MAINT: Improve speed of ufunc kwargs parsing. #11333
Conversation
Using the realization from inspecting similar CPython code that as we intern string names anyway, comparing keys with possible names by pointer will generally just work. Reduces the overhead as follows: ``` import numpy as np a = np.array(1) b = np.empty_like(a) %timeit np.positive(a) # 352->348 ns %timeit np.positive(a, subok=True) # 606->501 ns %timeit np.positive(a, where=True) # 586->503 ns %timeit np.positive(a, where=True, subok=True) # 695->531 ns %timeit np.positive(a, b) # 354->352 ns %timeit np.positive(a, out=b) # 557->480 ns %timeit np.positive(a, out=b, subok=True) # 668->506 ns %timeit np.positive(a, out=b, where=True, subok=True) # 752->536 ns ```
|
p.s. Quite obviously, more to be gained by doing this elsewhere, and by looking for |
| for (kw_id = 0; kw_id < ufunc_badkey; kw_id++) { | ||
| int cmp = PyObject_RichCompareBool(key, ufunc_kwnames[kw_id], Py_EQ); | ||
| if (cmp > 0) { | ||
| goto kw_found; |
There was a problem hiding this comment.
Instead of using a goto, just make the above a helper function returning kw_id
There was a problem hiding this comment.
I followed CPython's ceval quite exactly! But with a helper function, this may become useful for GenericReduction as well.
There was a problem hiding this comment.
Can you link to the code you're imitating?
There was a problem hiding this comment.
There was a problem hiding this comment.
p.s. Do think the helper function is a good idea, but was waiting with making changes as I hoped someone would have a better idea of how to handle those name arrays and the enum -- I like the names but not that multiple lists have to be maintained separately. Can one iterate over an enum?
|
I'm curious how much the |
|
The one issue would be that then the slow path becomes a whole set of |
|
I think the real solution here would be to mimic |
|
Closing in favour of #11351 |
This uses the realization from inspecting similar CPython code that as we intern string names anyway, comparing keys with possible names by pointer will generally just work, and is of course much faster than first converting a unicode object to
char*and then usingstrcmp.@eric-wieser - my C remains limited: is there a cleverer way than the
enumI used, ideally one that allows on to type things just once (there are now five places one has to keep up to date, three for the interns, one enum, and one array of the interned strings; this seems too much!)Reduces the overhead as follows: