Get rid of np.exp(...) to generate Bluestein Fourier Matrix about fastmat HOT 5 CLOSED

from fastmat.

You're right, I'm referring to this part. Since this is called only once at this part of the code, you can take it as an annotation in general. ;-)

from fastmat.

I can confirm this is sort of performance-critical for larger orders. Assume you have an Fourier-based interpolator, this generation takes a couple of hundred ms for orders around ~1e6, which is not THAT large (but also not that small). I once observed bad performance when generating interpolator objects for varying data sizes but didn't attribute it to the exponential generation. If it is verified that that's the cause we should address it.

from fastmat.

I made a short example, where you can see that there are faster solutions than np.exp(). Allthough I don't know exactly what this numexpr module does.
Even a straight forward solution with sine and cosine is faster than np.exp().

Example:

x = np.random.randn(1000000,1)1j
%timeit np.exp(x)
70.4 ms ± 450 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
%timeit numexpr.evaluate('exp(x)')
16 ms ± 55 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
x = np.random.randn(1000000,1)
%timeit np.cos(x)+1jnp.sin(x)
46.6 ms ± 130 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

from fastmat.

Modifications were introduced in v0.2 by d22813f. Closing.

from fastmat.