CudaDMA: Optimizing GPU Memory Bandwidth via Warp Specialization