[Bug 109216] 4-27% performance drop in Vulkan benchmarks

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=109216

--- Comment #25 from Jason Ekstrand <jason at jlekstrand.net> ---
(In reply to Ian Romanick from comment #24)
> There is a mostly reviewed MR that I just need to finish.  I'm really, sorry
> but since I've been back from FOSDEM, I've been very ill.  If this was such
> a pressing issue for you, you could have also reviewed that patch series
> when I sent it out and tagged you on it. :(

I'm sorry if my comments have been a bit curt and I haven't expressed my
meaning fully.  The difficulty for me isn't just the urgency (Vulkan basically
has no loop unrolling right now) but that I'm not convinced the currently
proposed fixes add up to a complete fix.  What we had before worked and it did
so very reliably as long as the loop contained no non-trivial continues.  With
8fb8ebfbb05d33, loops now frequently end up with a different form that makes
unrolling more difficult.  Because of the nature of the change to the IR that
can result from peephole select eating the continue block if statement, I'm not
actually convinced that the sum total of the proposed fixes is enough to fix it
in general.

The proper fix for this is probably to do something smarter with continues in
general.  I'm not denying the utility of 8fb8ebfbb05d33; any patch which gets >
0.1% boost in shader-db is pretty impressive.  I'm just claiming that there are
clearly problems, things aren't baked well enough, and we may need to step back
and try again.

As a side-note, it's very clear that not having SPIR-V support in shader-db is
a serious problem.  There is vkpipeline-db but I'm not sure if that's really
what we want as it records all the pipelines and doesn't do any de-duplication.

-- 
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/20190208/6675ed17/attachment-0001.html>