[Mesa-dev] NIR, SCons, and Gallium

[Ilia Mirkin]

On Mon, Jan 11, 2016 at 7:11 PM, Ian Romanick <idr at freedesktop.org> wrote:
> On 01/11/2016 06:21 AM, Jose Fonseca wrote:
>> FWIW, I updated SCons to build NIR, both with GCC and MSVC:
>>
>>   http://cgit.freedesktop.org/~jrfonseca/mesa/log/?h=scons-nir
>>
>> It was actually simpler than I anticipated.
>>
>> But I hit a wall -- there's actually no way to get NIR used with
>> softpipe/llvmpipe, not even as an intermediate IR somewhere between GLSL
>> IR and TGSI, is there?
>>
>> Without this I can't actually test it.  And I'm afraid the scons
>> integration will rot again unless it is used.
>>
>>
>> I know other gallium drivers already use NIR, but IIUC, they use NIR
>> internally, ie., TGSI -> NIR-> HW.
>>
>>
>> So what is exactly the long term plan for NIR in Mesa general, and
>> Gallium in particular?
>> - replace GLSL IR completely?
>
> Not likely.  GLSL IR has a lot of GLSL-specific semantics that aren't
> really appropriate for an IR at NIR's level.
>
>> - use NIR as intermediate IR betweem GLSL IR and TGSI, and run
>> optimizations in there?
>> - use NIR instead of TGSI at the gallium interface?
>
> I think folks have expressed a (slight?) preference for the latter
> rather than the former.  However, the former may provide an easier
> stepping stone.
>
> I know that Ilia has expressed some concerns about how NIR will effect
> nouveau.  I don't recall hearing an opinion expressed by Marek or any of
> the other AMD guys, but that probably just means that I wasn't paying
> enough attention.

To add a bit more precision -- my main concern is performance --
throwing an extra FOO IR <-> BAR IR stage seems, intuitively, like
it'll make things slower. Perhaps that could be mitigated by directly
accepting some other IR in nouveau, but... seems like effort would be
better spent towards accepting something like SPIR or SPIR-V than some
other intermediate IR. I just don't see any benefit to NIR for a
backend that already has an optimizing compiler.

Having had some experience already dealing with NIR as a total newbie
(through some freedreno work since Rob adopted NIR as the only input
IR), I'm less-than-encouraged... at this point I very much prefer to
deal with TGSI or (gasp!) GLSL IR. Both of those are relatively easy
to deal with, even if imperfect. (Not to mention the fact that TGSI
can be used as a transport IR, which AFAIK NIR cannot.) An argument
could be made that it's because I have more experience with them, and
it's probably true for GLSL IR, which I distinctly remember being
confused by. But it was all followable code. And TGSI was intuitive
from day one.

  -ilia