Gpu devicetextblock

Development build. This page describes main, not a released package. The latest published Lodestar.Gpu is 0.1.0 — read its documentation.

Home › Gpu › GPU kernels

DeviceTextBlock

A batch of strings renamed to a dense alphabet and held on the accelerator.

public sealed class DeviceTextBlock : IDisposable

Example — the shape a caller writes.

using Lodestar.Gpu.Compute;

using var context = GpuContext.Create(preferCpu: true);
using var block = DeviceTextBlock.Upload(context, "kitten", ["sitting", "mitten"]);

int held = block.Count;  // => 2
int alphabet = DeviceTextBlock.MaxPatternAlphabet;  // => 255

Members — one page each.

Member	What it does
DeviceTextBlock.Upload	Renames a batch against a pattern's alphabet and uploads it
DeviceTextBlock.Dispose	Frees the two device buffers

Properties — Count is how many strings the block holds. MaxPatternAlphabet is the constant 255, how many distinct characters a pattern may hold.

Remarks — a kernel parameter has to be blittable, so nothing here carries a string. The batch is renamed on the host into one flat array of symbol codes plus the offsets cutting it back into rows. That renaming is also what lets the equality masks be a 256-entry table rather than a lookup per character.

A character the pattern does not hold renames to one reserved code, because a mask no pattern position sets is the same mask for all of them. 255 rather than 256 distinct pattern characters for exactly that reason: one slot is spoken for.

The renaming is the expensive half, and it is on the host. Measured: scoring a resident batch ran 28× to 146× faster than the CPU path, while renaming and uploading per call ran 1.7× to 2.9× slower — a pass over every character, in the language the baseline is already written in. Upload once, score many times, or do not use this kernel.

Applies to — net10.0, netstandard2.1.

See also — BitParallelEditDistance, the namespace index.