The papers referenced in the BLIS repo are the authoritative reference to understand this stuff. I've no idea why people think optimized BLASes aren't performa

The papers referenced in the BLIS repo are the authoritative reference to understand this stuff. I've no idea why people think optimized BLASes aren't performant; you should expect >90% of CPU peak for sufficiently large matrices, and serial OpenBLAS was generally on a par with MKL last I saw. (BLAS implements GEMM, not matmul, as the basic linear algebra building block.) I don't understand using numpy rather than the usual benchmark frameworks, and on Zen surely you should compare with AMD's BLAS (based on BLIS). BLIS had a better parallelization story than OpenBLAS last I knew. AMD BLIS also has the implementation switch for "small" dimensions, and I don't know if current OpenBLAS does. SIMD intrinsics aren't needed for micro-kernel vectorization, as a decent C compiler will fully vectorize and unroll it. BLIS' pure C micro-kernel gets >80% of the performance of the hand-optimized implementation on Haswell with appropriate block sizes. The difference is likely to be due to prefecth, but I don't properly understand it.