|
|
|
|
1. Perhaps we should switch completely to external indices being of
|
|
|
|
|
kind LPK and internal indices being of type IPK, and the two morph
|
|
|
|
|
independently in int32/int64 (although admittedly LPK=32 and IPK=64
|
|
|
|
|
does not make much sense)
|
|
|
|
|
|
|
|
|
|
2. Change sizeof_ constant names accordingly to sizeof_ipk and
|
|
|
|
|
sizeof_lpk.
|
|
|
|
|
|
|
|
|
|
3. Should we define a psb_l_vect_type? But then, if I==L how can we
|
|
|
|
|
distinguish? Answer: if I==L the two vect type are still considered
|
|
|
|
|
different, even when internally they are the same.
|
|
|
|
|
|
|
|
|
|
4. So, let's rewrite under these rules:
|
|
|
|
|
psb_mpk_: Always 32 bits, used for MPI related stuff.
|
|
|
|
|
psb_ipk_: Can be 32 or 64 bits, always used for "local" indices and
|
|
|
|
|
sizes
|
|
|
|
|
psb_lpk_: Can be 32 or 64 bits, always used for "global" indices
|
|
|
|
|
and sizes, must be psb_lpk_ >= psb_ipk_
|
|
|
|
|
psb_epk_: always 64 bits, used for SIZEOF & friends.
|
|
|
|
|
|
|
|
|
|
5. Let's define the SND/RCV/SUM/MAX & friends in terms of M and E, the
|
|
|
|
|
compiler will remap I and L onto them automatically
|
|
|
|
|
|
|
|
|
|
6. Similar for sort; except for the inner routines of heap, where we
|
|
|
|
|
provide heap types I_IDX_HEAP, they have to be written
|
|
|
|
|
independently beccause the encapsulated types are always
|
|
|
|
|
different.
|
|
|
|
|
|
|
|
|
|
7. For communication stuff: let us define psb_i_base_vect and
|
|
|
|
|
psb_l_base_vect; the communication routines will work in terms of
|
|
|
|
|
them, then remap onto the array routines, which are going to be
|
|
|
|
|
written in terms of E and M. This guarantees working no matter what
|
|
|
|
|
the exact values for I and L are.
|
