Poisson random number generation: under the hood

How does numpy.random.poisson actually generate random samples from a poisson distribution? What technique does it use? I'm optimizing my code right now and it's the largest bottleneck. If numpy isn't using the fastest possible sampling techniques, it might be worth coding poisson random number generation myself.

I can't seem to find any information regarding this. The documentation doesn't shed any light on how the numbers are actually getting generated.

Thanks!