Merged fermionmatrix kernels including dslash broken? (Unit Test #323)


Added by Christopher Pinke almost 7 years ago. Updated over 4 years ago.


Status:New Start date:25 Jan 2013
Priority:Normal Due date:25 Jan 2013
Assignee:Christopher Pinke % Done:

0%

Category:-
Target version:-

Description

With cdf48d884ad7aa9d3bd068382ee0ec00af45b296, the merged fermion kernels including dslash seem to be broken.

One can see that in the failing test for the dslash + m_tm_inverse_sitediagonal(_minus) kernels. If one sets mu=0, the first two of these should give the same result as a pure dslash!

The other kernel, dslash + gamma5, passes the tests, however, it seems strange that both tests give the same result, where in the one the argument to the dslash is even and odd in the other. Thus, they should use different gaugelinks and give different results!


Related issues

precedes CL2QCD - Defect #416: Merged spinor kernels broken Feedback 27 Jan 2013
follows CL2QCD - Feature #412: Make own test module for merged fermion kernel Done 24 Jan 2013

Associated revisions

Revision 5e5e4f0c
Added by Christopher Pinke over 4 years ago

removed merged dslash and gamma5 kernel
refs #323

Revision 49ac64a6
Added by Christopher Pinke over 4 years ago

fixed last commit
refs #323

Revision 91254495
Added by Christopher Pinke over 4 years ago

removed leftovers from dslash_AND_gamma5 kernel
refs #323

History

Updated by Christopher Pinke over 6 years ago

  • Due date set to 25 Jan 2013
  • Start date changed from 22 Oct 2012 to 25 Jan 2013

Updated by Christopher Pinke over 4 years ago

Actually, it does not really make sense to merge gamma5 and dslash anyway.
The fermionmatrix gamma_5*Aee (and similar Aee_minus) has (in the pure Wilson case) a structure like
  • dslash
  • dslash
  • saxpy
  • gamma5
    Hence, merging gamma5 with dslash would result in
  • dslash
  • dslash + gamma5 merged
  • gamma5 * 1
  • saxpy

which would yield no benefit as gamma5 is still called alone!

Updated by Christopher Pinke over 4 years ago

I removed the dslash+gamma5 kernel.

Now remaining are the merged kernels of the dslash with the diagonal matrices in the twisted mass case.
Actually, these test fail (at least partly) on the CPU! Hence, there seems to be a more fundamental problem!

  • Assignee set to Christopher Pinke

Updated by Christopher Pinke over 4 years ago

For the moment, speedup is not needed for twisted mass fermions as they are not used.

  • Priority changed from High to Normal

Also available in: Atom PDF