Update reweighting routines for change in kappa

[Finkenrath]

Update reweighting routines to enable reweighting in "kappa"

• update global kappa during the calculations
• reset global kappa after factors are calculated
• update kappa within DDalphaAMG interface
  Note: Check if similar actions are needed within the QUDA interface (kappa is usually not changed during HMC / inversions etc.)

[Marcogarofalo]

I checked that QUDA in the light sector works out of the box. For the non-degenerete it was not working, I add the interface using the multishift solver. The result seems unchanged to me

[Marcogarofalo]

I change the lines here such that now with a:

• NDCLOVERDETRATIO monomial we call sw_trace_nd
• CLOVERDETRATIORW monomial we call sw_trace
• NDDETRATIO monomial there is no trlog to compute, also this now can call QUDA inverter

[Marcogarofalo]

now we pass the test that the reweighting factor of an NDCLOVERDETRATIO monomial with 2KappaEpsBar = 0 is equal to the one a CLOVERDETRATIORW with 2KappaMu = 2Kappamubar. The trlog part is exactly the same, the stochastic part is only the same up to stochastic noise anyhow for the parameter I tested the first two digits coincide.

In more detail NDCLOVERDETRATIO monomial

BeginMonomial NDCLOVERDETRATIO
  Timescale = 1
  CSW = 1.69
  # nominator parameters
  2KappaEpsBar  = 0.00
  2Kappamubar   = 0.06
  kappa = 0.1394265
  # denominator parameters
  2KappaEpsBar2 = 0.0
  2Kappamubar2  = 0.06001
  kappa2 = 0.1394265
  AcceptancePrecision =  1.e-26
  Name = ndcloverdetratio
  # useexternalinverter = quda
  UseSloppyPrecision = single
  RefinementPrecision = single
  solver= cgmmsnd
EndMonomial

BeginMonomial CLOVERDETRATIORW
  Timescale = 1
  CSW = 1.69
  # nominator parameters
  2KappaMu  = 0.06000
  kappa = 0.1394265
  rho = 0.0
  # denominator parameters
  2KappaMu2 = 0.06001
  ## for reweighting only rho=rho2 makes sense
  rho2 = 0.0
  kappa2 = 0.1394265
  AcceptancePrecision =  1.e-26
  Name = cloverdetratiorw
  useexternalinverter = quda
  UseSloppyPrecision = single
  #solver = mg
  solver = cg
EndMonomial

[Finkenrath]

Good spotted.

[kostrzewa]

@Marcogarofalo for NDCLOVERDETRATIO you can use

  UseSloppyPrecision = half
  RefinementPrecision = half

for an extra speed-up as this will be basically like running double-half mixed-precision CG. (with some small overhead)