use cuda version torch.randperm to avoid copy from gpu to cpu and a fatal bug in multi-thread cpu version
