The "usual" way to use 3dDeconvolve is to provide a timing file that gives the start times for each stimulus (which are
not required to be on the TR timing grid) -- this file is specified using the
-stim_times option.
For the HRF, you choose a model function to define the shape of the response for the stimulus. There are many options for that.
The "best" way to run 3dDeconvolve (or its daughter program, 3dREMLfit) is to use the AFNI
super script afni_proc.py -- whose
help is here. This script takes as input the anatomical and EPI datasets, plus the timing files, and will run the entire analysis for you: aligning the anatomical and EPI datasets together and to MNI/Talairach space, running the GLM analysis, and writing out some simple diagnostic reports about the data.