Numerical precision differences in jitted vs. non-jitted code! about jax HOT 3 CLOSED

Hi - thanks for the question! This kind of difference is expected: in general, floating point operations are only accurate to within a particular precision that depends on the width of the float representation. When you compute the "same" result in two ways, the results will not in general be bitwise-equivalent. JIT-compilation replaces your original sequence of operations with a more efficient compiled kernel, and so in general you should not expect bitwise-equivalent outputs.

You can see the approximate expected precision using finfo:

In [1]: import numpy as np

In [2]: np.finfo(np.float64).eps
Out[2]: np.float64(2.220446049250313e-16)

The differences you're seeing are consistent with those expectations.

from jax.

Also, a side note: you might wonder why the compiler doesn't just simplify x + 0 * y to x: the reason for this is that for floating point math, these two expressions may return different results! For example, if you plug-in y = np.inf or y = np.nan, the first and second expression are not equivalent.

from jax.

I think we should probably close this as working-as-intended, and Jake's answer covers the reasoning well.

from jax.