|
|
|
|
Most of the subdirectories here contain material that was originally
|
|
|
|
|
distributed through the serial sparse blas SPARKER project, plus a
|
|
|
|
|
number of extensions of our own design. The original file spblas.f can
|
|
|
|
|
be downloaded from matisa.cc.rl.ac.uk; of course any bugs in our
|
|
|
|
|
implementation are our own to fix.
|
|
|
|
|
|
|
|
|
|
The main reference for the serial
|
|
|
|
|
sparse BLAS is:
|
|
|
|
|
Duff, I., Marrone, M., Radicati, G., and Vittoli, C.
|
|
|
|
|
Level 3 basic linear algebra subprograms for sparse matrices: a user
|
|
|
|
|
level interface
|
|
|
|
|
ACM Trans. Math. Softw., 23(3), 379-401, 1997.
|
|
|
|
|
|
|
|
|
|
We also include SMMP by Randolph E. Bank and
|
|
|
|
|
Craig C. Douglas na.bank@na-net.ornl.gov and
|
|
|
|
|
na.cdouglas@na-net.ornl.gov; we wrapped it in a Fortran 95 interface
|
|
|
|
|
with dynamic memory allocation.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
How do you add a new storage format? Here is your checklist.
|
|
|
|
|
1. You have to invent a name for it, a name that can fit into 3
|
|
|
|
|
letters.
|
|
|
|
|
2. You have to decide whether you need fields into INFOA() for it, and
|
|
|
|
|
invent their names/positions accordingly.
|
|
|
|
|
3. You have to add an estimate (as good as possible) for the memory
|
|
|
|
|
occupation into psb_cest.
|
|
|
|
|
4. You have to be able to generate it. For our purposes, the minimum
|
|
|
|
|
effort that still meets the requirements is the ability to convert
|
|
|
|
|
to/from COOrdinate; the relevant code (properly encapsulated in
|
|
|
|
|
subroutines) should be added into psb_Xcsdp; you can take Xcocr as
|
|
|
|
|
the model. You may or may not be able to do it "in place";
|
|
|
|
|
ipcoo2csr & friends do just that; if you can, please add it because
|
|
|
|
|
it may come in handy.
|
|
|
|
|
5. If you need to sort entries, it usually better to use imsrx than
|
|
|
|
|
isrx. This is because coefficients lists are likely to contain
|
|
|
|
|
multiple ordered sublists ("runs"), and the merge sort algorithm
|
|
|
|
|
implemented in imsrx is capable of taking advantage of them,
|
|
|
|
|
whereas the isrx quicksort cannot.
|
|
|
|
|
6. You have to be able to convert it back to COOrdinate. This is
|
|
|
|
|
usually not hard, because the code that does the MV multiply is
|
|
|
|
|
already (maybe implicitly) walking through the coefficients while
|
|
|
|
|
keeping track of both row and column index, thus it suffices to
|
|
|
|
|
copy them as you meet them, and then call psb_fixcoo (which sorts
|
|
|
|
|
and compacts the output) on the result.
|
|
|
|
|
7. You have to embed the MV and SV codes in the infrastructure; this
|
|
|
|
|
is done first in the xSWMM and xSWSM routines that do a select on
|
|
|
|
|
the FIDA value. Below this level you're free to do as you please.
|
|
|
|
|
8. You have to extend the output capability of psb_Xcsprt. Again this
|
|
|
|
|
is easy for the same reasons as in 6.
|
|
|
|
|
9. You have to extend psb_Xcsrws for DIAG preconditioner.
|
|
|
|
|
10. Provide an UPDATE functionality, and add it to psb_update_mod;
|
|
|
|
|
currently JAD is out of this.
|
|
|
|
|
11. You have to provide the query facilities used in PSB_XSPINFO
|
|
|
|
|
(included in base/modules/psb_spmat_type).
|
|
|
|
|
12. You have to provide the functionality to extract a block of
|
|
|
|
|
rows in psb_getrow_mod.f90. This is used in the GETROW/GETBLK/CLIP
|
|
|
|
|
chain; GETROW is (all that is) used by the ILU factorization.
|
|
|
|
|
13. You have to provide the GETDIAG functionality.
|
|
|
|
|
14. RWEXTD: what do we do here? Should we switch to/from COO?????
|
|
|
|
|
The current interface is ugly!
|
|
|
|
|
15. CSC is currently a second class citizen; it is only needed for
|
|
|
|
|
interfacing UMFPACK in the mld2p4 preconditioners.
|
|
|
|
|
|
|
|
|
|
Is this complete? I sure hope so....
|
