Bugfix in cloverdet_monomial

[kottnad]

Fixes a bug in cloverdet monomial that is triggered if both of the following conditions are true
1.) external_library == QUDA_LIB && external_inverter == QUDA_INVERTER && no strict residual check is required
2.) one or more cloverrat monomials are present

[kostrzewa]

@kottnad thanks!

Wasn't the problem already triggered with just the CLOVERDET present when g_strict_residual_check=0 and (external_library == QUDA_LIB && external_inverter == QUDA_INVERTER) == true ?

For completeness: it is caused in the acceptance step by the call to Qm which requires sw_invert to have been called.

[kottnad]

Indeed, it is also enough to just have a single CLOVERDET to trigger it. However, running a two-flavor simulation including one or more CLOVERDETRATIOs is fine. Going to 2+1 with a CLOVERRAT is how it showed up originally.

[kostrzewa]

Indeed. I think we actually need to perform a high statistics run to evaluate the effect this had on production simulations.

The way I understand it, sw_inv as used by Qm in the acceptance step for the CLOVERDET would have been slightly wrong (supplied by whichever CLOVERDETRATIO acceptance step ran right before the CLOVERDET acceptance step), resulting in an unquantified effect on dH and thus the sampled probability distribution. The effect must have been (very) small, otherwise we would have lost acceptance.

[kostrzewa]

I would suggest an nf=2 twisted clover simulation on a 24c48 volume based on one of the old nf=2 ensembles at beta=2.1

Best would probably be cA2.60.24

[kostrzewa]

Actually, now that we discussed, given that sw_inv under the problematic condition will come from the CLOVERDETRATIO and this, usually, will have set the same g_mu, the resulting sw_inv will also be correct for the CLOVERDET acceptance step.

However, when the last function that filled sw_inv happens to be from an NDCLOVERRAT, then I don't have a feeling for how badly the resulting sw_inv would affect the acceptance step of CLOVERDET...

[kostrzewa]

cscs-ci run default

[kostrzewa]

system-wide CSCS maintenance...