Phoebe,
Granger causality analysis identifies connectivity based on consecutive time points (time series). What kind of experiment do you have, event-related or blocked design? Looks like more likely the former since you modeled the HRF with TENT. If you're focusing on TaskA, it's hard to tell whether you have to extract those time points relevant to TaskA or not.
There are two approaches to handling the other two tasks: (1) regressing out before extracting out those time points using 3dSynthesize; (2) extracting those regressors with the design matrix generated by 3dDeconvolve with option -xsave.
(1) is probably more favorable because you may run out of degree of freedom in Granger causality analysis.
Gang