[Bug 94477] Too many temporary expressions in shader freeze glLinkProgram

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=94477

--- Comment #9 from Tapani Pälli <lemody at gmail.com> ---
(In reply to Matt Turner from comment #8)
> (In reply to Tapani Pälli from comment #6)
> > Created attachment 126170 [details] [review] [review]
> > fix
> > 
> > This should result in same effect as the first experimentation, I will run
> > testing on this and if it works fine then I'll send it to mesa-dev mailing
> > list.
> 
> That's pretty clever, but unfortunately it cannot work. Cases of dividing by
> zero, inf, nan, etc cannot be handled. We don't do this optimization in
> opt_algebraic for exactly that reason.

right .. I knew there to be something missing as it was too good to be true :)

> On top of that, the purpose of the test is to check whether you can compile
> a huge expression, and the one it uses is hilariously unrepresentative and
> prone to abuse by hacks like this one :)
> 
> The test has uncovered real flaws in our compiler: GLSL IR's copy
> propagation runs in O(n^2), and the channel_expressions pass creates a huge
> number of expressions.
> 
> Ken and I have spent time recently trying to handle the shader-db
> regressions by disabling channel_expressions. It's slow going and other
> tasks keep taking priority.
> 
> The most direct way forward then, is to improve copy propagate's run time
> like I described in comment 3. Maybe you want to take a stab at replacing
> the linked lists in the pass with hash tables? I spent some time on it...
> maybe you'll have better luck. :)

I made a wrapper class for 'acp' that internally uses hash table but provides
some methods of exec_list (just to easily experiment with this). I'm seeing
some performance gains but not nearly as big as with my hack. I'll try to do
the same for 'propagation_elements' and see if total effect would be big enough
to continue with this exercise.

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-3d-bugs/attachments/20160905/c8838c60/attachment.html>