Gpu devicetokenhashes

Home › Gpu › GPU kernels

DeviceTokenHashes

One document's token hashes per row, held flat on the accelerator.

public sealed class DeviceTokenHashes : IDisposable

Example — the shape a caller writes.

using Lodestar.Gpu.Compute;

using var context = GpuContext.Create(preferCpu: true);
uint[][] documents = [[11u, 22u, 33u], [22u, 44u]];
using var resident = DeviceTokenHashes.Upload(context, documents);

int held = resident.Count;  // => 2

Members — one page each.

Member	What it does
DeviceTokenHashes.Upload	Flattens a batch of documents and uploads it
DeviceTokenHashes.Dispose	Frees the two device buffers

Properties — Count is how many documents the block holds.

Remarks — hashes rather than tokens, for the reason every other type here takes data rather than a Lodestar type: a kernel parameter must be blittable. A host that hashes first also spares this package an edge to Lodestar.Text, which owns the hash.

That is also the split that makes the kernel worth running. Hashing a token is cheap and sequential, so the host keeps it; minimising over the permutations is O(tokens × permutations), which is the part an accelerator is for.

The hash has to be the one the CPU path uses, or the signatures will not match. Lodestar.Text.Similarity.MinHash hashes a token as the first four bytes of its SHA-1, little-endian — the construction datasketch exports as sha1_hash32. Supply those and the two agree exactly; supply anything else and you have a different sketch, which is a legitimate thing to want and is not parity.

A document may be empty, and repeats are harmless: a minimum is idempotent, which is what makes this a set sketch rather than a bag one.

Applies to — net10.0, netstandard2.1.

See also — TiledMinHashSignatures, the namespace index.