Minutes QUDA call 2015-04-09

• first update to develop branch, reminder for everyone to try to merge develop into feature branches on a regular basis

• dynamic_cover_invert is about to be ready to be merged, Alex will create a pull request soon (needs to merge develop in the branch first)

• Alexei reported on issues when running with cuda-memcheck. Might be MPI related. He will check again and create an issue.

• HISQ EigCG does not work with mixed precision, full precision and other fermion action look good. Might be an issue with the combination of staggered and mixed precision, similar to issues we had for CG ???

  • Alexei will keep looking into this and create an issue if necessary.

• MILC Interface does not expose reconstruction of links as an option. We should discuss with MILC whether this should be an option. Mathias will create an issue to track this discussion.

• For changes to the MILC interface that require changes to MILC: Please use the repository at https://github.com/lattice/milc_qcd . This is our QUDA fork of Carleton's MILC repository. Create a feature branch here and we can then submit a pull request so Carleton can pull that into MILC.

• Alexei reminded us to get the regression tests started. We will have a separate point in out next call.

• Alexei will work on optimization of GMRES and communication avoid orthogonalization to bring performance to a good level.

• NEXT CALL April 23, 2015

  ** FOLLOW-UP** on cuda-memcheck, HISQ mixed precision EigCG, regression tests