Execute tensor product functions with RowFn

[connortsui20]

Rationale for this change

Moves inner product and cosine similarity to the shared row executor. This removes duplicated batch machinery and prepares constant-dependent work once.

• Progress towards: Epic: Row-oriented scalar functions #9128

What changes are included in this PR?

Cosine similarity computes each batch-constant norm once. Inner product and cosine keep their Normalized reductions and authoritative stored-norm semantics. Tests cover dense, constant, nullable, vector, tensor, and normalized inputs.

Rust 1.97.1 one-CGU fat-LTO measurements show the width-256 dense inner product within 4.1% of develop; most other product paths improve by 14–92%, with the largest gains on constant operands and narrow vectors.

What APIs are changed? Are there any user-facing changes?

There are no public API or behavior changes. Both functions implement RowFn and receive the standard scalar-function vtable automatically.

[codspeed-hq]

Merging this PR will regress 1 benchmark

⚠️ 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.

⚡ 11 improved benchmarks
❌ 1 regressed benchmark
✅ 1986 untouched benchmarks
🆕 15 new benchmarks
⏩ 529 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	words_gather_scalar[65536]	8.2 µs	9.4 µs	-12.64%
⚡	Simulation	column_x_constant[2]	503 µs	150.9 µs	×3.3
⚡	Simulation	column_x_constant[32]	457.5 µs	155.7 µs	×2.9
⚡	Simulation	column_x_column[2]	230.9 µs	102 µs	×2.3
⚡	Simulation	non_nullable[2]	645.4 µs	294 µs	×2.2
⚡	Simulation	nullable[2]	643 µs	314.9 µs	×2
⚡	Simulation	column_x_constant[256]	723.1 µs	375 µs	+92.84%
⚡	Simulation	column_x_column[32]	144.8 µs	83.8 µs	+72.65%
⚡	Simulation	column_x_column[256]	139.6 µs	82.6 µs	+68.89%
⚡	Simulation	column_x_extension_constant[2]	300.1 µs	191.9 µs	+56.4%
⚡	Simulation	column_x_extension_constant[256]	657.4 µs	441 µs	+49.08%
⚡	Simulation	column_x_extension_constant[32]	253.1 µs	191.6 µs	+32.1%
🆕	Simulation	column_x_nullable_constant[2]	N/A	168.9 µs	N/A
🆕	Simulation	column_x_nullable_constant[256]	N/A	394.2 µs	N/A
🆕	Simulation	column_x_nullable_constant[32]	N/A	174.1 µs	N/A
🆕	Simulation	constant_x_column[2]	N/A	152.3 µs	N/A
🆕	Simulation	constant_x_column[256]	N/A	375.2 µs	N/A
🆕	Simulation	constant_x_column[32]	N/A	156.3 µs	N/A
🆕	Simulation	column_x_constant[2]	N/A	335.1 µs	N/A
🆕	Simulation	column_x_constant[256]	N/A	461.2 µs	N/A
...	...	...	...	...	...

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-tensor-products (5114763) with ct/row-fn-tensor-l2 (f46a367)}

Footnotes

1. 529 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. ↩

[connortsui20]

Here are the local tensor-product benchmark results from the final Rust 1.97.1 run.

The comparison used a develop baseline with benchmark-only backfills (1edd0fca) and the full RowFn stack (8ca9fc73). Later stack layers do not change tensor-product source, but unrelated code-placement effects can still affect these binaries.

The machine was an AMD Ryzen 9 7950X running Linux. The build used rustc 1.97.1, LLVM 22.1.6, one CGU, fat LTO, and -C target-cpu=native. Each target used two warm runs and seven alternating measured pairs. Each process used 100 samples with a 0.25–0.5 second measurement window on CPU 4.

Negative changes are faster. The values are medians across the seven paired runs.

Width-256 dense inner products remain within 4.11%. Narrow, constant, and nullable-constant product cases improve by 14.68–91.92%. Every cosine-similarity case improves, while the Normalized controls remain within 1.34%.

Inner product: 15 cases

Benchmark	Baseline	RowFn stack	Change
non_nullable/256	5.119 µs	5.339 µs	+4.11%
nullable/256	5.899 µs	5.689 µs	-3.73%
non_nullable/32	4.019 µs	3.429 µs	-14.68%
nullable/32	4.809 µs	3.779 µs	-21.63%
constant_x_column/256	96.580 µs	30.980 µs	-67.83%
column_x_constant/256	96.450 µs	27.610 µs	-71.56%
column_x_nullable_constant/256	97.440 µs	24.460 µs	-74.92%
nullable/2	31.490 µs	7.659 µs	-75.71%
non_nullable/2	31.620 µs	7.329 µs	-76.79%
column_x_nullable_constant/2	112.600 µs	11.750 µs	-89.59%
column_x_constant/2	110.500 µs	11.440 µs	-89.66%
constant_x_column/2	117.200 µs	11.540 µs	-90.20%
constant_x_column/32	88.670 µs	7.869 µs	-91.14%
column_x_constant/32	86.890 µs	7.349 µs	-91.55%
column_x_nullable_constant/32	89.320 µs	7.239 µs	-91.92%

Cosine similarity: 15 cases

Benchmark	Baseline	RowFn stack	Change
column_x_extension_constant/256	34.630 µs	25.460 µs	-26.48%
column_x_extension_constant/32	8.759 µs	6.199 µs	-29.31%
constant_x_column/256	44.820 µs	28.290 µs	-36.88%
column_x_constant/256	44.940 µs	24.690 µs	-45.04%
column_x_column/32	4.569 µs	2.289 µs	-49.68%
column_x_column/256	5.359 µs	2.279 µs	-57.47%
column_x_extension_constant/2	16.460 µs	5.999 µs	-63.86%
column_x_nullable_constant/256	68.540 µs	21.860 µs	-68.02%
column_x_column/2	17.850 µs	4.329 µs	-75.78%
constant_x_column/32	28.570 µs	6.349 µs	-77.62%
column_x_constant/32	28.770 µs	5.849 µs	-79.41%
column_x_constant/2	36.100 µs	5.679 µs	-84.27%
constant_x_column/2	37.210 µs	5.849 µs	-84.33%
column_x_nullable_constant/32	53.640 µs	5.809 µs	-88.97%
column_x_nullable_constant/2	65.410 µs	5.989 µs	-90.89%

Normalized-array controls: 6 cases

Benchmark	Baseline	RowFn stack	Change
nullable/32	8.179 µs	8.279 µs	+1.34%
nullable/256	8.089 µs	8.119 µs	+0.87%
non_nullable/2	17.260 µs	17.340 µs	+0.58%
nullable/2	18.010 µs	18.160 µs	+0.50%
non_nullable/32	7.529 µs	7.579 µs	+0.40%
non_nullable/256	7.439 µs	7.409 µs	-0.27%