Thank you rick for your reply:)

I was fiddling to see whether upsampling a little bit before running 3dvolreg would be helpful for getting more accurate realignment, especially when only a limited portion of the brain is covered in the functional images.
In the above circumstance, I found 3dvolreg actually became the bottleneck in the pipeline, consuming much more time than 3dAllineate and 3dDeconvolve, contrary to my expectation.
Without upsampling, 3dvolreg is fast, as you mentioned.

I think 3dvolreg is a perfect candidate for parallelization in term of the nature of its computation, given that each sub-brick is independently (please correct me if this is not the case) registered to the template.

Do you think this feature would be on your list in the future?:)
Thanks~