Hopefully fixed Intel-regression in release mode: if the data points are very close to linear the cubic interpolation may fail and generate a NaN in highly optimised code.