Checkerboarding of Wilson and Clover operators looks different

[martin-ueding]

While browsing the code, I noticed some strangeness around the checkerboarding.

In wilson.h, there is the following operator:

void operator()(FourSpinorBlock *res, const FourSpinorBlock *in, int isign) {
    D->dslash(tmp, in, u[1], isign, 1);
    D->dslashAChiMinusBDPsi(res, tmp, in, u[0], mass_factor_alpha,
                            mass_factor_beta, isign, 0);
}

Depending on the mapping {0, 1} → {even, odd} (Peter says 0 = odd), this would then be an odd-odd operator.

However, in clover.h, it is the other way around:

void operator()(FourSpinorBlock *res, const FourSpinorBlock *in, int isign) {
    double beta = (double)0.25;

    D->dslash(tmp, in, u[0], invclov, isign, 0);
    D->dslashAChiMinusBDPsi(res, tmp, in, u[1], clov, beta, isign, 1);
}

From my understanding, this makes it an even-even operator. The solver does not care which operator was passed in. So one of the versions uses the gauge field on the wrong checkerboard‽

[bjoo]

Hi Martin, This is a good catch: I note that in the testDslashFull.cc I also test for this: dslash(chi2,u_test,psi, isign, 1); dslash(ltmp,u_test,chi2, isign, 0); chi2[rb[0]] = massFactorpsi - betaFactorltmp;

However, the chroma convention is for this to be on cb[1] which is why I guess I test it that way in the clover tests in testClovDslashFull.cc

  M( chi_s[1],
     psi_s[1],
     isign);

  qdp_unpack_spinor<> (chi_s[0], chi_s[1], chi, geom);

  // Apply QDP Dslash
  chi2 = zero;

  dslash(chi2,u_test,psi, isign, 0);
  invclov_qdp_ap.apply(clov_chi2, chi2, isign, 0);
  dslash(ltmp,u_test,clov_chi2, isign, 1);

  clov_qdp_ap.apply(chi2, psi, isign, 1);

  chi2[rb[1]] -= betaFactor*ltmp;

In Chroma I use Clover exclusively, and so probably the Dslash wrapper won’t work (If I even have one, I may have just integrated the Clover) — the chroma convention is that the Schur preconditioned operator lives on rb[1] (odd).

I should add to the to-do to make the Wilson the same as the Clover.

Best, B

On Apr 21, 2017, at 10:51 AM, Martin Ueding notifications@github.com wrote:

While browsing the code, I noticed some strangeness around the checkerboarding.

In wilson.h, there is the following operator:

void operator()(FourSpinorBlock res, const FourSpinorBlock in, int isign) { D-> dslash(tmp, in, u[1], isign, 1 ); D-> dslashAChiMinusBDPsi(res, tmp, in, u[0 ], mass_factor_alpha, mass_factor_beta, isign, 0 ); }

Depending on the mapping {0, 1} → {even, odd} (Peter says 0 = odd), this would then be an odd-odd operator.

However, in clover.h, it is the other way around:

void operator()(FourSpinorBlock res, const FourSpinorBlock in, int isign) {

double beta = (double)0.25 ;

D->

dslash(tmp, in, u[0], invclov, isign, 0 ); D-> dslashAChiMinusBDPsi(res, tmp, in, u[1], clov, beta, isign, 1 ); }

From my understanding, this makes it an even-even operator. The solver does not care which operator was passed in. So one of the versions uses the gauge field on the wrong checkerboard‽

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

---

Dr Balint Joo High Performance Computational Scientist Jefferson Lab 12000 Jefferson Ave, Suite 3, MS 12B2, Room F217, Newport News, VA 23606, USA Tel: +1-757-269-5339, Fax: +1-757-269-5427 email: bjoo@jlab.org

[bjoo]

fixed in 9031fbe by adding cb-index to the EvenOdd Operators and solvers