Per offline conversations, the default precision in the MILC Makefile is now 2, corresponding to double precision. This reflects the increasing need for higher precision in MD evolution to observe the expected 4th order scaling. Single precision is still available by setting PRECISION = 1 at the top of the MILC Makefile.
A new script in the ks_spectrum directory that simplifies compiling ks_spectrum_hisq with GPU offload via QUDA, including options for opting into multigrid support.
Both the new script and the compile script in ks_imp_rhmc have additional error checking, making sure various environment variables are appropriately set.
@detar , I noticed a few leftover changes from very early commits to the feature/staggMG branch, in particular in generic_ks/io_scidac_ks_eigen.c and generic_u1/io_u1lat.c. Should these be rolled back? Let me know.
This PR introduces a few changes in MILC:
2, corresponding todoubleprecision. This reflects the increasing need for higher precision in MD evolution to observe the expected 4th order scaling. Single precision is still available by settingPRECISION = 1at the top of the MILC Makefile.ks_spectrumdirectory that simplifies compilingks_spectrum_hisqwith GPU offload via QUDA, including options for opting into multigrid support.ks_imp_rhmchave additional error checking, making sure various environment variables are appropriately set.Documentation for MILC HISQ MG is available at https://github.com/lattice/quda/wiki/HISQ-MG-for-Measurements , where there are a few workable examples. More details will be provided soon.
@detar , I noticed a few leftover changes from very early commits to the
feature/staggMGbranch, in particular ingeneric_ks/io_scidac_ks_eigen.candgeneric_u1/io_u1lat.c. Should these be rolled back? Let me know.