You want the "-warp" dataset to be the +tlrc dataset that you defined via the markers, etc. -- this is the one that has the transformation matrices between +orig and +tlrc coordinates. If you marked on dataset spgr+orig, then the "-warp" dataset should be spgr+tlrc.
If your mask contains more than one value (you drew several ROIs, each marked with a different dataset value), then you also want to use the "-vote" option to preserve these values.
Thus, I'd recommend something more like
3dfractionize \
-template 3dtimeseries+orig \
-warp spgr+tlrc \
-input ROImasks+tlrc \
-clip 0.2 -vote -prefix ROImask