SaltyChiang
commented
7 months ago
- I wonder if we can have a single
applyWilsonfunction, that simply takes an extra template parameterdistance_pcto avoid having two different definitions ofapplyWilson.- Then instead of
if (d == 3) { out += ratio_fwd * (U * in.project(d, proj_dir)).reconstruct(d, proj_dir); } else { out += (U * in.project(d, proj_dir)).reconstruct(d, proj_dir); }we could use something like
out += ratio_fwd[d] * (U * in.project(d, proj_dir)).reconstruct(d, proj_dir);where
ratio_fwdwould be aconstexprarray that would trivially evaluate as1.0whendistance_pc = 1or ifd < 3. I think that would allow us to have a singleapplyWilsondefinition.
@maddyscientist That's sounds better than my implementation. Thanks and I'll try to do that these days.
- For the different .cu files: perhaps we could have a single .hpp file that contains all the boilerplate and have this file included in the stub files dslash_wilson.cu and dslash_wilson_distance.cu, for example, where the only different between the two is a booled template parameter. E.g., like we have dslash_coarse.cu and dslash_coarse_dagger.cu, etc.
I'll check these files, thank you for your advice.
Have you given any consideration to the light quark preconditioning technique outlined in the paper?
Sure. The light quark preconditioning described in the paper requires a 4D $\alpha(x)$. Notice the coefficient is $\frac{1}{\cosh[\alpha0(x\mu-L_\mu/2)]}$, which is the inverse of that in the heavy quark case. Preconditioning in space dimensions seems to make sense only if we solve the propagator on a point source, which is not a common situation. Also, you might notice that I force the alpha0 parameter to be positive here:
https://github.com/lattice/quda/blob/b9be5842d08c5eb911931f68ced621e9a4f7e293/lib/interface_quda.cpp#L2742-L2745
In fact, leaving the alpha0 negative will trigger a t-direction light quark preconditioning. (I use the sign of alpha0 to decide whether to apply the coefficient or the inverse) I've tried this with the CG solver, but don't see a significant improvement in iteration numbers. Using multigrid for light quarks should be a much better strategy in my opinion.
maddyscientist
We have been working on the bottom quark with the clover action recently. We cannot obtain a stable plateau in the effective mass diagram even though we have set the residual to be 1e-15. I implement the distance preconditioning algorithm proposed in https://arxiv.org/pdf/1006.4028.pdf.
Two parameters
distance_pc_alpha0anddistance_pc_t0are added toQudaInverParam, which follow Eq.(9) in the paper. I treated the preconditioning as a special solver instead of a new Dirac action, so I just modified the Wilson Dslash and didn't make a newDiracclass. The preconditioning is only enabled when callinginverQuda.I use a point source to solve propagators with different
The effective mass plot becomes much better with alpha0=1.5.
distance_pc_alpha0and draw the effective mass diagram generated from the pseudoscalar meson correlation functions. The effect of the preconditioning is shown below.The preconditioning is only enabled for CG inverter now due to a lack of testing. This could be enabled in more inverters later.
I created wilson_distance.cu, wilson_clover_distance.cu and wilson_clover_distance_preconditioned.cu, because the compilation time will be extremely long if I add distance and non-distance functions in one file.