Quicker ADC : Unlocking the Hidden Potential of Product Quantization With SIMD

Efficient Nearest Neighbor (NN) search in high-dimensional spaces is a foundation of many multimedia retrieval systems. A common approach to rely on Product Quantization, which allows the storage large vector databases memory and efficient distance computations. Yet, implementations nearest neighbor with Quantization have their performance limited by accesses they perform. Following this observation, Andr\'e et al. proposed Quick ADC up $6\times$ faster $m\times{}4$ product quantizers (PQ) leveraging specific SIMD instructions. Quicker generalization not codes supporting AVX-512, latest revision instruction set. In doing so, faces challenge using efficiently 5,6 7-bit shuffles that do align computer bytes or words. To end, we introduce (i) irregular combining sub-quantizers different granularity (ii) split tables allowing lookup larger than registers. We evaluate multiple indexes including Inverted Multi-Indexes IVF HNSW show it outperforms reference optimized (i.e., FAISS polysemous codes) for numerous configurations. Finally, release an open-source fork enhanced at http://github.com/nlescoua/faiss-quickeradc.