Generating optimal CUDA sparse matrix-vector product implementations for evolving GPU hardware