I second @jdtournier's advice to stick with the (single-shell) single-tissue pipeline for the time being. The confusion in the manual is with respect to an upcoming release of single-shell 3-tissue CSD (SS3T-CSD, see this reference, we used it for instance successfully in this, and this / this study). But to be clear: this is not yet publicly available in MRtrix3. Our internal results have been produced with an internal prototype of the method that I built specifically for this purpose. Ironically, this is what's been taking up a lot of my time and has delayed work on a publication / public implementation of the method... let's hope that changes soon.
Apart from that, @jdtournier suggestion to use MSMT-CSD with just WM and CSF as tissue types is a useful one too. This will buy you free-water elimination, which is not per se unimportant at b=2000, and in the presence of certain specific pathologies. Furthermore, using the MSMT-CSD implementation in this way, will also buy you a "hard-constrained" implementation of CSD inherently. Our experiences over here with the latter aspect have been nothing but positive compared to the default "soft-constrained" default CSD. I would personally highly recommend it. I can provide you with further details privately if you wish.
For further related information on these topics, see also this, this, this (recommended), and this post elsewhere on this forum.
However, an important word of caution: however strange it may sound at first (maybe), if pursuing this strategy (
dwi2response dhollander followed by
dwi2fod msmt_csd with only the WM and CSF responses), I personally recommend plugging this step into the current single-tissue pipeline, only using the WM outcome of this all after this very step. The reason for this is that the multi-tissue pipeline does the
mtbin bias field and intensity normalisation step right after this (and no bias field and intensity normalisation step before the CSD step). This very step, however, definitely requires the full 3-tissue CSD results (WM-GM-CSF) to work correctly at higher b-values (b=2000 is high enough in this context). Ironically, at a low b-value (say below b=1000), this wouldn't be so much of an issue.
Finally, if you'd want to pursue using SS3T-CSD itself, just contact me privately; we can always arrange something if and when it makes sense.