Deflation

Jump to bottom Edit New page

Carsten Urbach edited this page Sep 20, 2013 · 10 revisions

Deflation part in tmLQCD

bugs found / questions

deflation does not work together with SSE 2/3 (28.8.2013)
FIXED (20.9.2013) the scalar version (no MPI) does not work properly, it produces either a wrong little Dirac operator or the block boundaries are not treated correctly. With MPI different parallelisations but the same global block numbers give identical results, though.
- little_D appears to be identical in between scalar and MPI version
does it work for odd number of blocks in some or all directions per MPI task? More specifically, does the little GCR even/odd solver work properly then? --> only even no of blocks per MPI task and direction are allowed. This is checked for now explicitly in the code.

new features

changed the preconditioning in FGMRES as suggested in arXiv:1303.1377
included a deflated PCG solver with deflation subspace generation
added a CG even/odd smoother
included a deflated FGMRES solver for Qsq