Exploiting Online Locality and Reduction Parallelism for Sampled Dense Matrix Multiplication on GPUs.