Optimizing Byte Matrix Multiplication with AVX-VNNI

This article explores optimizing byte matrix multiplication using the AVX-VNNI instruction set. The author begins with a naive implementation, then uses the gemmology and xsimd libraries to create optimized versions employing transposition and a custom layout. Benchmark results show the custom layout achieves the best performance, leveraging the vpdpbusd instruction for significant efficiency gains. The article delves into the implementation details of gemmology's maddw function and its architectural variations.