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.