[Mesa-dev] Find common patterns in real GLSL shaders

[Ian Romanick]

On 08/03/2014 01:59 PM, Monis Javed wrote:
> 
> Hello
> I am IVth year B.Tech (Computer Engg.) Student at Jamia Millia Islamia
> I am currently part of the *GSoC 2014* program under *CERN SFT
> (http://www.google-melange.com/gsoc/proposal/public/google/gsoc2014/monisjaved/5668600916475904)*
> I was looking at the projects list for a suitable project for my minor
> project in college and I found the 
> *common patterns in real GLSL shader*s project which got my interest)
> I also looked at the EVoC program and was immediately interested due to
> the flexibility in the dates 
> I think I can undertake this project as a part of my minor project in
> college giving me more than sufficient time to work on it 
> Can you please give a little more details on the project 

I proposed the project, so I guess I should speak up. :)

Basically, the idea was to look at the output of MESA_GLSL=dump from a
large number of real-world shaders to look for long sequences of IR that
occur in many shaders.  With this information, we could:

A. Re-write the GLSL IR to be more optimal.

and / or

B. Modify code generators to detect these sequences and generate optimal
code.

This is roughly the idea behind a "peephole superoptimizer."  See
section 3.1 of http://theory.stanford.edu/~sbansal/superoptimizer.html
for more details.

Rather than having software automatically generate the optimized
sequences, as is typical of a superoptimizer, I figured we'd have some
smart humans optimize the few most common sequences by hand.

> -- 
> Regards
> Moonis Javed
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev