numpy.tanh gives different results on Windows and Linux

[tadeu]

Here is one example: the result for tanh(-0.35898351519709742) is -0.34431837261747228 on Linux and -0.34431837261747222 on Windows. Windows have the most approximate result.

[eric-wieser]

What does np.float64(math.tanh(-0.35898351519709742)) give on your linux machine?

[njsmith]

In general, numpy doesn't attempt to provide exactly the same results on different platforms. It's essentially impossible for floating point code. Even two different builds with the same compiler can give slightly different results depending on optimization level. In this case we're using the platform-provided tanh, so it's not even the same code doing the calculation.

One thing we can do if some platform's tanh is particularly inaccurate is to blacklist the platform-provided version in favor of our own code. I can't tell if that's the case here, though – just how far off are the Linux and windows results from the true answer? Which one is more accurate? What version of python, windows, and Linux are you using?

[tadeu]

@eric-wieser

>>> np.float64(math.tanh(-0.35898351519709742))
-0.34431837261747228

@njsmith Ok, got it that it's caused by the platform-provided functions. The results are usually far off by one ULP. Got some differences in the np.exp function too, e.g., np.exp(99.06494938359764).

I'm solving this by using numexpr, which uses MKL functions (VML), and MKL seems to have the same implementation for all platforms.

Just to let you know that perhaps it would be nice if numpy used MKL for these functions too.

[tadeu]

Just a curiosity: in these examples, for tanh, the Windows version is more accurate, and for exp, the Linux version is more accurate.

[njsmith]

1 ulp differences are definitely expected and ordinary. Using MKL is not a general solution either; e.g., even for trivial floating point calculations like a + b + c you can get different results depending on platforms (Windows uses different rounding rules than Linux/MacOS in some cases) and compiler settings. If you really, really need floating point determinism then it's possible but it generally requires you that every piece of code you use is written with this in mind, and that you carefully control your entire technology stack including exactly which compilers and target platforms you care about. The only folks I know who routinely do this is are game developers who need networked engines to remain in lockstep and otherwise can meet those critera (ref1, ref2).

Usually the best (and certainly most common) solution is to write your code so that it doesn't care about 1 ulp differences.