Hi,

> I'm implementing a neural network in the kernel, which needs floating
> point precision.

I wonder why fixed point rationals aren't enough for you...? You could
even get more precision out of them.

	Regards,
		-- Jachym Holecek