milc-qcd / milc_qcd

MILC collaboration code for lattice QCD calculations
Other
34 stars 33 forks source link

ks_spectrum_hisq solve does not terminate with error_for_propagator 0 rel_error_for_propagator 1e-05 #25

Closed james-simone closed 5 years ago

james-simone commented 5 years ago

Hi,

The hisq solver never reports convergence in application ks_spectrum_hisq with inputs:

== KSsolveSet ==

max_cg_iterations 2000 max_cg_restarts 80 check sourceonly momentum_twist 0 0 0 time_bc periodic precision 2 source 0 number_of_propagators 1

== propagator 0: KSsolveElement ==

mass 0.01 naik_term_epsilon 0 error_for_propagator 0 rel_error_for_propagator 1e-05 fresh_ksprop forget_ksprop

== KSsolveSet ==

max_cg_iterations 2000 max_cg_restarts 80 check yes momentum_twist 0 0 0 time_bc periodic precision 2 source 1 number_of_propagators 6

== propagator 1: KSsolveElement ==

mass 0.0363 naik_term_epsilon 0 error_for_propagator 0 rel_error_for_propagator 1e-05 fresh_ksprop forget_ksprop

Code in production is Git short tag ac033d6 branch. Float precision=double compiled with gc6 for cpu (no QUDA) gcc6 mvapich2_2.2

On startup, these options are printed: MIMD version 7.8.1 Machine = QMP (portable), with 96 nodes Host(0) = dsg0103 Username = (null) start: Fri Apr 26 10:54:27 2019

Options selected... Generic double precision C_GLOBAL_INLINE DBLSTORE_FN D_FN_GATHER13 FEWSUMS KS_MULTICG=HYBRID HISQ_REUNIT_ALLOW_SVD HISQ_REUNIT_SVD_REL_ERROR = 1e-08 HISQ_REUNIT_SVD_ABS_ERROR = 1e-08

james-simone commented 5 years ago

Convergence to specified rel_error_for_propagator has been fixed in later versions of the code.