Move fermionforce eo noneo comparision (Unit Test #331)


Added by Christopher Pinke almost 5 years ago. Updated almost 5 years ago.


Status:Done Start date:22 Oct 2012
Priority:Normal Due date:
Assignee:Christopher Pinke % Done:

100%

Category:-
Target version:-

Description

The test F_fermion_compare_noneo_eo.cpp should either be removed or moved into opencl_module_hmc_test


Related issues

related to CL2QCD - Test Setup #209: Precision Test of HMC Done 17 Oct 2011
blocks CL2QCD - Unit Test #325: test opencl_module_hmc Done 22 Oct 2012

History

Updated by Christopher Pinke almost 5 years ago

  • Status changed from New to In Progress

Updated by Christopher Pinke almost 5 years ago

With a4034f260d0c15d0992e70c2e4dbcc5486500c83 I created a comparision test in the hmc test module.

3 of 4 tests pass, the last one does not because there is a problem with the input vectors. If they are random, it is nontrivial to map the rnd numbers to be at the same sites. Therefore, I use the convert-to/from-eoprec kernels for this. However, the to-kernel gives strange results (the two vectors are the same), while the from kernel gives the same vectors (judging from their squarenorm), but a result of the force kernel that differs by .4%.

It think this might be left like this until the convert kernels have been tested!

  • % Done changed from 0 to 80

Updated by Christopher Pinke almost 5 years ago

In 1891e3d15db2a733ce339896271168c49bbe4b2d I removed the old executable.

Updated by Christopher Pinke almost 5 years ago

With 9e983d9c3336be5c80bd3d12795fca905e9e4cf1 I corrected the "error" with the input vectors: The rnd generator was initialized for every spinor, thus the resulting vector had only equal components. I also added an additional switch since one can either convert to or from eoprec. There are now 4 tests using rnd input, and all fail by ~1% between eo and non-eo result. This should not have something to do with the input vectors, since the same mechanism works fine for the fermionmatrix kernels.

  • % Done changed from 80 to 90

Updated by Christopher Pinke almost 5 years ago

Commit:57687c2229cc24396824bbe81afa6ccb167b9b1a fixes the "error" in the tests. It was just a twist in the convention of EVEN and ODD, basically interchanging even and odd solved the problem.
This was not an error in the kernel, but in its usage.

All tests pass.

  • % Done changed from 90 to 100
  • Status changed from In Progress to Done

Also available in: Atom PDF