add AVX512 support for filtering in place

[connortsui20]

Reorganizes some of the filter implementations and adds AVX512 filter implementation and benchmarks

[codecov]

Codecov Report

❌ Patch coverage is 75.83643% with 130 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.45%. Comparing base (c75c8a3) to head (eae0511).

Files with missing lines	Patch %	Lines
vortex-compute/src/filter/slice/simd_compress.rs	9.57%	85 Missing ⚠️
vortex-compute/src/filter/slice/out/by_mask.rs	55.55%	28 Missing ⚠️
vortex-compute/src/filter/slice/out/avx512.rs	91.07%	5 Missing ⚠️
vortex-compute/src/filter/slice/out/by_bitview.rs	87.50%	5 Missing ⚠️
vortex-compute/src/filter/slice/in_place/avx512.rs	92.30%	4 Missing ⚠️
vortex-compute/src/filter/slice/out/mod.rs	98.38%	2 Missing ⚠️
vortex-compute/src/filter/slice/in_place/mod.rs	99.08%	1 Missing ⚠️

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[a10y]

Does GitHub give us CI runners with AVX512 support? My recollection from when I did the AVX2 take kernel is that they don't

[connortsui20]

not really sure how to deal with this, I think sometimes they have it, sometimes not?

[robert3005]

CI runs in AWS, you can get whatever machine you want, right now by default we use m7i/m7a which are sapphire rapids/zen4 which have AVX512. On benchmarks we run c6i which is ice-lake sp which also has avx512

[0ax1]

Heads up that we can't micro-benchmark avx512 with codspeed, as valgrind/callgrind does not support avx512. Saying you'll get numbers for avx2 if it is implemented:

      - name: Build benchmarks (shard 1)
        env:
          RUSTFLAGS: "-C target-feature=+avx2"
      ...

[connortsui20]

Given that were pivoting to batch I'm going to put a pin in this.

[gatesn]

Nooo I like this!

[connortsui20]

The items here are moved to the out module in slice

[codspeed-hq]

CodSpeed Performance Report

Merging #5399 will not alter performance

_{Comparing ct/avx512-filter (eae0511) with develop (c75c8a3)}

Summary

✅ 1478 untouched
🆕 56 new
⏩ 214 skipped¹

Benchmarks breakdown

	Benchmark	BASE	HEAD	Change
🆕	in_place_scalar[(1024, 0.0)]	N/A	4.6 µs	N/A
🆕	in_place_scalar[(1024, 0.1)]	N/A	6.2 µs	N/A
🆕	in_place_scalar[(1024, 0.25)]	N/A	7 µs	N/A
🆕	in_place_scalar[(1024, 0.5)]	N/A	7.6 µs	N/A
🆕	in_place_scalar[(1024, 0.75)]	N/A	8.2 µs	N/A
🆕	in_place_scalar[(1024, 0.9)]	N/A	8.5 µs	N/A
🆕	in_place_scalar[(1024, 1.0)]	N/A	8.8 µs	N/A
🆕	in_place_scalar[(131072, 0.0)]	N/A	553.9 µs	N/A
🆕	in_place_scalar[(131072, 0.1)]	N/A	784.8 µs	N/A
🆕	in_place_scalar[(131072, 0.25)]	N/A	866 µs	N/A
🆕	in_place_scalar[(131072, 0.5)]	N/A	943.8 µs	N/A
🆕	in_place_scalar[(131072, 0.75)]	N/A	1 ms	N/A
🆕	in_place_scalar[(131072, 0.9)]	N/A	1.1 ms	N/A
🆕	in_place_scalar[(131072, 1.0)]	N/A	1.1 ms	N/A
🆕	in_place_scalar[(16384, 0.0)]	N/A	69.5 µs	N/A
🆕	in_place_scalar[(16384, 0.1)]	N/A	97.8 µs	N/A
🆕	in_place_scalar[(16384, 0.25)]	N/A	108.1 µs	N/A
🆕	in_place_scalar[(16384, 0.5)]	N/A	117.7 µs	N/A
🆕	in_place_scalar[(16384, 0.75)]	N/A	126.8 µs	N/A
🆕	in_place_scalar[(16384, 0.9)]	N/A	132.2 µs	N/A
...	...	...	...	...

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

Footnotes

1. 214 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]

@0ax1 is there a way for me to disable the benchmark just for codspeed?

[a10y]

note: i think you can do this faster by always doing speculative write to data[write_pos], and then just incrementing write_pos by the mask bit on each loop iteration. FSST uses a similar trick in its compression loop.

maybe save for FLUP though so we can benchmark the difference

[connortsui20]

yeah this entire function is super inefficient but I'm going to wait on optimizing all of these further until we actually start using it in operators and can get better / more realistic benchmarks

[gatesn]

We probably shouldn't merge SIMD code unless it is fully optimized. It's a bit misleading

[connortsui20]

I could just remove this then? This isn't SIMD code

[0ax1]

@0ax1 is there a way for me to disable the benchmark just for codspeed?

You shouldn't need to. We compile for AVX2, so the AVX512 version shouldn't be picked up.

Ah I see that we're running into failed to execute the benchmark process, exit code: 132 which I assume is picked based on runtime capabilities of the CPU?

Discussed to exclude the benchmark via: #[cfg(not(codspeed))]

[connortsui20]

hmm it seems I've done something wrong here then: https://github.com/vortex-data/vortex/actions/runs/19644046057/job/56254751182?pr=5399