Absolutely. Numpy's performance model is so brittle that one needs to be very selective when they cherry-pick their microbenchmarks. My rule of thumb is that i

Absolutely. Numpy's performance model is so brittle that one needs to be very selective when they cherry-pick their microbenchmarks. My rule of thumb is that if your cherrypicked microbenchmark involves one specialized Numpy function call, it'll be fast, but if there is two separate specialized Numpy function calls it'll be slow.