Accelerating Sparse Approximate Matrix Multiplication on GPUs