Started out as what intended to be a quick look at how to extend the
fences to support the priority inheritance ended up completing the final
few steps to implement the most trivial priority based scheduler. A lot
of the policy decisions have been left out (such as e.g. only submitting
one request from each context each interrupt) so that it is not very
fine-grained, nor does it implement time-slicing and fair load
balancing, and continues to live in blissful ignorance of the challenge
of preemption. It is also execlists only. It purely serves as fodder to
take the next steps towards a real scheduler, and in the meantime offers
one benefit to current desktops, giving uber priority to the rendering
of the pageflip.
This adds the dependency tracking of the requests required to do
inflight reordering (as distinct from the dependency tracking required
to decide when we can submit the request to hardware) and implements a
priority sorted rbtree of the ready-to-execute requests to be submitted
on the next context-switch.
-Chris
___
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
