[Bug 91926] [SKL bisected] texsubimage pbo intermittent failures

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=91926

--- Comment #36 from Ben Widawsky <ben at bwidawsk.net> ---
On Wed, Oct 21, 2015 at 01:21:12PM +0000, bugzilla-daemon at freedesktop.org
wrote:
> https://bugs.freedesktop.org/show_bug.cgi?id=91926
> 
> --- Comment #35 from Topi Pohjolainen <topi.pohjolainen at intel.com> ---
> (In reply to Topi Pohjolainen from comment #34)
> > (In reply to Topi Pohjolainen from comment #13)
> > > I was able to narrow down the combination of _mesa_update_state() and
> > > brw_emit_mi_flush(). I got it simplified to a call _mesa_update_texture()
> > > followed by part of the logic in brw_context.c:intel_update_state() - namely
> > > calling brw_render_cache_set_check_flush() at least once. That call in turn
> > > is simple check for existence of the underlying buffer object of the texture
> > > and then a call to brw_emit_mi_flush().
> > > 
> > > So, after more simplification, I found out that if I'm calling
> > > brw_emit_mi_flush() _twice_ after the call for _mesa_meta_pbo_TexSubImage()
> > > in intelTexSubImage() I'm not able to reproduce the error. Recall that
> > > calling brw_emit_mi_flush() isn't enough.
> > > 
> > > This is turn means that I'm mostly likely in square one - I'm simple
> > > changing the dynamics enough to hide the real bug.
> > 
> > I have a little more to add here. I started reading bspec again regarding
> > flushing and related pipe control options. I added another
> > brw_emit_pipe_control_flush() into brw_emit_mi_flush() in order to
> > experiment with individual bits of the pipe control command. (All that
> > brw_emit_mi_flush() does is to emit pipe-control commands).
> > 
> > Neither on IVB or on SKL I cannot see any failures if I just submit another
> > pipe-control command with only a single bit set, namely
> > PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE. Any other bit didn't seem to make any
> > difference.
> 
> And here it should be noted that the preceding (and already existing
> pipe-control command) in brw_emit_mi_flush() contains this bit among others.
> 
> Also I noticed bspec telling me that on SKL TEXTURE_CACHE_INVALIDATE must
> always be accompanied with CS_STALL bit. I didn't have it, but the command
> still made a difference.
> 

Invalidating the texture cache is a pretty large hammer, particularly if you're
only updating a small region of your texture. (CS stalls are used all over and
probably won't make too much difference). Have you already tried throwing in a
bunch of clflushes or a wbinvd before the upload?

-- 
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/20151021/5ff24dc5/attachment.html>