feat[cuda]: filter for Decimal/VarBinView

[a10y]

• Wire in BufferHandles for VarBinView
• Wire up cub filtering logic for decimals/strings

[codspeed-hq]

CodSpeed Performance Report

Merging this PR will degrade performance by 31.56%

_{Comparing aduffy/filter-decimals (08acb19) with develop (f568de5)}

⚠️ 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.

Summary

⚡ 62 improved benchmarks
❌ 12 regressed benchmarks
✅ 1105 untouched benchmarks
⏩ 1323 skipped benchmarks¹

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	10M_10pct[1000000]	131.9 µs	165.7 µs	-20.41%
❌	WallTime	1M_90pct[1000000]	57.1 µs	83.4 µs	-31.56%
❌	Simulation	chunked_varbinview_canonical_into[(10, 1000)]	1.6 ms	2.1 ms	-23.07%
❌	Simulation	chunked_varbinview_into_canonical[(10, 1000)]	1.6 ms	2.1 ms	-22.96%
⚡	Simulation	bench_compare_sliced_dict_varbinview[(1000, 10000)]	632.7 µs	156.5 µs	×4
⚡	Simulation	bench_compare_sliced_dict_varbinview[(10000, 10000)]	876.9 µs	491.2 µs	+78.5%
⚡	Simulation	bench_compare_sliced_dict_varbinview[(2000, 10000)]	741.7 µs	218 µs	×3.4
⚡	Simulation	bench_compare_sliced_dict_varbinview[(20000, 10000)]	1,212.3 µs	687 µs	+76.47%
⚡	Simulation	bench_compare_sliced_dict_varbinview[(3333, 10000)]	887.3 µs	299 µs	×3
⚡	Simulation	bench_compare_sliced_dict_varbinview[(5000, 10000)]	1,069.7 µs	398.7 µs	×2.7
⚡	Simulation	bench_compare_sliced_dict_varbinview[(2500, 10000)]	796.4 µs	248.3 µs	×3.2
⚡	Simulation	bench_compare_sliced_dict_varbinview[(7500, 10000)]	856 µs	470.7 µs	+81.87%
⚡	Simulation	bench_compare_varbin[(10000, 2048)]	457.7 µs	335.4 µs	+36.48%
⚡	Simulation	bench_compare_sliced_dict_varbinview[(9999, 10000)]	877.1 µs	491.2 µs	+78.54%
⚡	Simulation	bench_compare_varbin[(10000, 512)]	258.4 µs	228.8 µs	+12.92%
⚡	Simulation	bench_compare_varbinview[(10000, 512)]	258.2 µs	228.1 µs	+13.19%
⚡	Simulation	bench_compare_varbinview[(10000, 2048)]	457.7 µs	335.6 µs	+36.38%
⚡	Simulation	decode_varbin[(1000, 128)]	82.9 µs	28.9 µs	×2.9
⚡	Simulation	decode_varbin[(10000, 2)]	677.2 µs	142.2 µs	×4.8
⚡	Simulation	decode_varbin[(1000, 32)]	82.4 µs	28.4 µs	×2.9
...	...	...	...	...	...

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

Footnotes

1. 1323 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. ↩

[0ax1]

Looks great! Couple of questions / remarks.

[0ax1]

I forget to remove this in my PR. We don't want to do a blocking wait but just return the handle. The caller can opt-into waiting for prev operations on the stream to complete, by registering its own compiler driver callback or by calling await_stream_callback(stream).await?;.

[0ax1]

Could collect directly into host_fields, then it doesn't need to be mut.