Repository navigation
performance improvements to np.full and np.ones - #7682
Merged
Merged
Conversation
rishi-kulkarni
requested review from
esc,
sklam and
stuartarchibald
as code owners
December 21, 2021 22:31
Member
|
Many thanks for the PR! I'll take a look at this on Wednesday when I'm back at work. |
Member
|
It will probably be of relevance to take the following into account when reviewing this issue: |
Member
|
@rishi-kulkarni do you have any benchmarks to illustrate the performance gains here? |
Contributor
Author
|
Yep, here it is. This is on current master branch: import numpy as np
from numba import jit
@jit
def nb_full(shape, fill_value, dtype):
ans = np.empty(shape, dtype)
fl = ans.flat
for idx, v in enumerate(fl):
fl[idx] = fill_value
return ans
@jit
def f4(shape):
return nb_full(shape, np.nan, np.float64)
@jit
def f1(shape):
return np.full(shape, np.nan, np.float64)
shape = (128000, 1000)
f1(shape)
f4(shape)
%timeit np.full(shape, np.nan, np.float64)
%timeit f1(shape)
%timeit f4(shape)
# output
116 ms ± 4.79 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) # numpy
206 ms ± 1.28 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) # current implementation
110 ms ± 2.27 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) # array.flat implementation On this branch: 116 ms ± 5.48 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) # numpy
114 ms ± 7.35 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) # proposed implementation
114 ms ± 2.79 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) # array.flat implementation |
gmarkall
previously approved these changes
Jan 20, 2022
gmarkall
left a comment
gmarkall
left a comment
Member
There was a problem hiding this comment.
This looks good to me. Whilst fixing #7753 might also fix this issue, this change doesn't appear to add significant complexity and fixes #7665 right now, so I think it would be OK to merge.
Note that we still have a big difference between Numba and NumPy for zeros, because NumPy uses calloc:
import numpy as np
from numba import jit
@jit
def f0(shape):
return np.zeros(shape, dtype=np.float64)
shape = (128000, 1000)
f0(shape)
%timeit f0(shape)
%timeit np.zeros(shape, dtype=np.float64)produces:
401 ms ± 119 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
8.33 µs ± 40.6 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
(and is unaffected by this branch).
stuartarchibald
approved these changes
Jan 21, 2022
stuartarchibald
left a comment
stuartarchibald
left a comment
Contributor
There was a problem hiding this comment.
Thanks for the patch and fixes.
stuartarchibald
added
5 - Ready to merge
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #7665. np.full and np.ones use the slow np.ndindex iterator, which results in both functions being about twice as slow as their numpy counterparts. Iterating over the flat array instead makes the functions as fast as their numpy versions.