I'd intended to submit this after seeing the results recently while referring to it apropos BLAS implementation. I've quoted matmul here as the classic example

I'd intended to submit this after seeing the results recently while referring to it apropos BLAS implementation. I've quoted matmul here as the classic example of compiler optimization failures and won't again, at least until I've understood these results. I couldn't see how to reproduce the results and it's not clear what the various matrix sizes are, related to the dips in MKL performance that I don't understand where polly wins. Also, it's not clear what the POWER results are, as BLIS doesn't support it other than via the very slow generic configuration; I haven't got round to enquiring.