This is partially true (or at least once was, the performance gap is pretty slim nowadays), but most of the high-performance BLAS implementations like OpenBLAS

This is partially true (or at least once was, the performance gap is pretty slim nowadays), but most of the high-performance BLAS implementations like OpenBLAS or MKL have their hot path code written in hand-tuned assembler anyway.