perf(crypto): vectorized BLAKE3 / LtHash and cached key parsing #112

New issue

Closed

opened 2026-09-24 12:28:55 +00:00 by Grandiras · 0 comments

Grandiras commented

2026-09-24 12:28:55 +00:00

Owner

Part of the v0.7.0 update cycle. The CPU cost of Spaces syncing and signature verification.

Scope

BLAKE3: keep the compression state in locals, precompute the message schedule, and allocate nothing per hash. XOF output blocks are independent, so compress 8 counters at once with Vector256<uint>, with a scalar fallback. A prototype measured 11.9 µs → 0.72 µs (16.6×) for 2048 B of output.
LtHash: add and subtract lanes with Vector256<ushort> over MemoryMarshal.Cast of the XOF output (little-endian, with a scalar fallback). A prototype measured LtHash.Add at 13.1 µs → 0.77 µs (17×) with 0 B allocated. A 10k-record SpaceRepoCommit.FromIndex drops from 131 ms to about 8 ms.
Signature verification: cache parsed did:key public keys (bounded, keyed by the did:key string). Make the curve parameters static readonly, precompute the square-root exponent, decode Base58 into a span, and stop IsLowS allocating. A P-256 verify measured 90 µs → 41 µs.
Internal duplication in AtProtoCrypto: AtProtoKey.GetCompressedPublicKey vs CompressPublicKey; DecompressPoint re-deriving the curve from OIDs; repeated OID literals; FromMultikey repeating MultibaseToBytes; and the curve→JWS alg mapping written twice (ServiceAuthGenerator, SpaceTokens), which becomes one helper.
Remove the unused Blake3.KeyLen.

Tests

Add the atproto-interop-tests crypto fixtures (w3c_didkey_P256.json, w3c_didkey_K256.json, signature-fixtures.json), including low-S/high-S cases.
Fix K256_GenerateAndSign_IfSupported, which passes silently on PlatformNotSupportedException.
The existing 27 BLAKE3 reference vectors and the lane-for-lane LtHash tests must stay green. Add input lengths 0…1024 compared against the scalar path.

Acceptance

Bit-exact results against all vectors.
Before/after BenchmarkDotNet numbers in the PR description.
No public API change.

Part of the v0.7.0 update cycle. The CPU cost of Spaces syncing and signature verification. ## Scope - [ ] **BLAKE3:** keep the compression state in locals, precompute the message schedule, and allocate nothing per hash. XOF output blocks are independent, so compress 8 counters at once with `Vector256<uint>`, with a scalar fallback. A prototype measured 11.9 µs → 0.72 µs (16.6×) for 2048 B of output. - [ ] **LtHash:** add and subtract lanes with `Vector256<ushort>` over `MemoryMarshal.Cast` of the XOF output (little-endian, with a scalar fallback). A prototype measured `LtHash.Add` at 13.1 µs → 0.77 µs (17×) with 0 B allocated. A 10k-record `SpaceRepoCommit.FromIndex` drops from 131 ms to about 8 ms. - [ ] **Signature verification:** cache parsed `did:key` public keys (bounded, keyed by the did:key string). Make the curve parameters `static readonly`, precompute the square-root exponent, decode Base58 into a span, and stop `IsLowS` allocating. A P-256 verify measured 90 µs → 41 µs. - [ ] **Internal duplication in `AtProtoCrypto`:** `AtProtoKey.GetCompressedPublicKey` vs `CompressPublicKey`; `DecompressPoint` re-deriving the curve from OIDs; repeated OID literals; `FromMultikey` repeating `MultibaseToBytes`; and the curve→JWS `alg` mapping written twice (`ServiceAuthGenerator`, `SpaceTokens`), which becomes one helper. - [ ] Remove the unused `Blake3.KeyLen`. ## Tests - [ ] Add the `atproto-interop-tests` crypto fixtures (`w3c_didkey_P256.json`, `w3c_didkey_K256.json`, `signature-fixtures.json`), including low-S/high-S cases. - [ ] Fix `K256_GenerateAndSign_IfSupported`, which passes silently on `PlatformNotSupportedException`. - [ ] The existing 27 BLAKE3 reference vectors and the lane-for-lane LtHash tests must stay green. Add input lengths 0…1024 compared against the scalar path. ## Acceptance - Bit-exact results against all vectors. - Before/after BenchmarkDotNet numbers in the PR description. - No public API change.

Grandiras referenced this issue

2026-09-24 12:39:04 +00:00

feat(identity)!: resolver abstraction, identity caching and SSRF-safe fetching #120

Grandiras referenced this issue

2026-09-24 12:39:33 +00:00

Tracking: v0.7.0 update cycle — protocol catch-up, optimization, debloat, API reshaping #141

Grandiras referenced this issue from a commit

2026-09-24 18:14:56 +00:00

perf(crypto): vectorized BLAKE3 / LtHash and cached did:key parsing (closes #112)

Grandiras referenced this issue from a pull request that will close it,

2026-09-24 18:14:58 +00:00

perf(crypto): vectorized BLAKE3 / LtHash and cached did:key parsing #150