[joseph-isaacs]

Summary

Part of the fearless_simd migration, following #10327. Bitmap rank-select's 64-byte chunk scan and in-chunk select had separate kernels per architecture: NEON vcnt intrinsics, AVX-512 vpopcntq intrinsics and a scalar fallback, all behind cfg(target_arch) and CpuKernel dispatch. This replaces them with one portable #[simd] kernel each, used on every target, including Miri.

Changes

• vortex-buffer/src/bit/select.rs:
  • scan_chunks popcounts each 64-byte chunk as u64x8 lanes.
  • select_in_chunk uses the same lane popcounts to find the word.
  • Both dispatch via Level::new().
• Removed:
  • scan_chunks_neon, scan_chunks_avx512_vpopcnt and scan_chunks_scalar;
  • select_in_chunk_vbmi2 and select_in_chunk_scalar;
  • their CpuKernels and every cfg(target_arch).
• Kept: the BMI2 pdep/tzcnt in-word select. It is a scalar bit-deposit instruction with no fearless_simd equivalent, so its one x86 cfg remains.
• Adds a select_last_vortex_buffer benchmark, which selects the last set bit in 512-, 65,536- and 1,048,576-bit buffers.

Why u64x8 lanes

The earlier revision summed bytes per 16-byte vector. That was faster on ARM64, but on AVX2 it took 2.1–2.5× as long as develop. Lane popcounts avoid that:

• x86 AVX2: matches develop (measured below).
• x86 AVX-512: lowers to vpopcntq, like the removed native kernel.
• ARM64: lowers to roughly the same vcnt and pairwise-add sequence as the removed NEON kernel. Expect parity with develop rather than the earlier 0.60–0.86×.

Benchmarks

select_last_vortex_buffer on an AVX2-only x86 VM (no AVX-512, so develop uses the scalar chunk scan). Each figure is the fastest of 4 alternating runs × 200 samples, pinned to one core:

Bits	develop	This PR	Ratio
512	11.6 ns	9.8 ns	0.84×
65,536	150.4 ns	151.5 ns	1.01×
1,048,576	2,098 ns	2,103 ns	1.00×

The byte-sum revision measured 2.14× and 2.36× at 65,536 and 1,048,576 bits, which is why it was replaced. The AVX-512 and NEON results come from the CodSpeed runners on this PR.

Testing

• cargo nextest run -p vortex-buffer --release: 891 passed (x86_64).
• cargo nextest run -p vortex-buffer --target aarch64-unknown-linux-gnu --release under qemu-aarch64: 889 passed.
• cargo miri nextest run -p vortex-buffer with CI's MIRIFLAGS: 221 passed. All 23 select tests also passed under Miri with --run-ignored all.
• cargo clippy -p vortex-buffer -- -D warnings: clean on x86_64 (--all-targets --all-features), aarch64 and wasm32.
• Not run: native ARM64 or AVX-512 benchmarks.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YUbVgtk4iks9jBGhqe5pTF

[codspeed]

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.

⚠️ 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

⚡ 28 improved benchmarks
❌ 1 regressed benchmark
✅ 2072 untouched benchmarks
🆕 3 new benchmarks
⏩ 518 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	bitwise_not_vortex_buffer[128]	4 µs	5.4 µs	-25.69%
⚡	Simulation	fsst_decompress_string	526 µs	188.4 µs	×2.8
⚡	Simulation	chunked_canonicalize_into[(20, 100, 16, 4)]	739.9 µs	340.6 µs	×2.2
⚡	Simulation	chunked_dict_fsst_into_canonical[(250, 10, 2)]	681.9 µs	341.1 µs	+99.91%
⚡	Simulation	chunked_canonicalize_into[(20, 50, 8, 4)]	274.4 µs	213.1 µs	+28.78%
⚡	Simulation	bitwise_not_vortex_buffer[2048]	5.8 µs	4.5 µs	+28.59%
⚡	Simulation	chunked_canonicalize_into[(20, 50, 8, 16)]	279.5 µs	218.3 µs	+28.05%
⚡	Simulation	chunked_canonicalize_into[(20, 50, 8, 64)]	306.3 µs	242.8 µs	+26.16%
⚡	Simulation	chunked_into_canonical[(20, 50, 8, 4)]	314.3 µs	252.2 µs	+24.59%
⚡	Simulation	chunked_into_canonical[(20, 50, 8, 16)]	323.3 µs	261.5 µs	+23.63%
⚡	Simulation	chunked_into_canonical[(20, 50, 8, 64)]	348.2 µs	285.7 µs	+21.87%
⚡	Simulation	decompress_fsst[(200, 4, 4)]	60.8 µs	51 µs	+19.21%
⚡	Simulation	decompress_fsst[(200, 4, 8)]	61.6 µs	51.9 µs	+18.76%
⚡	Simulation	chunked_canonicalize_into[(10, 150, 4, 4)]	235.3 µs	199.5 µs	+17.94%
⚡	Simulation	decompress_fsst[(200, 16, 4)]	66.4 µs	56.6 µs	+17.3%
⚡	Simulation	decompress_fsst[(200, 16, 8)]	69.4 µs	59.7 µs	+16.21%
⚡	Simulation	chunked_canonicalize_into[(10, 150, 16, 4)]	266.6 µs	230.7 µs	+15.57%
⚡	Simulation	chunked_canonicalize_into[(20, 100, 16, 16)]	461.3 µs	399.7 µs	+15.43%
⚡	Simulation	chunked_into_canonical[(10, 150, 4, 4)]	264.9 µs	230.4 µs	+15%
⚡	Simulation	chunked_into_canonical[(20, 100, 16, 4)]	452.7 µs	397.6 µs	+13.85%
...	...	...	...	...	...

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 ji/fearless-simd-wins (07f1e33) with develop (3c87da0)²}

Footnotes

1. 518 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 develop (d14fa26) during the generation of this report, so 3c87da0 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩