The minimum for 3dDeconvolve is to do
make libmri.a libcoxplot.a libf2c.a 3dDeconvolve
If you looked at the rule for making 3dDeconvolve in Makefile.INCLUDE, you could probably figure this out.
3dDeconvolve shouldn't link to Motif, but the link line for it will probably include -lXm since they pretty much all do. If that causes a problem, you'll have to manually remove the offending library.
As far as MPI-izing the program, that is a worthy goal. The main issue is that you have to have a way to collect the output from each process back to the common array regions which will be used to make the final datasets. The current version uses shared memory, allocated with the mmap() Unix library function before the fork()-ing. The structure of the program is not well set up for MPI -- the results from the function regression_analysis() and glt_analysis() go into these arrays in save_voxel(). To avoid huge communication overhead, you'll have to have each process save its results locally, and then send them back in bulk to be assembled into the final volumes.
The speed test on the web site is too small for a real test of a big job. I can let you have a much bigger run, that needs 2.3 GB of scratch space (and only runs when 3dDeconvolve is compiled in 64-bit mode because of that).