5 ms·

Very nice write up. Those are the kind of matrix sizes that MKL is fairly good at, might be worth a comparison as well? Also, if you were designing for smalle

Very nice write up. Those are the kind of matrix sizes that MKL is fairly good at, might be worth a comparison as well? Also, if you were designing for smaller cases, say MNK=16 or 32, how would you approach it differently? I'm implementing neural ODEs and this is one point I've been considering.

pmeira 2y ago
Don't forget BLIS itself!
gnufx 2y ago
For very small sizes on amd64, you can, and likely should, use libxsmm. MKL's improved performance in that region was due to libxsmm originally showing it up, but this isn't an Intel CPU anyway.