The GMSH file format has changed in version 4. FeeLLGood currently understands files made by GMSH version 2. Compatibility with files from version 3 has not been tested.

The unit tests ut_tetra.cpp, ut_node.cpp and ut_pt3D.cpp sometimes fail with errors above the tolerance. The statistics of the errors suggest they are rounding errors. Here is a graph, collected after a large number of runs of make test (I did not count how many), showing the empirical tail distribution (complementary cumulative distribution function):

• the abscissa is the reported error
• the ordinate is the rank of the error, sorted from largest (rank = 1) to smallest

For ut_tetra.cpp and ut_node.cpp, the issue can be solved by rising the tolerance. The case of ut_pt3D.cpp seems more worrisome, as its errors can get quite large. This test, unlike the other two, measures a relative error. As shown in the figure below, the relative error is large only when the numbers being compared are very small:

A possible solution could be to switch to measuring an absolute error. Alternatively, we could try to figure out how the rounding errors are supposed to scale.

This test is comparing triple products. It can be expected for the error to scale like the product of the norms of the vectors involved. The following graph shows that, when the errors are divided by this scaling factor, they indeed remain below a few DBL_EPSILON. This graph and the next one are based on another series of tests (10⁶ tests total).

This scaling makes the tail distribution better behaved: