9 ms·
I'm sure there's more to it, but just comparing the profile output shows aggressive use of prefetch and broadcast instructions.
by marshallward 2y ago
I'm sure there's more to it, but just comparing the profile output shows aggressive use of prefetch and broadcast instructions.
- jart 2y agoBLIS does that in their kernels. I've tried doing that but was never able to get something better than half as good as MKL. The BLIS technique of tiling across k also requires atomics or an array of locks to write output.