[libre-riscv-dev] [isa-dev] Re: FP transcendentals (trigonometry, root/exp/log) proposal

[lkcl]

On Thursday, August 8, 2019 at 2:30:38 PM UTC+8, waterman wrote:
> Hi folks,
> 
> 
> We would seem to be putting the cart before the horse.  ISA-level support for correctly rounded transcendentals is speciously attractive, but its utility is not clearly evident and is possibly negative.  It does not make sense to allocate opcode space under these circumstances.

There are definitely alternative (conflicting) directions here which are driven by price and performance in radically different markets.

3D graphics is seriously optimised, in completely different directions from those that drove the IEEE754 standard, and unfortunately it has been left up to secretive proprietary companies to lead that, as the NREs kept going up and up, driving out Number Nine, Matrox, ATI getting bought by AMD and so on.

A new Open 3D Alliance initiative is in its early stage of being formed and the plan is to get some feedback from members on what they want, here.

This proposal is therefore part of "planning ahead", and there are *going* to be diverse and highly specialist requirements for which IEEE754 compliance is just not going to fly.... *and* there are going to be adopters for whom IEEE754 is absolutely essential.

Recognising this, by creating separate Platform Specs (specially crafted for 3D implementors that distinguish them from the Embedded and UNIX specs) is, realistically, the pragmatic way forward.

L.