[petsc-dev] better regular testing on accelerators

[Smith, Barry F.]

> On Jun 12, 2019, at 9:58 AM, Jed Brown <jed at jedbrown.org> wrote:
> 
> Would it be sufficient to add the CUDA arguments to PETSC_OPTIONS when running the test suite on those machines?

  You mean -vec_type cuda -mat_type cuda -dm_vec_type cuda -dm_mat_type cuda ? We should definitely try it, it may break something we'll see. It will miss all the code that uses directly MatCreateAIJ....() but then maybe we should change that code :-)


  Barry

> 
> "Smith, Barry F. via petsc-dev" <petsc-dev at mcs.anl.gov> writes:
> 
>>   In order to get better testing on the accelerators I think we need to abandon the -vec_type cuda approach scattered through a handful of examples and instead test ALL examples that are feasible automatically with the various accelerator options.  I think essentially any examples that use AIJ are feasible for testing (those that use BAIJ, SBAIJ, Ell are not) I am not sure if there is an automatic way to determine all of these cases. Labeling all such test cases manually would likely miss some and be out of date immediately.
>> 
>>   Any thoughts?
>> 
>>   Thanks
>> 
>>   Barry
>> 
>> Maybe we could short circuit the issue by having a mode of configuring or compiling or running where MatCreate_AIJ and VecCreate_ are bypassed directly to the accelerator cousins (this is not completely trivial because the cousins generally construct the basic ones and then convert themselves to the cousin).