[Bug 106390] Invalid imageAtomicExchange() writes.

bugzilla-daemon at freedesktop.org bugzilla-daemon at freedesktop.org
Fri Oct 26 09:15:41 UTC 2018


https://bugs.freedesktop.org/show_bug.cgi?id=106390

Danylo <danylo.piliaiev at gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |danylo.piliaiev at gmail.com

--- Comment #11 from Danylo <danylo.piliaiev at gmail.com> ---
Hi,

Without this commit gradients are the same for all patterns:

    store  8x8 | atomic 4x4
    atomic 1x1 | atomic 8x8

I have looked at a native code of the shader with and without the commit
and the only difference is in the send command:


        Incorrect gradient (with commit)
send(8)         null<1>UW       g1<8,8,1>UD
    dp data 1 ( DC typed atomic, Surface = 1, SIMD16, mov) mlen 3 rlen 0 {
align1 1Q };

        Correct gradient (without commit)
send(8)         g20<1>UW        g1<8,8,1>UD
    dp data 1 ( DC typed atomic, Surface = 1, SIMD16, mov) mlen 3 rlen 1 {
align1 1Q };


Here we can see that optimization from that commit is working as expected since
return value of imageAtomicExchange is unused; <dest> is being null in such
case
and response length is 0.
In "Send Message" section of Volume 7: 3D-Media-GPGPU it is explicitly said
that it's correct to do this.

However something going wrong here.

One interesting observation is that writing to the red channel yields correct
gradient,
while writing to the green or blue channel does not.

I cannot find anything related to such issue in docs, so someone who has more
knowledge
of inner workings of GPU is needed here.

You can find slightly modified app to reproduce the issue in attachments.
Change the writing from the green channel to the red to see the difference.

-- 
You are receiving this mail because:
You are the assignee for the bug.
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/20181026/a147245b/attachment.html>


More information about the intel-3d-bugs mailing list