Closed maddyscientist closed 5 years ago
Thank you @detar for the merge.
The default is to use no reconstruction, so it would have worked fine by default previously. What broke was if the environment variable QUDA_MILC_HISQ_RECONSTRUCT=13
was set, then the long link would pick up the wrong scaling. It now works as it should regardless of naik-epsilon value.
This pull request fixes support for long-link reconstruction when using HISQ fermions with QUDA offload.
ks_param
struct, the regular CG solver and dslash calls did not have this: with this in mind I have augmented thefn_links_t
struct to include the naik_epsilon value when the links are constructed. This is then passed to the MILC-QUDA interface.I've tested this code with su3_hisq_rhmd, and I see complete agreement now between running with
QUDA_MILC_HISQ_RECONSTRUCT=13
andQUDA_MILC_HISQ_RECONSTRUCT_18
(well to solver tolerance anyway). Both for the multi-shift solver and the CG used for psi-bar-psi at the end of the trajectories.This pull request is dependent on matching changed being made in QUDA (https://github.com/lattice/quda/pull/720).