perf: enable POPCNT in the BMI2, AVX2 and AVX-512 kernels

[mightsleep]

Summary

#[target_feature(enable = "bmi2")], "avx2" and "avx512f,..." do not imply POPCNT, so count_ones() inside these kernels compiles to a SWAR sequence (with imul for narrow masks) in any build without target-cpu. The bench runners use target-cpu=native, so they won't show a difference.

Changes

Add popcnt to the target features and to the runtime detection of the intersect_by_rank BMI2 kernels and the x86 simd_compress kernels. No other #[target_feature] function in the tree counts bits without it (the fearless_simd levels already enable POPCNT).

Default build, Zen 5, divan medians:

bench	before	after
intersect_by_rank 100k random	3.17 µs	1.93 µs	1.64x
random_rotating 0.5/0.5	9.71 µs	5.67 µs	1.71x
very_sparse_mask_cached 10%/1%	1.91 µs	0.81 µs	2.36x
filter_fixed_width random_i16 0.8	1.60 µs	1.06 µs	1.51x
filter_fixed_width random_i32 0.5	1.49 µs	1.03 µs	1.45x
filter_fixed_width random_i64 0.5	1.13 µs	0.99 µs	1.14x
filter_fixed_width random_i128 0.5 (no kernel)	1.45 µs	1.42 µs	1.02x

The filter numbers are the AVX-512 kernels; the AVX2 ones should show up in the CodSpeed simulation shard, where AFAIK valgrind has no AVX-512.

[robert3005]

Thanks!

[codspeed]

Merging this PR will regress 3 benchmarks

⚠️ 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.

⚠️ 10 benchmarks spent significant time in system calls

System calls cannot be consistently instrumented, so they are not included in the measure, which understates the real cost. Please switch to the Walltime instrument to accurately measure system calls.

Measurement and system calls

• random_i128[0.01]
• filter_run_end[len=4096_run=16_density=0.1]
• decompress[u32, (1000, 256)]
• decompress[u16, (4000, 256)]
• zstd_compress_string
• multipolygon_mixed_validity
• pcodec_compress_f64
• take_varbin
• ilike_contains
• from_vec_drop_vortex[65536]

⚠️ 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

⚡ 29 improved benchmarks
❌ 3 regressed benchmarks
✅ 2181 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	filter_avx2[80]	2.5 µs	2.9 µs	-13.72%
❌	WallTime	filter_avx2[64]	2.4 µs	2.7 µs	-12.25%
❌	WallTime	threshold_neon[8]	2.9 µs	3.3 µs	-11.83%
⚡	WallTime	random_rotating_avx512[(0.05, 0.5, "self_5pct_mask_50pct")]	17.9 µs	9.2 µs	+94.28%
⚡	WallTime	random_rotating_avx512[(0.1, 0.5, "self_10pct_mask_50pct")]	19.8 µs	10.3 µs	+91.75%
⚡	WallTime	random_rotating_avx2[(0.02, 0.5, "self_2pct_mask_50pct")]	16.3 µs	8.6 µs	+89.11%
⚡	WallTime	random_rotating_avx512[(0.02, 0.5, "self_2pct_mask_50pct")]	16.1 µs	8.5 µs	+89%
⚡	WallTime	random_rotating_avx2[(0.05, 0.5, "self_5pct_mask_50pct")]	17.9 µs	9.5 µs	+88.52%
⚡	WallTime	random_rotating_avx2[(0.1, 0.5, "self_10pct_mask_50pct")]	19.4 µs	10.3 µs	+88.08%
⚡	WallTime	random_rotating_avx2[(0.5, 0.5, "self_50pct_mask_50pct")]	19.7 µs	10.8 µs	+82.69%
⚡	WallTime	random_rotating_avx512[(0.5, 0.5, "self_50pct_mask_50pct")]	19.7 µs	10.8 µs	+81.76%
⚡	WallTime	random_rotating_avx512[(0.9, 0.9, "self_90pct_mask_90pct")]	22.2 µs	13.8 µs	+60.3%
⚡	WallTime	random_rotating_avx2[(0.9, 0.9, "self_90pct_mask_90pct")]	22.2 µs	14.1 µs	+57.91%
⚡	Simulation	rank_indices[(0.05, 0.05, "self_sparse_rank_sparse")]	60.1 µs	48.5 µs	+23.89%
⚡	WallTime	threshold_avx512[8]	2.1 µs	1.8 µs	+16.83%
⚡	WallTime	random_rotating_avx512[(0.5, 0.02, "self_50pct_mask_2pct")]	58.3 µs	50.7 µs	+14.88%
⚡	Simulation	rank_indices[(0.5, 0.05, "self_dense_rank_sparse")]	92.4 µs	80.8 µs	+14.39%
⚡	Simulation	density_matrix[(0.05, 0.5, "self_sparse_mask_dense")]	62.8 µs	54.9 µs	+14.33%
⚡	Simulation	density_matrix[(0.05, 0.05, "self_sparse_mask_sparse")]	63.1 µs	55.2 µs	+14.24%
⚡	Simulation	rank_indices[(0.05, 0.5, "self_sparse_rank_dense")]	62.9 µs	55.1 µs	+14.22%
...	...	...	...	...	...

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 mightsleep:popcnt-kernels (3be756d) with develop (d42e7c9)}