MueLu failing to compile on cuda-9 #2683
Comments
bartlettroscoe
added
type: bug
bartlettroscoe
added this to the Initial cleanup of new ATDM builds of Trilinos milestone
|
@trilinos/muelu |
|
Looks like the issue is here in kokkos_kernels. Is there some issue with include paths? |
|
@bathmatt The nightly CUDA 8 MueLu builds are ok. I will see if I can reproduce under CUDA 9. |
|
@trilinos/kokkos-kernels @srajama1 |
|
This is most likely an issue of including both "math.h" and which will screw stuff up in ways dependent on include order. This is outside of KokkosKernels control unfortunately. Solution: NEVER EVER EVER include math.h in C++ files. Because math.h is C. |
|
There's 167 of math.h in all sorts of c++ file files. List is included Here are ones in kokkos directories... ./packages/kokkos-kernels/example/fenl/fenl_impl.hpp |
|
should I sed script all of them away in one big push? |
|
So, looking a bit deeper, lots of those are amp_math.h and quadmath.h. Are those o.k.? |
|
this is just <math.h> files. 110 files but only kokkos-kernels examples are there |
|
Looking at the list most are in examples or tests. For your use case can it be belos/src/* ? |
|
For me, the issue is that I can't compile trilinos with cuda 9 and muelu kokkos flags set. It reports errors in muelu which look like they are in kokkos-kernels. The rest doesn't matter to me but is a means to an end. |
|
@bathmatt I had no luck reproducing the problem on waterman with the CUDA 9 modules loaded -- trilinos built without error with the script On the CEE lan, I module loaded SPARC's CUDA 9 environment: Trilinos compilation fails in Kokkos: |
|
@jhux2 one thing to try for the CEE lan issue is to try checking that |
|
@ndellingwood Thanks for the tip. The variable was indeed set incorrectly, but even pointing to Do you know if there another platform where CUDA 9 is available that I can try building on? |
|
@jhux2 - CUDA9 is available on Hansen and Shiller as well as the POWER machines. |
Yea, that has been official Trilinos policy for at least 12+ years. If you go to: there is a link to: and there is the policy:
@trilinos/framework team. Could we consider adding some simple static analysis checks to the auto PR testing for things like this? Other projects do that (e.g. INL MOOSE). Otherwise, how do we educate Trilinos developers about policies like this? Is there not a checklist for this stuff? |
Can't reproduce on Shiller either. I may not have the correct Trilinos configure script. |
|
Is cuda9 on shiller? It works in 8 |
@bathmatt, there are no CUDA9 builds for the ATDM Trilinos builds yet. We have been focusing on duplicating all of the EMPIRE Trilinos builds running on jenkins-srn and we did not see any CUDA 9 builds there. We can add a CUDA 9.0 build on shiller/hansen. The only issue is that then we need to extend the keyword parsing to not just look for |
|
I have a cuda9 build on shiller going, if it fails I'll send the mods to Ross's procedure to get it going. |
|
This seems to be only tickled by my system configuration and not on shiller. I think we can close this, but first @bartlettroscoe can you confirm that this should work with the scripts? Basically, swap out the cuda 9 modules after one sets the env. Seemed to do the right thing. IF there is no obvious error thre I will look at reopening once I can reproduce on a different public system |
@bathmatt, basically, this is just replacing the the compilers, etc. This should work. Should we go ahead and add a |
|
I think we can close this until I nail down exactly what is diffrent on my system causing this. |
bartlettroscoe
added
the
PA: Data Services
Not sure how long it's been failing
The text was updated successfully, but these errors were encountered: