[davidwendt]

Description

Improves the performance of nvtext::jaccard_index by replacing the segmented-sort based counting of unique and intersecting substring hashes with cuco static_set_ref sets in shared memory.

The substring hashes for each row are still computed once. The hash kernel now compacts the hashes using a warp ballot instead of a serial copy by lane 0. Each row is then counted by one of two kernels:

• Rows whose sets fit in shared memory (up to 384 substrings) are processed by a warp per row. Each warp inserts the first column's hashes into one set and the second column's hashes into another set. Each successful insert into the second set is checked against the first set to count the intersection.
• Larger rows are partitioned by the high bits of their hashes in a single scatter step, and each (row, partition) is processed by a block using the same set logic. The counts from all of a row's partitions are added together. If a partition's set fills up, the block reprocesses that partition in smaller sub-partitions.

This removes the CUB segmented sort, which also removes the workaround for sorting more than 2B elements in sections.

A hash value of 0 is now stored as 1, since 0 is used as the empty-slot sentinel. Previously, substrings that hashed to 0 were dropped and their slot in the row was left uninitialized.

The benchmark is updated so that each pair of rows shares half of its text, so the intersection counts are no longer always 0. It also adds row widths of 8192 and 32768 bytes to measure long rows, and skips configurations larger than 1GB. A new gtest compares the results for rows of up to 100K characters against a host reference implementation.

Benchmark results on an RTX A6000 (substring_width=5):

max_width	num_rows	before (ms)	after (ms)	speedup
128	262144	3.23	2.30	1.4x
512	262144	100.79	8.36	12.1x
2048	262144	127.56	31.95	4.0x
8192	131072	116.40	72.29	1.6x
32768	32768	166.11	75.07	2.2x

The large improvement at 512 is because the CUB segmented sort handles segments of a few hundred items poorly. Peak memory is about half of before when all rows fit in a warp's sets, and about the same otherwise.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

🤖 Generated with Claude Code

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[davidwendt]

/ok to test

[davidwendt]

@coderabbitai full review

[coderabbitai]

✅ Action performed

Full review finished.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Repository: NVIDIA/cudf/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 8c943d37-6a8e-40d4-ade9-2422f6bba721

📥 Commits

Reviewing files that changed from the base of the PR and between 1b906f3 and b89727d.

📒 Files selected for processing: 3

• cpp/benchmarks/text/jaccard.cpp
• cpp/src/text/jaccard.cu
• cpp/tests/text/jaccard_tests.cpp

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

📝 Summary

Summary by CodeRabbit

• Improvements
  • Updated Jaccard index calculations to support rows containing very long strings, including empty strings, while preserving the existing similarity ratio and returning zero when the union is empty. The revised approach also handles rows with different amounts of distinct content. These changes are covered by additional comparisons against a reference implementation across a range of input lengths.

The Jaccard implementation now counts substring hashes with sets instead of a sorted-hash pipeline. It uses warp-level counting for smaller rows and partitioned counting for larger rows. Tests compare results with a host reference, and the benchmark adds shared-substring inputs and wider benchmark sizes.

Changes: Jaccard Counting

Layer / File(s)	Summary
Hash representation and compaction cpp/src/text/jaccard.cu	The code adds set hash types and capacities, and compacts valid substring hashes into unsorted row outputs.
Warp-level row counting cpp/src/text/jaccard.cu	Rows within the warp-set capacity use separate sets to count unique hashes and intersections. Warp reductions write counts to row outputs.
Partitioned counting for larger rows cpp/src/text/jaccard.cu	Larger rows are partitioned and counted with block-level sets. The counting path retries with finer sub-partitions when a set fills.
Jaccard integration and coverage cpp/src/text/jaccard.cu, cpp/tests/text/jaccard_tests.cpp, cpp/benchmarks/text/jaccard.cpp	jaccard_index connects hashing and counting to the result calculation. Tests compare results with an ASCII reference. Benchmarks use paired inputs with a shared string and include larger width values.

Priority: ➖ Normal

Estimated code review effort: 4 (Complex) | ~45 minutes

Change: Refactor

Suggested reviewers: vyasr, kylefromnvidia, bdice

Merge Risk: ⚪ Minimal · up to b8972

The faster set-based Jaccard counting keeps 64-bit offsets for large inputs and computes the same ratio as before. No concrete correctness or stability defect was found, so the change appears ready to merge after normal CI.

Pre-merge checks: 4 · 1

Failed checks: 1 warning

Check name	Status	Explanation	Resolution
Docstring Coverage		Docstring coverage is 72.22% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 18 functions across 3 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

Passed checks: 4 passed

Check name	Status	Explanation
Title check		The title clearly and concisely describes the main change: improving nvtext::jaccard_index performance with cuco shared-memory sets.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.
Description check		The pull request description directly explains the Jaccard implementation changes, shared-memory set approach, benchmark updates, tests, and reported performance results.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

/ok to test

[davidwendt]

/ok to test