GPUs: Anatomy of high performance matmul kernels