SEACAS Stray character prevents compilation #100
Comments
|
Thanks, I will fix. |
|
Great, thanks! |
|
What compiler were you using for this? It looks like it is getting confused by the C++ comments instead of C /* */ comments... |
|
Just gcc 4.7.2 |
|
The issue is that the file contained c++ comments '//' instead of c comments '/* */' and the compiler was using the -std=c90 option which seems to ignore the c++ comments which means it was treating the text following the comments as valid code which doesn't work very well for copyright notices and most other comment text. Comment characters have been converted to c format and comitted to Sierra repository. Will be synced to Trilinos soon. |
|
Oh ok I misunderstood the error message (and I rarely compile any C code to run into this ...). Thanks for identifying and fixing. |
|
I will make every effort I can to get another Sierra to Trilinos sync done today. I'm currently in the middle of the integration though so doing so right now isn't ideal. |
|
Shall I push a short term fix into Trilinos so that our downstream tests like albany and Nalu don't break in the nightlies? Its simple enough to do? |
|
You will have to clear that with @bmpersc ; I'm not sure what that does to his syncing process. |
|
I did talk to him, and he is trying to get a sync done today. |
|
This is fixed closing. |
Fixes trilinos#6937 Addresses CDOFA trilinos#100 This is an issue to track and attach a PR for the removal of `LD_PRELOAD` from the `jsrun_wrapper` The original implementation of the `jsrun_wrapper` did not have any `LD_PRELOAD`. The variable was added to enable unit tests to succeed when they allocated memory before calling `MPI_Init`. (see trilinos#6724 ). When codes allocate memory, the Cuda Hooks will interact with libpami (so MPI knows where the memory lives). This requires symbols to be loaded during `MPI_Init` (namely libpami). When a code allocates before `MPI_Init`, they will get an obscure error: (perhaps others) ``` CUDA Hook Library: Failed to find symbol mem_find_dreg_entries. ``` This is because the cuda hook library assumes MPI has been initialized. By adding an `LD_PRELOAD` to the wrapper, this allows memory allocations to find the needed symbols, but we are clearly outside the way Spectrum intended the system to be used. This shouldn't block ATDM apps, as they have been functioning up to this point. This may cause a headache for some unit tests, but it would be good to make sure developers are aware that most cuda-aware MPI implementations tack on extra rules regarding initializing the library.++ (see note below) We decided to leave the `LD_PRELOAD` in place until other testing issues were ironed out. This issue and PR should be blocked until ATDM devops decides they are ready to move forward. By removing LD_PRELOAD several unit tests will crash with errors about missing symbols. The fix will be to ensure that `MPI_Init` is called before `Kokkos::initialize`, which can be easier said than done. ++ Note: this pattern of requiring `MPI_Init` first is not new. E.g., OpenMPI w/UCX caused these obscure bugs that arose from allocs before initialize. trilinos#5033 (comment) @bartlettroscoe
Hi
after the current update of seacas I get the following errors:
Deleting the license header from both
Fixes that, so I guess its just an issue with a windows character or so coming in.
@trilinos/seacas
The text was updated successfully, but these errors were encountered: