[wip] PatchedArray

[a10y]

Summary

• new PatchedArray
• On read, BitPackedArray w/Patches -> PatchedArray(BitPackedArray)
• Slice
• Filter
• Take
• ALP
• ALP-RD (?)

Closes: #000

Testing

[codspeed-hq]

Merging this PR will improve performance by 22.59%

⚡ 3 improved benchmarks
✅ 391 untouched benchmarks
🆕 6 new benchmarks
⏩ 2052 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
🆕	Simulation	bench_patch_transpose[100]	N/A	201 µs	N/A
🆕	Simulation	bench_patch_transpose[10]	N/A	197.2 µs	N/A
🆕	Simulation	bench_patch_transpose[1]	N/A	197.3 µs	N/A
🆕	Simulation	bench_patch_transpose[2048]	N/A	277.5 µs	N/A
🆕	Simulation	bench_patch_transpose[65536]	N/A	2.7 ms	N/A
🆕	Simulation	bench_patch_transpose[1024]	N/A	238.4 µs	N/A
⚡	Simulation	take_map[(0.1, 1.0)]	4.2 ms	3.6 ms	+17.09%
⚡	Simulation	take_map[(0.1, 0.5)]	2.6 ms	2.1 ms	+22.59%
⚡	Simulation	take_map[(0.1, 0.1)]	1,007.5 µs	911.1 µs	+10.57%

---

_{Comparing aduffy/patched (5580564) with develop (5d6a3c8)²}

Footnotes

1. 2052 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 (761c404) during the generation of this report, so 5d6a3c8 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[gatesn]

Moving my comment from another PR:

Unless we have some way of passing down an "is_defined" mask into execution, there's no way for the child array to know whether some checked operation should fail or not.

e.g. the child of a Patches array may overflow on some arithmetic, but the patch would have succeeded. We need a way to tell the child to ignore the overflow for the invalid positions.

[a10y]

do you think that's something we could ferry in the executionctx?

[a10y]

Alternatively you can just not pushdown scalarfns to the child and force it to be done after patching? It's clear to me why you'd wanna push down cardinality reducing ops (e.g. Slice) but idk why it matters if fallible scalar fns like Sum or Binary need the patched result

[gatesn]

The push-down makes sense, but it seems a shame that we cannot have a specialized kernel for ALP::compare because there's a patch array in the middle. I guess compare can be pushed down? Because it isn't a checked operation?

In fact, I think we already have this defined as ScalarFnVTable::is_fallible

[gatesn]

Other option ofc is ArrayVTable::execute(selection: Mask, defined: Mask) - but 🤷