[HN Gopher] When NumPy is too slow
___________________________________________________________________
When NumPy is too slow
Author : ingve
Score : 15 points
Date : 2023-06-28 19:35 UTC (3 hours ago)
(HTM) web link (pythonspeed.com)
(TXT) w3m dump (pythonspeed.com)
| credit_guy wrote:
| I don't get it. If code written in numpy is slow because the
| algorithm is bad, and you need to rewrite it, we can't say "numpy
| is too slow". A good chunk of the blogpost talks about this
| particular case, which is a non-case.
|
| The rest is some vague suggestions to use numba, or jax. But if
| your code is vectorizable in my experience you don't get any
| benefit from numba and jax. With the exception that jax is by
| default single-precision so you get some speed up from that. If
| you make it double precision, you get back to numpy speed. I
| suppose the author's conclusion is not all that different,
| otherwise they'd put a code snippet and show some numbers.
| tnecniv wrote:
| I have definitely gotten significant speed ups using Jax. I try
| to vectorize my code as much as possible but that's not always
| feasible for every aspect of an algorithm. Even for the
| vectorized operations, the JIT compilation gets me a speed up
| when used correctly (I.e. structuring code to minimize
| recompiling). Jax also has the added bonus of trivially running
| your algorithm on the GPU.
___________________________________________________________________
(page generated 2023-06-28 23:02 UTC)