-
Notifications
You must be signed in to change notification settings - Fork 101
Laphnn work
maddyscientist edited this page Sep 10, 2024
·
15 revisions
- add support for passing gauge field without anti periodic factors included and add them in QUDA (optional?)
- quibble: in performGaugeSmearQuda, if you don't ask for topo charge to be computing, it still outputs it, but output is nonsense
-
eigensolveQuda
on smeared config- expose boolean to eig pram struct whether to use smeared or not
- then in set Dirac param can set if needed to use smeared or not
-
plaqQuda
on smeared config - need to set
block_size = 1
to download the first evector - kappa normalization for 3-d Laplace is borked
- routine should return a bool to indicate if converged or not??
- eigenvalues returned in
eigensolveQuda
are not correct for trlm3d: only nev values, not one for each time: in EigenSolver::cleanUpEigensolver, there is a line evals.resize(n_conv); - eigenvalues are not in ascending order on each time (I think time slices that converged earlier are ordered, but the last several that converged last are not)
- get tuning working for time slices split across ranks
- in the short term switch off autotuning at the point of divergence and reenable when converged
- a little clean up of parameter checking (checks fail for fields not needed, like inv_type)
- will need an applyLaphPhaseConvention() to the Laph eigenvectors
-- the 0th color element for site (0,0,0) on each time slice is rephased to be real and positive (rephasing each ev is tantamount to changing the noise) -- this is already written for the CPU so this is low priority - the number of restarts with the quda version is significantly higher than the current chroma_laph version: try to understand why (on E250, chroma_laph takes just a handful, quda takes over 100)
- timing goal: E250 96^3 x 192 lattice, 768 eigvec:
-- summit CPU solves: MPI ranks = 1536 Total wall time 7781 secs -- time on summit just to READ the evs from file is 320 seconds for MPI ranks = 384. - Dirac-Pauli doesn't with the outer solver