Multipole convergence order test is sensitive to roundoff

The integration convergence test performed by Multipole is sensitive to roundoff, as the relative error is ~1e-7 in the Simpson case. This is important only during the correctness test, not for general use of the thorn. The computation of the convergence order includes a subtraction of the exact from the numerical result, and this loses many digits of precision in the final convergence order. This causes the simpson test to fail on several machines. So, while it is useful to have the convergence order output, this is not a good regression test. The attached set of patches output the integration results, and raise the tolerance of the convergence order to 1e-3. The tolerance for the integration results is left at the default. The tests all pass on Datura, though they also did before this change.

OK to commit?

Keyword:

Comments (4)