Matrix Multiplications on GPUs Run Faster When Given "Predictable" Data