[Bug 101910] [BYT] ES31-CTS.functional.copy_image.non_compressed.viewclass_96_bits.rgb32f_rgb32f

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=101910

--- Comment #11 from Topi Pohjolainen <topi.pohjolainen at intel.com> ---
(In reply to Jason Ekstrand from comment #10)
> (In reply to Topi Pohjolainen from comment #9)
> > While the failing tests themself don't create RGB32F surfaces to back
> > renderbuffers (in which case they would probably fail as i965 marks RGB32F
> > as non-renderable), the surfaces nonetheless end up as render targets. As
> > the titles of all the failing tests say, tests copy two images. Image copies
> > in turn are implemented in i965 using blorp (_mesa_CopyImageSubData() ->
> > brw_blorp_copy_miptrees()) which ends up using RGB32F as GPU render target.
> 
> No, it doesn't.  It should be using R32_UINT and manually offsetting to the
> slice to be copied.  There may be bugs there though.

You are correct, it does change the format. But the tiling vs align remains.
Old uses Y-tiling with valign 2 and new uses X-tiling with valign 2. Latter
should be correct while the former shouldn't.

> 
> > With respect to PRM, isl does the right thing and chooses X-tiling with
> > valign of 2. This for some reason, however, fails. Old miptree used contrary
> > to the spec Y-tiling with valign of 2 which for some reason works (at least
> > in these tests).
> > Also as mentioned a few comments earlier, for some reason at least
> > rgb32f_rgb32f.cubemap_to_cubemap works if one uses X-tiling and valign of 4
> > (which also is against the spec).
> 
> My gut suspects something in our code to calculate offsets though I'm not
> sure.

That would explain it. I'll take a look tomorrow.

-- 
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/20170726/a56511b6/attachment.html>