[Bug 89580] Implement a NIR -> vec4 pass

[bugzilla-daemon at freedesktop.org]

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

--- Comment #34 from Eduardo Lima Mitev <elima at igalia.com> ---
A little update:

The backend is already upstream. Jason reviewed the series and after some
polishing, he pushed it upstream himself (thanks!).

There were some regressions on gen4, gen5 and gen8 which we (the Igalia team)
couldn't detect due to lack of available HW. However Jason sent a mini-series
fixing all these issues
<http://lists.freedesktop.org/archives/mesa-dev/2015-August/090617.html>. With
these patches, the backend can be considered to have 0 regressions on all Intel
gens.

Now we have started work on improving quality of emitted code, so that we don't
perform worse than vec4_visitor. This is a requirement to activate the backend
by default.

I think we can keep this bug open to track this work too, but I'm ok with
closing this and opening a new one just for the optimization work.

For reference, this is more or less where we stand today comparing the NIR-vec4
path against non-NIR, on HSW:

total instructions in shared programs: 1853749 -> 1838155 (-0.84%)
instructions in affected programs:     1707212 -> 1691618 (-0.91%)
helped:                                5310
HURT:                                  13531
GAINED:                                0
LOST:                                  0

In the coming days week, we are going to work on identifying optimization
opportunities and implementing them. We will create a series of patches, and at
some point submit them to mesa-dev for review. If you want to follow progress
or do early testing, we will be pushing to this branch:
https://github.com/Igalia/mesa/tree/nir-vec4-quality

cheers

-- 
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/20150810/02405956/attachment.html>