[Beignet] XOR in a data swap kernel hangs Ivy Bridge GPU

[Edward Ching]

I'm experimenting with sorting algorithms on Ivy Bridge GPU using OpenCL
and finally get to one that matches/outperforms (at least when using Intel
Windows SDK for IvyBr GPU) the combined quad CPU cores on an Ivy Bridge
using OpenMP ) which is encouraging.

However, the same OpenCL kernel when run on Beignet causes the GPU to hang
with the following message in system log:

[root at ivy4 utests]# dmesg
[17909.175965] [drm:i915_hangcheck_hung] *ERROR* Hangcheck timer elapsed...
GPU hung

I've reduced and simplifed the reproducing scenario down to the attached
test case, which when reproduces the GPU hang (ignore the SUCCESS message
on utest cmd console - instead see the hang reported in the system log with
dmesg).  The weird thing is, if the XOR in line 14 in compiler_swap.cl is
changed to either a AND or an OR, the GPU would not hang.

Is there anything illegal about the way the constructs are used? Could
anyone help shed some light on this?

Thanks,
/Ed
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/beignet/attachments/20130828/408afa4a/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-XOR-in-data-swap-kernel-causes-GPU-hang.patch
Type: application/octet-stream
Size: 2283 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/beignet/attachments/20130828/408afa4a/attachment.obj>