QUDA call 5th March, 2015

Attendees: Alexei, Jim, Mathias, Thorsten, Balint, Nuno, Steve, Heechang, Mike

1. Scope and schedule of the calls. General consensus to have developer focussed calls every two weeks. Exception is next meeting which will be March 26th to avoid clash with GTC.

2. Reminder: quda development model. Homework for everyone: read blogs posts listed on here to get an idea of the development model we are switching to. In summary: master branch should always correspond to latest official release. Use pull requests for merging new features into main development branch. Unit tests for all new features. Hot fixes done by branching from master, pull requesting back into developer and master branches (and make new release).

3. Development updates

• Alexei:
  • eigCG, now fully supports low precision eigenvectors (single precision). All e-vectors in single precision, reduces memory footprint hugely. Works well with Wilson and twisted-mass.
  • MILC eigCG interface: tested for staggered 16^3x48 speedup is 4x, 32^3x48 with 378 e-vectors only gives a 2x speedup.
  • GMRES-DR: full precision working well. mixed-precision in development, works until 1e-5 residual, investigating. Aim to merge into development about April. Future work for DD, multi-shift, flexibile, multigrid.
• Heechang: ramping up on staggered domain decomposition, picking up where Chris Winteroud left off.

4. Upcoming tasks:

• Mike to clean up stale branches. Won't be done until post GTC.
• Plan to resurrect QUDA mailing list for all developer communication: [email protected].

Next meeting: 26th March 11:00 PDT / 14:00 EDT.