), INTENT (In)            :: ila_z
! Units are the ones used in prism_set_points
! For PSMILe_Trs_set_epio3d_double, dla_z will be declared DOUBLE
INTEGER, INTENT (IN)                             :: id_mask
! Flag to indicate if the array ila_mask really contains 
! a mask (id_mask = 1) or not (id_mask = 0) 
INTEGER, DIMENSION(il_size), INTENT (In)         :: ila_mask
! Mask convention?
! OUTPUT ARGUMENTS
INTEGER, INTENT (Out)               :: id_epio_id
! Returned EPIO Id
INTEGER, INTENT (Out)               :: il_err   
! Returned error value

- PSMILe_Trs_give_neighbors3d

! Routine called by source process to define neighbours
! INPUT ARGUMENTS:
INTEGER, INTENT (In)                     :: il_epio_id
! Corresponding EPIO Id
INTEGER, INTENT (In)                     :: il_epiot_size
! Size of the target EPIO
INTEGER, INTENT (In)                     :: il_nb_nei  
! Number of neighbors
INTEGER, DIMENSION(il_nb_nei*il_epiot_size), INTENT (In)  :: ila_neighbors
! For all target EPIO points, neighbor addresses in the source EPIO
! OUTPUT ARGUMENTS:
INTEGER, INTENT (Out)                    :: il_err   
! Returned error value

- PSMILe_Trs_put_int, PSMILe_Trs_put_real, PSMILe_Trs_put_double

! INPUT ARGUMENTS:
INTEGER, INTENT (In)                        :: il_transient_out_id
! Transient out global Id (transi_out_id in type_smioc.F90) 
INTEGER, INTENT (In)                        :: il_epio_id
! Source EPIO id
INTEGER, INTENT (In)                        :: il_epios_size
! Size of source EPIO 
REAL, DIMENSION(il_epios_size), INTENT (In) :: rla_src_field
! ila_src_field declared INTEGER for PSMILe_Trs_put_int
! dla_src_field declared DOUBLE for PSMILe_Trs_put_double
! OUTPUT PARAMETERS:
INTEGER, INTENT (Out)                       :: il_err   
! Returned error value

-PSMILe_Trs_get_int, PSMILe_Trs_get_real, PSMILe_Trs_get_double

! INPUT ARGUMENTS:
INTEGER, INTENT (In)                        :: il_transient_in_id
! Transient in global Id (transi_in_id in type_smioc.F90) 
INTEGER, INTENT (In)                        :: il_epio_id
! Target EPIO id
INTEGER, INTENT (In)                        :: il_epiot_size
! Size of target EPIO 
REAL, DIMENSION(il_epios_size), INTENT (In) :: rla_tgt_field
! ila_tgt_field declared INTEGER for PSMILe_Trs_get_int
! dla_tgt_field declared DOUBLE for PSMILe_Trs_get_double
! OUTPUT PARAMETERS:
INTEGER, INTENT (Out)                       :: il_err   
! Returned error value