data parallel patching for FastLanes

[a10y]

Summary

This PR adds G-ALP style patches that allow for data parallel access. This allows removing the additional execute_patches kernel launch, and instead inserts patch values inside of the unpacking kernels, at the step when we dispatch from shared -> global memory.

The benchmarks added in #6712 show a significant speedup with this change, and performance is pretty much invariant to patch count.

Background

The G-ALP paper's contribution is modifying the standard layout of "exceptions" (what we call Patches in Vortex) to allow for fully data parallel access. Their target is an f32 ALP decoding kernel, but it is equally applicable for our unpacking kernels.

The core insight is that storing patches in sorted order, which is great for single-threaded execution on a super-scalar CPU, results in very poor GPU performance.

Instead, we need to shuffle them so that they can be accessed ordered by (chunk, lane). The chunk is the normal FastLanes vector chunk size (1024). The lane depends on the width of the type. Doing this means that we get O(1) access to the patches within our kernel.

Replicating Figure 2 from the paper:

This diff looks large, but in reality the big pieces

• Adds a new patches.h header file with shared definitions between the CUDA kernel and the Rust code.
• Adds a patches.cuh with some C++ code to make it easier to seek/iterate over a range of patch values
• Updates the unpacking kernels to accept an optional set of patches

This does not change the memory format for the Patches type, rather this just does a D2H -> transpose on CPU -> H2D transformation.

Testing

Uses the existing test suite which includes a lot of bit-unpacking with patches.

[codspeed-hq]

Merging this PR will improve performance by 17.64%

⚡ 1 improved benchmark
✅ 953 untouched benchmarks
⏩ 1466 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	chunked_opt_bool_canonical_into[(10, 1000)]	1.6 ms	1.4 ms	+17.64%

---

_{Comparing aduffy/patches-transposed (3e53f51) with develop (0abfb33)}

Footnotes

1. 1466 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. ↩