radau5 solver does not support sundials 64 bit index

[jschueller]

if sundials is compiled with SUNDIALS_INDEX_SIZE=64 (default) radau5 will fail in superlu_mt method get_perm_c
with message error from colamd

[PeterMeisrimelModelon]

Can you provide some more information such as the error message and sundials version?

[jschueller]

yes, thats for all sundials versions >=5:
tests.test_examples.Test_Examples.test_cvode_with_jac_sparse ... COLAMD failed at line 58 in file /tmp/superlu_mt-4.0.1/SRC/get_perm_c.c

[PeterMeisrimelModelon]

Just for clarity, your initial comment states Radau5, the test you quote is CVode, so you observe both solvers being affected?

[jschueller]

maybe the error message does not match the output, because only radau uses superlu, right ?

[PeterMeisrimelModelon]

CVode + sparse also uses superLU.

[jschueller]

oh then that might be it too, sorry if the report wasnt super clear

[PeterMeisrimelModelon]

I can reproduce the issue with CVode , looking into it.

[PeterMeisrimelModelon]

This is somewhat of a known issue, see https://github.com/modelon-community/Assimulo/blob/master/setup.py#L394

Sticking with SUNDIALS_INDEX_SIZE=32 is likely the best option for now. If one used a superLU configured to match SUNDIALS_INDEX_SIZE=64, there might be issues with Radau5's sparse solver.

/Peter

[modelonrobinandersson]

@PeterMeisrimelModelon @jschueller is this issue solved by #108?

[jschueller]

no

[jschueller]

but with #108 now we configure sundials with the correct index type (32 bits) so that newer versions can be tested

[PeterMeisrimelModelon]

Also just for completeness, the issue here seems to be that the sparse jacobian callbacks for both Radau5 & CVode (+ IDA?) both work with generic Cython int + we currently assume superLU is using 32 bit indices. So, a few things not really in sync.