Altivec sws code intentionally overreads buffers for better performance,
so we need to allocate larger buffers to handle that.