[Beignet] Haswell issues

[Zhigang Gong]

> -----Original Message-----
> From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of
> Rebecca N. Palmer
> Sent: Monday, April 27, 2015 6:10 AM
> To: beignet at lists.freedesktop.org
> Subject: Re: [Beignet] Haswell issues
> 
> > I will raise this issue to Linux i915 kernel team. And hope we can get
> > a clean fix in Linux kernel side finally.
> 
> I agree that's a good idea, but given how long it is likely to take, I'd like to at
> least give a clear user-friendly error in the meantime.
Daniel already submitted a patchset to address part of this issue. The only remained
issue is the atomic in L3. Kernel team will investigate that issue, and if possible that
feature will be enabled by default as well. I am thinking about to revert the "atomic in L3"
patch in beignet and add a new known issue to describe this performance issue.

Then we will get a working unmodified kernel for all HSW machines soon!

> 
> -Is it likely that "drm_intel_gem_bo_context_exec() failed: Invalid argument"
> (since 9e8874c) affects _all_ Haswell hardware?  If so, I'd consider this an
> improvement over the previous "some work (as long as you don't use __local),
> some silently do nothing", and put the i915_cmd_parser workaround in the
> error message.
Disable __local support is somehow not a good idea for me. As it's an important and mandatory feature,
doing that will break most of the existing opencl applications. The user will simply get a broken beignet.icd too. 

> 
> -What happens when you do try to use __local (without
> https://01.org/zh/beignet/downloads/linux-kernel-patch-hsw-support): an
> explicit error, or wrong results?
> 
> -It was suggested some time ago that i915.enable_ppgtt=2 might help; does it?
> (This kernel parameter can only be set at boot time, and is an "unsafe
> parameter" i.e. "may crash your system", but would be easier than recompiling
> Linux.)

After revert the beignet's atomic in l3 patch, the i915.enable_ppgtt=2 should work for HSW platform.
CC to Mengmeng, could you help to identify the first kernel version which this enable_ppgtt=2 works for beignet?

Based on all the above facts, I think a better way to handle this issue for HSW platform:

1.  Revert the atomic in L3 patch in beignet. Add a known performance issue in the wiki.
2.  Check the whether the kernel version is newer than the first working one(enable_ppgtt=2),  if so then further check whether i915.enable_ppgtt is 2. If not, don't report any OCL device/platform.
3.  After the Daniel's patchset get merged, we can check the kernel version and remove the kernel parameter checking in step 2 if we get a good new kernel installed.

Any comments?

Thanks,
Zhigang Gong.

> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet