[Bug 89580] Implement a NIR -> vec4 pass

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=89580

--- Comment #21 from Eduardo Lima Mitev <elima at igalia.com> ---
(In reply to Jason Ekstrand from comment #20)
> (In reply to Eduardo Lima Mitev from comment #19)
> > We still have some work left improving certain implementations and getting
> > rid of some FIXMEs, but we will start cleaning and squashing patches already.
> > 
> > https://github.com/Igalia/mesa/commits/nir-vec4
> 
> Good work!  I look forward to seeing the patches.  I haven't had much chance
> to look over things for a while.  At this point, I'll probably just wait for
> patches unless you want feedback on something specific prior to sending them.
> 

Yes, it sounds good. We are preparing the patch set and hope to send it this
week already.

The only controversial part at this point is the uniform support. The way
vec4_visitor and fs_nir handle uniform the file differ significantly. Right now
we have the vec4_visitor implementation but to make it work well with nir, we 
have had to patch nir directly in some rather hackish ways. So Igo and Samuel
are evaluating a full rewrite of the uniform support to have it similar to
fs_nir layout instead. I will let them comment on it directly, if/when they see
fit

> Once we get the initial pass merged, it would be good to spend some time
> looking at how the NIR pass performs relative to the old fs_visitor code. 
> One tool we have for this is shader-db.
> 

Yes, we know shader-db and have used it already. We will definitely be looking
into performance after we get the support in. 

>
> Matt ran shader-db on your branch
> yesterday and the instruction count is at +1% vs. non-NIR in affected
> programs.  That is actually pretty fantastic.
> ...
>  It shouldn't take terribly long to figure out
> the major pain-points and get them fixed.  Talk to me, Matt, or Ken offline
> and we'll find a way to get shader-db to you.
> 

Ok great. Thank you both for the offer.

> Improving the shader-db numbers is *not* a prerequisite for getting it
> merged, just a prerequisite for turning it on by default.

Yes, it makes perfect sense.

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/intel-3d-bugs/attachments/20150617/71bf1385/attachment.html>