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.