BLIS 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

BLIS 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.