Cache and bandwidth aware matrix multiplication on the GPU