This is the same as the last posting, but hopefully now all the ground work is inplace that we get through CI cleanly (gem_eio is the bane of my existence). The goal of the patchset is to eliminate high latencies caused by execlists_submission_tasklet being deferred to ksoftirqd and scheduled as a normal task. This is achieved by writing to the ELSP directly from the process submitting the new work. From irq context, we continue to use the tasklet, so that we don't starve other users of the system, as here we expect to keep the hw supplied with a pair of contexts so that it has a request to work on as we queue the next, a slight delay in our processing of the irq should not result in a GPU stall. The consequence of rearranging the work to be split between different contexts is that we end up with even more under an irqsoff spinlock. To address this raise upon impact on the rest of the system (as well as worrying about contention between ourselves), we first make sure that we eliminate all unnecessary mmio and especially the forcewake. Later we apply some more microoptimisations -- that aren't so micro at the microsecond latency level! -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
