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

[lkcl]

On Thursday, August 8, 2019 at 2:28:43 PM UTC+8, Luis Vitorio Cargnini(OURS/RiVAI) wrote:
 
> No, just adopt IEEE-754, it is a standard, it is a standard for a reason. Anything out of IEEE-754, it does not conform with IEEE and for such you are on your own. 

Just to emphasise, Luis, Andrew: "on their own" is precisely what each of the proprietary 3D GPU Vendors have done, and made literally billions of dollars by doing so.

Saying "we are on our own" and expecting that to mean that not conforming to IEEE754 would kill the proposal, this is false logic.

MALI (ARM), Vivante, the hated PowerVR, NVidia, AMD/ATI, Samsung's new GPU (with Mitch's work in it), and many more, they *all* went "their own way", hid the hardware behind a proprietary library, and  *still made billions of dollars*.

This should tell you what you need to know, namely that a new 3D GPU Platform Spec which has specialist FP accuracy requirements to meet the specific needs of this *multi BILLION dollar market* is essential to the proposal's successful industry adoption.

If we restrict it to UNIX (IEEE754) it's dead.

If we restrict it to *not* require IEEE754, it's dead.

The way to meet all the different industry needs: new Platform Specs.

That does not affect the actual opcodes. They remain the same, no matter the Platform accuracy requirements.

Thus the software libraries and compilers all remain the same, as well.

L.