Avoid reduction overhead in hash join probe kernels

[PointKernel]

Description

Follow-up to #24637. Match contexts and join retrieval pass a null total_matches pointer, but the runtime guard still gives their probe kernel the register and shared-memory footprint of the fused size reduction.

Select separate sum and no-sum kernel specializations in the host launcher. Match contexts, ordinary retrieval, and partitioned retrieval use a kernel without the sum accumulator or block reduction, while size queries retain the fused reduction and its memory savings. Both specializations keep the same probing and full-join bookkeeping logic.

Checklist

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

[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.

[PointKernel]

/ok to test

[PointKernel]

Before/after performance for match-count contexts and partitioned joins, using nvbench_compare.py. One row per benchmark configuration; means and standard deviations are pooled from the existing 500 samples per configuration, including variation between runs.

Ref: merged #24637 (d625effc8b). Cmp: 53bd4916d1; current head ee54fb66f8 has identical join source. RTX PRO 6000 Blackwell, CUDA 13.3.73, GCC 14.3, Release sm120; clocks unlocked, no foreign GPU compute overlap observed.

left_size/right_size are probe/build rows; Nullable=1 means nullable. Negative %Diff means faster. pr24637_probe is the custom follow-up benchmark; inner_join is the stock benchmark.

python python/scripts/nvbench_compare.py --no-color baseline.json candidate.json

[0] NVIDIA RTX PRO 6000 Blackwell Workstation Edition

Nullable	DataType	left_size	right_size	num_keys	multiplicity	selectivity	kind	mode	nulls	Ref Time	Ref Noise	Cmp Time	Cmp Noise	Diff	%Diff	Status
1	INT32	10000000	100000	1	1	0.1	inner	context	unequal	177.812 us	1.68%	160.400 us	2.45%	-17.413 us	-9.79%	🟢 FAST
1	INT32	10000000	100000	1	1	0.1	left	context	unequal	179.140 us	1.65%	160.686 us	1.49%	-18.454 us	-10.30%	🟢 FAST
0	INT32	10000000	100000	1	1	0.1	left	context	unequal	205.921 us	0.98%	190.597 us	3.94%	-15.324 us	-7.44%	🟢 FAST
0	INT32	10000000	100000	1	1	0.1	inner	context	unequal	205.773 us	2.18%	187.277 us	1.63%	-18.496 us	-8.99%	🟢 FAST
1	INT32	10000000	100000	1	1	0.1	inner	partitioned	unequal	578.406 us	2.60%	543.241 us	2.25%	-35.165 us	-6.08%	🟢 FAST
0	INT32	10000000	100000	1	1	0.1	inner	partitioned	unequal	642.552 us	1.20%	613.061 us	3.37%	-29.491 us	-4.59%	🟢 FAST
1	INT32	10000000	100000	1	1	0.1	left	partitioned	unequal	710.642 us	3.83%	673.815 us	0.89%	-36.827 us	-5.18%	🟢 FAST
0	INT32	10000000	100000	1	1	0.1	left	partitioned	unequal	770.522 us	5.59%	736.486 us	5.56%	-34.036 us	-4.42%	🔵 SAME
0	STRING	10000000	100000	1	1	0.1	left	context	unequal	482.793 us	0.62%	458.734 us	0.58%	-24.060 us	-4.98%	🟢 FAST
1	STRING	10000000	100000	1	1	0.1	inner	context	unequal	475.375 us	0.80%	453.749 us	0.82%	-21.627 us	-4.55%	🟢 FAST
1	STRING	10000000	100000	1	1	0.1	left	context	unequal	477.565 us	0.68%	453.496 us	0.80%	-24.070 us	-5.04%	🟢 FAST
0	STRING	10000000	100000	1	1	0.1	inner	context	unequal	479.975 us	0.52%	458.313 us	0.59%	-21.662 us	-4.51%	🟢 FAST
0	INT32	10000000	100000	1	1	0.9	left	context	unequal	281.123 us	1.22%	284.581 us	1.04%	3.458 us	1.23%	🔴 SLOW
1	INT32	10000000	100000	1	1	0.9	left	context	unequal	237.922 us	0.98%	240.304 us	2.86%	2.382 us	1.00%	🔴 SLOW
0	LIST	100000	10000	1	1	0.3	inner	context	unequal	101.342 us	2.05%	100.661 us	2.07%	-0.681 us	-0.67%	🔵 SAME
0	LIST	100000	10000	1	1	0.3	left	partitioned	unequal	224.868 us	1.81%	226.929 us	2.57%	2.061 us	0.92%	🔵 SAME
1	LIST	100000	10000	1	1	0.3	full	context	unequal	128.141 us	2.64%	130.185 us	4.34%	2.044 us	1.60%	🔵 SAME
1	LIST	257	1000	1	1	0.3	inner	context	equal	52.951 us	6.30%	52.734 us	6.77%	-0.218 us	-0.41%	🔵 SAME
1	LIST	257	1000	1	1	0.3	left	partitioned	equal	162.421 us	3.43%	163.681 us	3.78%	1.259 us	0.78%	🔵 SAME
0	STRUCT	100000	10000	1	1	0.3	inner	context	unequal	214.525 us	4.23%	214.441 us	2.56%	-0.084 us	-0.04%	🔵 SAME
0	STRUCT	100000	10000	1	1	0.3	left	partitioned	unequal	473.896 us	2.28%	488.446 us	8.45%	14.551 us	3.07%	🔴 SLOW
1	STRUCT	100000	10000	1	1	0.3	full	context	unequal	340.246 us	3.67%	343.640 us	13.12%	3.394 us	1.00%	🔵 SAME
1	STRUCT	257	1000	1	1	0.3	inner	context	equal	309.042 us	2.49%	311.036 us	3.61%	1.994 us	0.65%	🔵 SAME
1	STRUCT	257	1000	1	1	0.3	left	partitioned	equal	673.450 us	2.84%	668.010 us	5.63%	-5.440 us	-0.81%	🔵 SAME
0	INT32	10000000	100000	1	1	0.1	full	partitioned	unequal	862.674 us	1.73%	826.637 us	1.35%	-36.037 us	-4.18%	🟢 FAST
1	INT32	10000000	100000	1	1	0.1	full	partitioned	unequal	804.291 us	0.72%	774.667 us	2.49%	-29.625 us	-3.68%	🟢 FAST
1	STRING	10000000	100000	1	1	0.1	full	partitioned	unequal	1.435 ms	1.91%	1.391 ms	1.54%	-43.637 us	-3.04%	🟢 FAST
1	FLOAT64	100000	10000	3	1	0.3	full	partitioned	unequal	251.945 us	14.09%	238.303 us	3.80%	-13.642 us	-5.41%	🟢 FAST
1	LIST	100000	10000	1	1	0.3	full	partitioned	unequal	339.734 us	4.75%	336.357 us	2.78%	-3.377 us	-0.99%	🔵 SAME
1	STRUCT	100000	10000	1	1	0.3	full	partitioned	unequal	771.611 us	9.73%	757.947 us	2.64%	-13.665 us	-1.77%	🔵 SAME

inner_join

[0] NVIDIA RTX PRO 6000 Blackwell Workstation Edition

Nullable	NullEquality	DataType	num_keys	left_size	right_size	skip_large_sizes	mode	Ref Time	Ref Noise	Cmp Time	Cmp Noise	Diff	%Diff	Status
1	NULLS_UNEQUAL	INT32	3	10000000	100000	0	partitioned	1.021 ms	1.42%	950.435 us	2.47%	-70.469 us	-6.90%	🟢 FAST
0	NULLS_UNEQUAL	INT32	3	10000000	100000	0	partitioned	1.394 ms	2.14%	1.294 ms	1.27%	-100.247 us	-7.19%	🟢 FAST

[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: 05076d03-cc4a-4afe-b5d7-33f7352381ab

📥 Commits

Reviewing files that changed from the base of the PR and between 87d2d78 and ee54fb6.

📒 Files selected for processing: 1

• cpp/src/join/hash_join/hash_csr_kernels.cuh

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

• Performance
  • Improved hash join execution efficiency when match totals are not requested.
  • Reduced unnecessary processing and resource usage for applicable join operations.
  • Preserved match-count calculation behavior when totals are requested.

The probe-count kernel now uses a ComputeSum template parameter to control total-match accumulation and reduction. The launcher selects the specialization based on whether total_matches is non-null.

Changes: Hash CSR probe-count specialization

Layer / File(s)	Summary
Kernel aggregation specialization cpp/src/join/hash_join/hash_csr_kernels.cuh	The kernel compiles match accumulation and block reduction only when ComputeSum is enabled. The launcher enables this specialization when total_matches is non-null.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Refactor

Suggested reviewers: scanhex12

Merge Risk: ⚪ Minimal · up to ee54f

This change specializes the hash join probe-count kernel so launches that do not request totals skip the reduction. No actionable merge-blocking risk was found.

Pre-merge checks: 5

Check name	Status	Explanation
Title check		The title clearly identifies the main change: removing reduction overhead from hash join probe kernels.
Description check		The description directly explains the kernel specialization changes, affected paths, preserved behavior, and reported validation results.
Docstring Coverage		No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
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.

Full details: Docstring Coverage

Explanation

No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0 files. (1 skipped: 1 unsupported.)

✨ 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.}