<feed xmlns='http://www.w3.org/2005/Atom'>
<title>ydb/library/cpp/dot_product/dot_product_vnni.cpp, branch main</title>
<subtitle>Mirror of YDB github repos</subtitle>
<id>https://code.mastervirt.ru/ydb/atom?h=main</id>
<link rel='self' href='https://code.mastervirt.ru/ydb/atom?h=main'/>
<link rel='alternate' type='text/html' href='https://code.mastervirt.ru/ydb/'/>
<updated>2026-07-09T19:25:51Z</updated>
<entry>
<title>Add VNNI int8 dot product fast path</title>
<updated>2026-07-09T19:25:51Z</updated>
<author>
<name>mirali777</name>
<email>mirali777@yandex-team.com</email>
</author>
<published>2026-07-09T18:46:42Z</published>
<link rel='alternate' type='text/html' href='https://code.mastervirt.ru/ydb/commit/?id=b7a0527229fb38319bd9f9a55d9e714c34a63b49'/>
<id>urn:sha1:b7a0527229fb38319bd9f9a55d9e714c34a63b49</id>
<content type='text'>
Adds AVX512 VNNI runtime dispatch for signed int8 dot products used by Link HNSW distance calculations, with safe fallback to the existing AVX2/SSE implementations. The signed int8 VNNI path corrects the u8*i8 bias per 64-byte chunk so accumulator semantics stay aligned with the existing int8 DotProduct implementation.

Verification:
- `ya make -tt library/cpp/dot_product/ut`
- `ya make library/cpp/dot_product/bench`
- `ya make -r library/cpp/dot_product/bench`
- `ya make -tt library/cpp/dot_product/ut --sanitize=address`
- `ya make -tt library/cpp/dot_product/ut --sanitize=undefined -F 'TDocProductTestSuite::TestDotProduct8' -F 'TDocProductTestSuite::TestDotProduct8VnniEdges' -F 'TDocProductTestSuite::TestDotProductCharStability'`
- `ya make -tt util/system/ut -F TestCpuId`
commit_hash:5624243c26df2a9e067e631477f6c3603ef45626
</content>
</entry>
</feed>
