-
Notifications
You must be signed in to change notification settings - Fork 41
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
SPARSE_seq* tests fail on PowerPC #83
Comments
Subsequently, the same happens with
|
This is concerning, but I cannot reproduce these segfaults on my machine. |
Now we know that OpenBLAS is not to blame here – same failures with vecLibFort (Accelerate). |
I have built and enabled
|
|
Apparently that
|
I suspect we were getting false test passes, all those with 0.05 sec pass time.
MPI errors look the same:
This is when running under old MPI, but under new everything looks as bad. |
Strange.
if this fails, can you then run valgrind on it? |
Been away from my PPC hardware. I will update on this in a couple of days. Valgrind is not within options though, it is broken on PPC. Apple GDB works. |
@pghysels Here it is:
|
@pghysels It looks like all MPI tests fail simply because wrong arguments are passed:
This should be fixable, I guess. |
@pghysels Without MPICH, aren’t there just tolerance threshold issues? On 32-bit precision is expected to be lower, also Darwin uses IBM long double (if that is relevant).
Without MPICH results are okayish
UPD. I see it was better initially, I added an unneeded flag now to modify alignments. |
I'm sorry for the delay. Since all the failed tests are for SPARSE_seq*, not for SPARSE_seq_scotch*, that makes me think there is something wrong with the METIS ordering. You can try your code/test that fails and switch the ordering from METIS to Scotch. |
METIS can be configured with 32 bit or 64 bit integer support. But that shouldn't make a difference, both should work. One common source of problems is that scotch installs the |
@pghysels I have suddenly found crash logs. Does this suggest something?
|
There is a recursive function to compute the number of levels in the tree. The number of levels is too large (at least 508) so the code probably crashes with a stack overflow. We should avoid this recursion. But the tree should ideally only have ~log(N) levels, so something is wrong with the tree. The tree is determined by the matrix sparsity pattern and the matrix ordering (which is computed by METIS, or scotch). |
Everything else passes,
SPARSE_seq*
all fail:10.6.8, ppc32, gcc 12.2.0, OpenBLAS 0.3.21 (linking to
vecLibFort
seems not to work, therefore had to link toOpenBLAS
).Build logs:
build_log.txt
Tests log:
tests_log.txt
The text was updated successfully, but these errors were encountered: