I think in order to do that the ground would have to be prepared with a modified version of SRFI 144 for half-floats.  I don't know which of its procedures, which result from a merger between R6RS and <math.h>, are still relevant, but someone sufficiently knowledgeable and/or experienced should take a look.  For that matter, there is no SRFI for 32-bit floats either.

I have toyed around with the idea of a fully or partly table-driven implementation of 8-bit floats in which all the values are integral (see the second table at <https://en.wikipedia.org/wiki/Minifloat>), but have put it aside as too low-priority for now.

On Fri, Apr 24, 2020 at 3:54 PM Jeffrey Mark Siskind <xxxxxx@purdue.edu> wrote:
One thing that came to mind is that the SRFI does not support half-floats
which are popular with GPUs.

    Jeff (http://engineering.purdue.edu/~qobi)