The test test_gfp2_2_2g fails
Imported issue: Initially reported by @gaudry in https://gforge.inria.fr/tracker/?group_id=2065&aid=21767
The test called test_gfp2_2_2g as it is written passes most of the times, but sometimes it fails.
In fact, to be honnest, it always fails, in the sense that when it ""passes"", the computed values for virtual logarithms are wrong (all zeros, or just plain wrong).
Several reasons have been identified:
-
Galois, as it is implemented can not work (or at least we see no theoretical reason why it should). It must be disabled.
-
There might be some problem with the free relations like in bug 21761, so they must be disabled.
-
Last but not least, on side 1, the SM currently chosen by the code is trivial on the units.
For Galois, an approach would be to see the corresponding knowledge as free relations, but they should be on a single side. Let's call them half-free relations. Each of them would include SM for the corresponding side, and not the whole SMs. Testing this would require to modify quite a bit the filter + SM machinery.
For the unfortunate choice of the SM, there is a brand new patch by E. Thomé (872b301a) to allow to play with various choices.
Note that we were able to have the test pass with a result that seems correct with another choice of SM.
There was no need to put the following two lines :
tasks.linalg.allow_zero_on_rhs = 1
tasks.reconstructlog.partial = true
so that many consistency checks were done.