`plotter.interpolate_current` result depends significantly on CPU architecture
I noticed that the test
kwant/tests/test_plotter.py::test_current_interpolation in c7a13533 fails on i386 while it passes on amd64. On both architectures, the environment is an up-to-date Debian testing, so that the software environment should be identical.
I narrowed down the problem to the following script. It calls
plotter.interpolate_current with the same hard-coded data, but the result is different.
I also attach a script that contains the output (
j0[y_axis]) from both architectures and produces the following plot.
As one can see, the results on both architectures are similar, but vastly bigger than what seems acceptable due to differences between the architectures.
Once this bug is fixed, the original test should be reviewed. Check out the history to see how I dealed with the failing test for 1.4.0.
@jbweston, you know this code best, can you have a look?