Understanding the efficiency of GPU algorithms for matrix-matrix multiplication