Use RowFn for primitive numeric comparisons

[connortsui20]

Progress towards #9128

Supersedes #9346. Depends on #9548.

Moves primitive numeric comparisons onto the associated-output RowKernel API. The implementation covers all 11 numeric physical types, all six operators, and array/array, array/constant, constant/array, and constant/constant inputs. Float equality remains bitwise and ordering matches total_cmp.

On x86-64, runtime dispatch selects AVX2 or AVX-512 and writes comparison masks directly into packed Boolean storage. Other targets retain the columnar implementation.

CodSpeed reports a -7.59% aggregate result with 34 regressions, 13 improvements, and 1,934 unchanged benchmarks. These are AVX2 simulation results, so the local AVX-512 results do not explain them away. In same-artifact comparisons against the retained columnar path, RowFn ranges from 7.75% faster to 12.27% slower with AVX2 and from 5.32% faster to 8.87% slower with AVX-512.

Local AVX2 and AVX-512 wall-time results

These results compare the RowFn and retained columnar paths in the same candidate artifact. They isolate the implementations within this PR rather than comparing develop against the full stack.

The benchmark processes 65,536 rows on a Ryzen 9 7950X pinned to CPU 20. Each entry is the median of five independent runs with 501 samples per run. AVX2 uses -C target-feature=+avx2. AVX-512 uses -C target-cpu=native.

Workload	AVX2 columnar	AVX2 RowFn	Delta	AVX-512 columnar	AVX-512 RowFn	Delta
f64 GTE, array/array	12.770 µs	11.780 µs	-7.75%	11.100 µs	10.510 µs	-5.32%
i64 GTE, array/array	10.170 µs	10.330 µs	+1.57%	9.780 µs	9.970 µs	+1.94%
i64 GTE, nullable array/array	10.560 µs	11.400 µs	+7.95%	10.260 µs	11.170 µs	+8.87%
i64 GTE, array/constant	4.400 µs	4.940 µs	+12.27%	4.350 µs	4.520 µs	+3.91%
i64 GTE, constant/array	4.510 µs	4.800 µs	+6.43%	4.580 µs	4.700 µs	+2.62%
u64 GTE, array/array	10.270 µs	10.330 µs	+0.58%	10.211 µs	10.380 µs	+1.66%

The native assembly uses mask comparisons and direct kmov stores without expanding results into Boolean lanes. The AVX2 loop uses YMM comparisons and vmovmskps, with constant broadcasts outside the inner loop. Neither loop contains calls.

CodSpeed simulation results

The Performance Analysis job reports:

• Aggregate result: -7.59%
• 34 regressed
• 13 improved
• 1,934 unchanged
• 47 new
• 54 skipped

Representative regressions include:

Benchmark	Base	Candidate	CodSpeed efficiency
eq_i64_constant	99.4 µs	120.5 µs	-17.55%
lt_i64_nullable	135.3 µs	170.6 µs	-20.69%
or_chain[16]	1.4 ms	2.1 ms	-31.45%

The simulation build selects the AVX2 implementation through the test harness's compile-time gating. There is no evidence that these regressions come from CodSpeed failing to simulate AVX-512. The separate AVX2, AVX-512, and NEON wall-time jobs complete successfully, but these benchmarks are new and therefore lack a base comparison in the aggregate report.

Verification

• 80 binary comparison tests
• Direct AVX2 and AVX-512 differential coverage for every numeric type, operator, input shape, and bitmap boundary
• Full vortex-array --all-features suite: 3,479 passed, one skipped
• cargo clippy -p vortex-array --all-targets --all-features -- -D warnings
• cargo test --doc -p vortex-array --all-features: 74 passed, 20 ignored
• cargo doc -p vortex-array --all-features --no-deps
• cargo +nightly fmt --all -- --check

[codspeed-hq]

Merging this PR will degrade performance by 7.19%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 14 improved benchmarks
❌ 34 regressed benchmarks
✅ 1933 untouched benchmarks
🆕 47 new benchmarks
⏩ 54 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	or_chain[16]	1.4 ms	2.1 ms	-31.45%
❌	Simulation	or_chain[1024]	1.5 ms	2.1 ms	-31.13%
❌	Simulation	float_gt[16]	292.3 µs	383.4 µs	-23.78%
❌	Simulation	case_when_nary_10_conditions[1000]	381.6 µs	498.2 µs	-23.39%
❌	Simulation	float_gt[1024]	322.7 µs	416.5 µs	-22.53%
❌	Simulation	int_gt[16]	113.4 µs	146.1 µs	-22.37%
❌	Simulation	int_gt[1024]	120.2 µs	151.8 µs	-20.78%
❌	Simulation	lt_i64_nullable	135.3 µs	170.6 µs	-20.69%
❌	Simulation	or_chain[8192]	2.7 ms	3.3 ms	-18.17%
❌	Simulation	is_not_null_pred[16]	232.5 µs	282.3 µs	-17.62%
❌	Simulation	eq_i64_constant	99.4 µs	120.5 µs	-17.55%
❌	Simulation	is_not_null_pred[1024]	249 µs	300.6 µs	-17.16%
❌	Simulation	case_when_nary_early_dominant[1000]	217.3 µs	258.2 µs	-15.87%
❌	Simulation	case_when_nary_equality_lookup[1000]	326.7 µs	387.2 µs	-15.64%
❌	Simulation	case_when_nary_3_conditions[1000]	221 µs	260.4 µs	-15.1%
❌	Simulation	int_gt[8192]	171.4 µs	201.7 µs	-15.02%
❌	Simulation	compare_int_nullable	207.5 µs	243.8 µs	-14.91%
❌	Simulation	baseline_eq[4, 1024]	98.1 µs	115.3 µs	-14.88%
❌	Simulation	baseline_eq[16, 1024]	99.5 µs	116.7 µs	-14.74%
❌	Simulation	float_gt[8192]	580.8 µs	678 µs	-14.33%
...	...	...	...	...	...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ct/row-fn-binary-numeric-comparison (e4d3617) with develop (332ae0b)²}

Footnotes

1. 54 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. No successful run was found on ct/row-fn-associated-output (4aa169c) during the generation of this report, so develop (332ae0b) was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩