Modify reprog_device::extract to return groups in a single pass #8460
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
davidwendt
added
2 - In Progress
Codecov Report
@@ Coverage Diff @@
## branch-21.08 #8460 +/- ##
===============================================
Coverage ? 82.95%
===============================================
Files ? 109
Lines ? 18226
Branches ? 0
===============================================
Hits ? 15120
Misses ? 3106
Partials ? 0 Continue to review full report at Codecov.
|
davidwendt
changed the title
davidwendt
added
3 - Ready for Review
vuule
requested changes
Jun 15, 2021
vuule
approved these changes
Jun 16, 2021
robertmaynard
approved these changes
Jun 17, 2021
cwharris
approved these changes
Jun 18, 2021
|
@gpucibot merge |
This was referenced Jun 21, 2021
rapids-bot bot
pushed a commit
that referenced
this pull request
Jun 25, 2021
) Closes #8569 This essentially undoes the performance improvement made in #8460 since the logic mishandles a greedy quantifier pattern when it occurs inside an extract group. The internal regex logic is only able to track a single extract group when such a quantifier is specified. This PR does improve the interface for the internal extract call and adds some gtests for this issue. Authors: - David Wendt (https://github.com/davidwendt) Approvers: - Ram (Ramakrishna Prabhu) (https://github.com/rgsl888prabhu) - MithunR (https://github.com/mythrocks) URL: #8575
rapids-bot bot
pushed a commit
that referenced
this pull request
Oct 13, 2021
) This is a less ambitious version of #8460 which had to be reverted in #8575 because it did not work with greedy quantifiers. The change here involves calling the underlying `reprog_device::extract` to retrieve each group result within a single kernel rather than launching a kernel for each group. The output is placed contiguously in a 2d span (wrapped uvector) and a permutation iterator is used to build the output columns (one column per group). Like it's predecessor, the performance improvement is mostly when specifying more than 1 group in the regex pattern. The benchmark results showed no change for single groups but was 2x faster for multiple groups over long (8K) strings and up to 4x faster for multiple groups over many (16M) strings. The benchmark test for extract was also updated to better report the number of groups being used when measuring results. Authors: - David Wendt (https://github.com/davidwendt) Approvers: - Mark Harris (https://github.com/harrism) - Nghia Truong (https://github.com/ttnghia) URL: #9358
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 PR modifies the internal regex
reprog_device::extractfunction to return all matching groups in a single call. Previously, retrieving each group range required individual calls to thisextractfunction resulted in re-matching the entire given pattern for each group. The code logic would identify each group but only return the range for the specified group.The code change here passes a pre-allocated global memory array to capture each group range in a single pass. The extract is an all-or-nothing process. In fact, a
findfunction must first be executed to retrieve the bounds of the given pattern. So if any of the groups are missing or do not match, no groups are returned for that row. Retrieving the last group would always require processing the previous groups and the code logic now records those positions in the global memory array. The memory array can then be used directly to build the output columns.This simplifies the code around extract and also improves performance especially for long strings or patterns with many groups. For small strings and a small number of groups, the gbenchmark showed equivalent performance to the previous implementation. For larger strings and more groups, the gbenchmark showed a 2-3x improvement.