Argsort performance improvement #1859
Merged
Conversation
This file contains 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
Instead of implementing argsort as sort over structures (index, value), with subsequent projection to index, it is now implemented as sort over linear indices themselves, with dereferencing comparator, and subsequent mapping from linear index to row-wise index. On Iris Xe, argsort call took 215 ms to argsort 5670000 elements of type int32, and it now takes 117 ms. The new implementation also has smaller temporary allocation footprint. Previously, it would allocate 2*(sizeof(ValueT) + sizeof(IndexT)), now it only allocates sizeof(IndexT) for storing linear indices.
Eliminate use of temporary allocation altogether, cutting argsort execution time from 116 ms to 110 ms for 5670000 element array of type int32_t.
|
Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞 |
|
Array API standard conformance tests for dpctl=0.19.0dev0=py310hdf72452_118 ran successfully. |
|
Array API standard conformance tests for dpctl=0.19.0dev0=py310hdf72452_119 ran successfully. |
|
Drop in coverage must be related to some changes in coverall analysis code. PR did not make changes to files coverall reports have reduced coverage. |
ndgrigorian
approved these changes
Oct 9, 2024
There was a problem hiding this comment.
This LGTM, the performance improvement is very nice. Thank you @oleksandr-pavlyk !
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
This change modifies implementation of
tensor.argsortfunctions making is about 2x faster.Instead of implementing argsort as sort over structures (index, value),
with subsequent projection to index, it is now implemented as sort
over linear indices themselves, with dereferencing comparator, and
subsequent mapping from linear index to row-wise index.
On Iris Xe,
tensor.argsortcall took 215 ms to find sorting permutation forvector of 5670000 elements of type int32, and it now takes 106 ms.
The new implementation no longer makes temporary allocations for
storing indices. Previously, it would allocate
2*(sizeof(ValueT) + sizeof(IndexT)), now it just uses the outputallocation.
Tests exist.