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

[lkcl]

On Sunday, September 15, 2019 at 1:57:26 AM UTC+1, Allen Baum wrote:
>
> Actually, it doesn't matter if its more accurate or less; the problem is 
> that it's different,
> That difference probably doesn't matter for some programs, but unless 
> you're only running some very small number of programs where you might be 
> able to tell if it matters  - then the difference matters.
>

3D games / applications are supposed to be very specifically designed not 
to depend on the variation in accuracy.  time and time again i have seen 
advice which specifically states that it is the responsibility of the 
[cross-platform] 3D application developer to run their application on 
multiple independent hardware platforms.

marginally OT: the situation however analogous to the "browser wars", where 
the only organisation that has not given up is Mozilla.  Opera gave up and 
used webkit.  even Microsoft gave up and used webkit

[definitely OT: sigh they should never have abandoned the original IE 
engine in the first place...]


in OpenCL, the moment that the spec says "ULP < 3", this is a clear sign 
that the applications writers are *required* to take that into account.  
this was *clearly* built-in to the standard, right from the start, such 
that any application writer *knows* - should know - that they have a 
responsibility to take variations in implementations into account.

only where there is a de-facto "forced" standard (like Ahmdahl copying the 
IBM mainframes) will applications writers end up *potentially* writing code 
where bit-for-bit compatibility is "expected".


so i think we're ok here, i will however make a note that only Standards 
with clear ULP rules are to be included.


l.

>