feat(fastlanes): compute kernels for BitPackedV2

[mhk197]

Summary

Stacked on #9750. Ports the BitPacked compute kernels to BitPackedV2 so the new encoding no longer falls back to canonicalisation for slice, take, filter, compare, between, cast and is_constant.

• compute/: slice (reduce and execute), take, filter, fused compare against a constant, between, streaming predicate, cast (nullability-only reduce and widening execute), is_constant.
• vtable: reduce_parent rules for slice and cast, kernel registration in initialize, and the is_constant aggregate kernel.
• unpack_iter gains back the chunk-walking helpers only the kernels use (for_each_packed_chunk, for_each_unpacked_chunk, decode_map_into) and bitpack_decompress the mapped unpack used by cast.

Each kernel is the v1 kernel with the chunk lookup changed to go through the per-chunk width table; after renaming, the non-test kernel code differs from v1 by 2 to 19 lines per file. The fused compare additionally handles zero-width chunks, which v1 could only see as a whole-array degenerate case.

Tests

The v1 kernel tests are carried over, including the fused-compare sweeps that assert the streaming kernel engages rather than falling back, plus the slice tests that go through reduce_parent. The per-chunk behavioural suite from #9750 now exercises the kernels instead of the fallbacks.

Validation

• cargo nextest run -p vortex-fastlanes (463 tests)
• cargo clippy -p vortex-fastlanes --all-targets --all-features, cargo +nightly fmt --all

[codspeed-hq]

Merging this PR will degrade performance by 14.48%

⚠️ 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.

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 10 improved benchmarks
❌ 60 regressed benchmarks
✅ 2184 untouched benchmarks
⏩ 284 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	slice_tight_loop_vortex[65536]	59.3 µs	121.8 µs	-51.32%
❌	Simulation	slice_empty_tight_loop_vortex	41.2 µs	71.3 µs	-42.26%
❌	Simulation	bitwise_and_vortex_buffer[128]	3.9 µs	6.4 µs	-39.33%
❌	Simulation	bitwise_or_vortex_buffer[128]	3.9 µs	6.4 µs	-38.8%
❌	Simulation	take_fsl_f16_force_per_index[2048, 10]	361.1 µs	581 µs	-37.84%
❌	Simulation	take_fsl_f16_force_per_index[1024, 10]	209.4 µs	320.2 µs	-34.59%
❌	Simulation	bitwise_not_vortex_buffer[128]	4.7 µs	6.8 µs	-31.27%
❌	Simulation	random_i8[0.5]	67.3 µs	96.5 µs	-30.29%
❌	Simulation	cached_indices_i32[0.01]	36.2 µs	52 µs	-30.23%
❌	Simulation	take_fsl_f16_force_per_index[512, 10]	134.1 µs	189.1 µs	-29.11%
❌	Simulation	in_place_nullable[16384]	36.2 µs	50.9 µs	-28.89%
❌	Simulation	bitand_owned_lhs_vortex_buffer[128]	4.6 µs	6 µs	-23.41%
❌	Simulation	take_fsl_f16_force_per_index[256, 10]	94.5 µs	122.7 µs	-23.03%
❌	Simulation	bitand_owned_lhs_vortex_buffer[1024]	4.8 µs	6.2 µs	-22.73%
❌	Simulation	bitand_owned_lhs_vortex_buffer[2048]	5 µs	6.4 µs	-21.89%
❌	Simulation	new_raw_prim_test_between[i32, 2048]	61.3 µs	78.1 µs	-21.45%
❌	Simulation	chunked_constant_i32_append_to_builder[(10, 100)]	62.4 µs	79.2 µs	-21.2%
❌	Simulation	slice_vortex_buffer	6.6 µs	8.2 µs	-19.81%
❌	Simulation	bitwise_and_vortex_buffer[1024]	4.2 µs	5.2 µs	-18.98%
❌	Simulation	bitwise_or_vortex_buffer[2048]	4.6 µs	5.6 µs	-17.81%
...	...	...	...	...	...

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 mk/bitpacked-v2-kernels (34bd0f3) with mk/bitpacked-v2 (7de5950)}

Footnotes

1. 284 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. ↩

[mhk197]

Superseded: the stack now follows the one-array model (one in-memory BitPacked, two wire formats) in #9750, so the kernels are edited in place there instead of ported to a second encoding.