scalapack: query lwork from ppocon

[davydden]

i was trying to fix scalapack_05 on some testers https://cdash.kyomu.43-1.org/testSummary.php?project=1&name=scalapack%2Fscalapack_05.mpirun%3D4.debug&date=2018-02-15
and was digging into the code.

Looks like lwork is wrong, IBM states this as:

 lwork ≥ 2np0 + 2nq0 + max(2, max(nb(max(1, iceil(nprow-1, npcol))), nq0 + nb(max(1, iceil(npcol-1, nprow)))))

mb = MB_A
nb = NB_A
iroff = mod(ia-1, mb)
icoff = mod(ja-1, nb)
iarow = mod(RSRC_A + (ia-1)/mb, nprow)
iacol = mod(CSRC_A + (ja-1)/nb, npcol)
np0 = NUMROC(n+iroff, mb, myrow, iarow, nprow)
nq0 = NUMROC(n+icoff, nb, mycol, iacol, npcol)

and netlib-scalapack as

       LWORK >= 2*LOCr(N+MOD(IA-1,MB_A)) + 2*LOCc(N+MOD(JA-1,NB_A))+
             MAX( 2, MAX(NB_A*CEIL(NPROW-1,NPCOL),LOCc(N+MOD(JA-1,NB_A)) +
             NB_A*CEIL(NPCOL-1,NPROW)) ).

in our case IA=JA=1 so things are simpler, but still there were no max and alike in

int lwork = 2 * n_local_rows + 3 * n_local_columns + column_block_size;

To be on the safe side we can query the right number from ppocon.

p.s. this does not fix the float test for me, it still hangs on MPI=4 run with a 32x32 matrix with block-size 32 (process grid will consist of one core only here). double works just fine, so does float with 1 MPI core...

[davydden]

/run-tests

[bangerth]

OK to merge once the tester is happy.

[davydden]

For the record, i know what's going on with that scalapack_05 test. It's another Fortran symbol conflict due to faulty pArpack. Should be pslamch (I submitted a patch, will find out if it helps soon). As soon as I disable pArpack in deal.ii, the test runs fine. @BenBrands , who extended ScaLAPACK wrappers to run with float and modified that test, did not see this issue as he is not building dealii with pArpack.

[bangerth]

Oh, are you saying that both SCALAPCK and PArpack have functions called pslamch and that the linker chooses one or the other randomly? If so, Ouch!

Can you post a link to your patch, if it's in some public repo, for posterity?

[bangerth]

Ah, found it: opencollab/arpack-ng#50

[davydden]

Oh, are you saying that both SCALAPCK and PArpack have functions called pslamch and that the linker chooses one or the other randomly? If so, Ouch!

yeap, Arpack used to have pdlamch but i fixed it in opencollab/arpack-ng#21

the fix for float is opencollab/arpack-ng#85