Quoting Chris Wilson (2019-01-30 16:08:59) > Quoting Tvrtko Ursulin (2019-01-30 15:57:06) > > Under what scenarios we end up not freeing active nodes sufficiently? It > > would have to be some user which keeps many contexts around, having only > > used them once? > > Never in a typical scenario :) Almost every allocation is served from > the last_request slot, you need to have a pair of concurrent references > to a vma for it to activate, and that takes multiple engines reading the > same object within a context/ppgtt. That being said, that applies to vma. HW semaphores is an example where the cacheline could be shared far and wide with uncertain lifetimes. Context barriers in contrast just have a few different timelines and idle almost immediately. -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
