Lodestar.Gpu 0.1.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

MinHash signatures for a batch of documents, one thread per permutation.

public sealed class TiledMinHashSignatures

Example — the shape a caller writes.

using Lodestar.Gpu.Compute;

using var context = GpuContext.Create(preferCpu: true);
using var resident = DeviceTokenHashes.Upload(context, [[11u, 22u], [22u, 33u]]);
var kernel = new TiledMinHashSignatures(context);

ulong[] multipliers = [3UL, 5UL];
ulong[] addends = [13UL, 17UL];
IReadOnlyList<uint[]> signatures = kernel.Signatures(resident, multipliers, addends);

int documents = signatures.Count;  // => 2
int length = signatures[0].Length;  // => 2

Members — one page each.

Remarks — the work splits where the parallelism is. Hashing a token is cheap and sequential, so the host keeps it and hands over DeviceTokenHashes; minimising over the permutations is O(tokens × permutations) and is what the accelerator takes. A thread owns one permutation of one document and reduces in a register, while the document's hashes are tiled through shared memory so a group reads them once rather than once per thread.

The coefficients are supplied rather than seeded, which is the call Lodestar.Text.Similarity.MinHashPermutations already makes for the same reason — deriving them from a seed would make parity depend on a random number generator instead of on an algorithm. Taking them as spans is also what keeps this package free of an edge.

The permuted value is masked with an AND, not a modulo. The two agree only below the mask, so the choice moves every signature rather than rounding one, and the CPU path records the measured pair of values that proves it.

Build this once. Loading compiles, and a cold launch measures ILGPU's compiler.

Applies to — net10.0, netstandard2.1.

See also — DeviceTokenHashes, the namespace index.