site and s loop into the kernels. This will save on function call overhead and
guarantee L2 prefetching strategy is right since OMP can't distribute the
sub-chunks of work.
turned up problems on the BlueWaters Cray.
Gets 75MB/s from home filesystem on parallel configuration read. Need to make the RNG IO parallel,
and also to look at aggregating bigger writes for the parallel write.
Not sure what the home filesystem is.
Thought I had already committed these.
Believe I have got the Gparity fermion force working.
* tests/Test_gpdwf_force.cc -- correctly predicts dS for two flavour pseudofermion
based on a small dt update of U field.
* tests/Test_hmc_EODWFRatio_Gparity.cc -- ran 1 trajectory on 8^4 with dH=0.21.
Need to accumulate a full plaquette log to believe fully which will take some hours of run time.
So valence sector looks ok.
FermionOperatorImpl.h provides the policy classes.
Expect HMC will introduce a smearing policy and a fermion representation change policy template
param. Will also probably need multi-precision work.
* HMC is running even-odd and non-checkerboarded (checked 4^4 wilson fermion/wilson gauge).
There appears to be a bug in the multi-level integrator -- <e-dH> passes with single level but
not with multi-level.
In any case there looks to be quite a bit to clean up.
This is the "const det" style implementation that is not appropriate yet for clover since
it assumes that Mee is indept of the gauge fields. Easily fixed in future.
Allows multi-precision work and paves the way for alternate BC's and such like
allowing for example G-parity which is important for K pipi programme.
In particular, can drive an extra flavour index into the fermion fields
using template types.
6000 matmuls CG unprec
2000 matmuls CG prec (4000 eo muls)
1050 matmuls PGCR on 16^3 x 32 x 8 m=.01
Substantial effort on timing and logging infrastructure
Tanh/Zolo * (Cayley/PartFrac/ContFrac) * (Mobius/Shamir/Wilson)
Approx Representation Kernel.
All are done with space-time taking part in checkerboarding, Ls uncheckerboarded
Have only so far tested the Domain Wall limit of mobius, and at that only checked
that it
i) Inverts
ii) 5dim DW == Ls copies of 4dim D2
iii) MeeInv Mee == 1
iv) Meo+Mee+Moe+Moo == M unprec.
v) MpcDagMpc is hermitan
vi) Mdag is the adjoint of M between stochastic vectors.
That said, the RB schur solve, RB MpcDagMpc solve, Unprec solve
all converge and the true residual becomes small; so pretty good tests.
