[Mesa-dev] NaN behavior in GLSL (was Re: [PATCH] glsl: always do sqrt(abs()) and inversesqrt(abs()))

[Ian Romanick]

On 01/12/2017 02:47 PM, Matteo Bruni wrote:
> 2017-01-12 23:41 GMT+01:00 Axel Davy <axel.davy at ens.fr>:
>>> Do you refer to the d3d9 MAD or the hardware instruction? If the
>>> former, just generating MUL and ADD separately should do the trick. In
>>> the latter case, I guess that means the "NaN switch" should also
>>> affect code generation (although I think that should be already
>>> covered by the "precise" qualifier.)
>>>
>> all radeon card released so far have special instructions for both mul and
>> mad to have 0*inf = 0.
>>
>>
>> I guess you'd need some gl extension to use them if available.
> 
> I see. Does it need to be a separate extension though? I.e. isn't it
> enough if the driver uses those special instructions when we don't
> want NaN and use the normal instructions otherwise?

Code generation is a complex matter.  Generating a separate multiply may
reduce register pressure and be better that generating a MAD.  This is
part of the reason GLSL 4.20 added the fma() built-in function.  This
provides some uniformity among what implementations can (correctly) do,
but it limits some optimization opportunities.  Freedom of the system to
optimize vs. uniformity of results is a pretty classic trade off.

> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev