Re: [PATCH v2 1/7] drm/i915/gen8: Add infrastructure to initialize WA batch buffers

[Dave Gordon <david.s.gordon@intel.com>]

On 02/06/15 19:36, Siluvery, Arun wrote:
> On 01/06/2015 11:22, Daniel, Thomas wrote:
>>>
>>> Indeed, allocating an extra scratch page in the context would simplify
>>> vma/mm management. A trick might be to allocate the scratch page at the
>>> start, then offset the lrc regs etc - that would then be consistent
>>> amongst gen and be easy enough to extend if we need more per-context
>>> scratch space in future.
>>> -Chris
>>
>> Yes, I think we already have another use for more per-context space at
>> the start.  The GuC is planning to do this.  Arun, you probably should
>> work with Alex Dai and Dave Gordon to avoid conflicts here.
>>
>> Thomas.
>>
> 
> Thanks for the heads-up Thomas.
> I have discussed with Dave and agreed to share this page;
> GuC probably doesn't need whole page so first half is reserved for it's
> use and second half is used for WA.
> 
> I have modified my patches to use context page for applying these WA and
> don't see any issues.
> 
> During the discussions Dave proposed another approach. Even though these
> WA are called per context they are only initialized once and not changed
> afterwards, same set of WA are applied for each context so instead of
> adding them in each context, does it make sense to create a separate
> page and share across all contexts? but of course GuC will anyway add a
> new page to context so I might as well share that page.
> 
> Chris/Dave, do you see any problems with sharing page with GuC or you
> prefer to allocate a separate page for these WA and share across all
> contexts? Please give your comments.
> 
> regards
> Arun

I think we have to consider which is more future-proof i.e. which is
least likely:
(1) the area shared with the GuC grows (definitely still in flux), or
(2) workarounds need to be context-specific (possible, but unlikely)

So I'd prefer a single area set up just once to contain the pre- and
post-context restore workaround batches. If necessary, the one area
could contain multiple batches at different offsets, so we could point
different contexts at different (shared) batches as required. I think
they're unlikely to actually need per-context customisation[*], but
there might be a need for different workarounds according to workload
type or privilege level or some other criterion ... ?

.Dave.

[*] unless they need per-context memory addresses coded into them?

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx