shyjsarah commented on PR #82: URL: https://github.com/apache/paimon-vector-index/pull/82#issuecomment-5492635452
Re-reviewed the latest head (`7296eaa`) after the eight follow-up commits since `314c6bf`. The generic AVX2/FMA work fixes the previously reported `dsub>4` regression on capable x86_64 hosts, and the new per-subspace exact fallback improves the large-offset/precomputed-table behavior. The new targeted tests, core suite, and workspace Clippy pass locally. Two blockers remain: 1. **Add/search numerical consistency is still not fixed across all paths.** The direct/FMA 8-bit encoder can choose code 1 while the ordinary, unoptimized-reader, ephemeral-reuse, resident-precomputed, and optimized serialized-reader search paths all rank code 0 first. The new fallback only detects individual distances near the cancellation bound; it does not protect a small pairwise margin between two large nonzero distances. I reproduced the same failure on `7296eaa` through all five paths. 2. **Small adds still regress on supported x86_64 hosts without AVX2/FMA.** `encode_batch_blocked` now always transposes the full codebook. For `d=768,m=192`, this writes about 0.75 MiB before encoding any row. On the scalar runtime fallback, rows 1/7/8/31 were approximately 39%/46%/37%/45% slower than `314c6bf`; the regression disappears at 32 rows, isolating the fixed transpose cost. Minor follow-up: the public distance-table fallback now uses direct/FMA arithmetic in cancellation cases, while 4-bit and byte-stable encoders still select codes with the legacy norm/dot arithmetic. An extreme finite fixture changes top-1 after upgrading the scorer, including after write/reopen. I suggest separating the legacy/byte-stable table contract from the direct-encoded 8-bit IVF-PQ table contract, or versioning the scoring semantics. I still recommend holding the merge until the two blockers above are addressed. Cross-architecture parity coverage for NEON, generic AVX2/FMA, and the scalar fallback would also be valuable. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
