[Mesa-dev] Proposal: MESA_debug_operations

[Rob Clark]

On Sun, Apr 14, 2019 at 2:22 PM Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>
> Hi all,
>
> I've put together a GLSL extension proposal at:
>
> https://people.freedesktop.org/~imirkin/MESA_debug_operations.spec

I could see some cases where something like this would be useful..

BR,
-R

>
> The situation this is trying to alleviate is the fact that it's an
> _enormous_ pain to actually execute shaders and collect results on any
> particular hardware without a surrounding driver to handle all the
> mundane bits. It's much easier to stick it into an already working
> driver, and basically pass a token to the shader compiler to do what
> it will.
>
> We've had numerous occasions in nouveau development whereby some
> instruction has a flag that we don't quite understand (or even a whole
> instruction). We could do one-off hacks to make the shader compiler
> "test it out", but it's tedious. This enables permanent infrastructure
> in mesa to pass such tokens through with enough parameters to be able
> to test anything. (Although I doubt it would ever be submitted for
> inclusion in the official KHR db.)
>
> Since this relies on careful interactions with shader compiler, what
> the debug ops really mean would need to be either baked in, or
> provided via env vars. The idea isn't to have anything checked in that
> would actually use this ext, but to enable driver developers with a
> simpler testing methodology. But you could have a glsl model of how an
> operation works and record any differences. Or you could just dump the
> results into a ssbo in a compute shader for offline analysis.
>
> It introduces an ALU-style op, texture, image, and memory. I think
> that covers it, except for interpolation, but I felt tackling that one
> would be unnecessarily complicated. And interpolation frequently
> involves bits outside of the shader as well in any case.
>
> I don't have an implementation for this yet, I wanted first to gather
> some feedback (and most importantly, any opposition) before doing any
> actual work on this.
>
> Cheers,
>
>   -ilia
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev