[Mesa-dev] [Bug 100613] Regression in Mesa 17 on s390x (zSystems)

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=100613

--- Comment #10 from Roland Scheidegger <sroland at vmware.com> ---
(In reply to Ray Strode from comment #8)
> With attachment 130980 [details] [review], things are mostly working now. 
> piglit found a couple regressions:
> 
> https://people.freedesktop.org/~halfline/results-summary-shift-change/
> regressions.html

Sure looks like all the same problem - the 3x16bit formats
(PIPE_FORMAT_R16G16G16_X) don't work.
Might be because they are handled differently by lp_build_gather: 2x16 and 4x16
are loaded as exactly that. Whereas 3x16 will load as 1x48 plus zext to 64bit
(vector_justify would add a shift).
(The same isn't true for 3x32, which is loaded as 3x32 in lp_build_gather()
then the vector padded.)
I'm not sure maybe loading as 3x16 plus vector padding, analogous to what is
done for 3x32 would work? The code for doing so should work, it's purely not
done this way because codegen on x86 produces a (albeit correct) mess - but it
might fare better on ppc... Otherwise some vector_justify in the right cases
might fix this too...

That said, the lp_build_gather() calls are supposed to give the same results
regardless if things are fetched as vectors or scalars - this is meant purely
for optimization purposes. And I've got some feeling that's not really the case
on BE (if it isn't, that code is very very likely to break again in the
future).

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20170423/824b7ff6/attachment.html>