[Intel-gfx] [PATCH 07/23] drm/i915/gt: Move the [class][inst] lookup for engines onto the GT

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-07-26 10:51:01)
> 
> On 26/07/2019 10:33, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2019-07-26 10:22:08)
> >>
> >> On 23/07/2019 19:38, Chris Wilson wrote:
> >> I read it, relatively rushed, since pressure keeps getting applied! :/
> >>
> >> There are some good parts and implementation looks okay, but I am not
> >> sure we need a tree. Nodes are bigger than pointers, management code is
> >> bigger, lookup is slower.. is it a win all things considered?
> > 
> > A big win imo. Consider that this interface is purely debug, the primary
> > interface runtime will be via gt->engines, the nodes are much smaller
> > than the sparse array.
> 
> I guess it depends. One rb_node is three pointers and can only be used 
> from a single tree. Nor does the patch replaces all sparse arrays.

There would be reasonable objection if I removed all the arrays in one
go :-p

> > I am adamant that we are not adding more sparse arrays. A 2D lookup
> > table since that matches the HW, but even then we may just end up with
> > LUT (1 extra pointer load to replace the sparse array with a compact?)
> 
> I feel it's too early for this specific patch.

It's too early? The whole point is to enable gt-centrification for the
later patches and lift gt initialisation out of GEM.
-Chris