[Intel-gfx] [PATCH] drm/i915/tgl: WaDisableGPGPUMidThreadPreemption

[Tvrtko Ursulin]

On 04/03/2020 16:02, Rafael Antognolli wrote:
> On Wed, Mar 04, 2020 at 03:31:44PM +0000, Tvrtko Ursulin wrote:
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>
>> Enable FtrPerCtxtPreemptionGranularityControl bit and select thread-
>> group as the default preemption level.
>>
>> v2:
>>   * Remove register whitelisting (Rafael, Tony).
>>
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> Cc: Michał Winiarski <michal.winiarski at intel.com>
>> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
>> Cc: piotr.zdunowski at intel.com
>> Cc: michal.mrozek at intel.com
>> Cc: Tony Ye <tony.ye at intel.com>
>> Cc: Rafael Antognolli <rafael.antognolli at intel.com>
> 
> Thanks for CC'ing me. I also saw a reply from Jason yesterday, but I
> don't see it in the list now (though my mail client a mess lately).

I saw nothing from Jason, but there was an email from you asking about 
interface descriptors and whitelisting which is why I copied you.

> But he asked whether it's possible for Media and OpenCL drivers to
> also disable mid-thread preemption through the
> INTERFACE_DESCRIPTOR_DATA, instead of from the kernel side, so we could
> try to experiment with it in the future.
> 
> Also, do you have an idea of how broken it is? Or is it just not tested
> because no driver is currently implementing it? And do you know if the
> windows 3D drivers implement it at all? I see code in the driver that
> seems to me that it's only disabled in certain cases...
> 
> To summarize, I think we should either:
>     1) Disable mid-thread preemption from the kernel and not whitelist
>     the register (just like you do in this patch); or
>     2) Not do anything at all from the kernel, and let userspace disable
>     it if needed.
> 
> I think 2) is better, if it's not an issue to the other userspace
> drivers (OpenCL and Media).

I know it is somewhat broken like in 
https://gitlab.freedesktop.org/drm/intel/issues/1293.

And I know OpenCL and Media would prefer i915 to handle it, but that's 
always the case. :) OpenCL and Media folks are on the thread so can 
comment if they are okay with handling this themselves.

Indeed a blanket ban in i915 means no one can try it out later without 
further kernel changes.

Regards,

Tvrtko