Open Source Summit + Embedded Linux Conference + OSPOCon 2021

Graphics and media workloads are deeply pipelined, deeply parallel, and deeply asynchronous. Whilst GPUs are extremely performant, their amazing performance implies a relatively high latency, which is problematic when we have a handful of milliseconds to go from the application submitting its first job, to delivering pixels into eyeballs. In order to fulfill this promise, Linux graphics users make many dependent submissions asynchronously to each other. The Linux graphics subsystem has traditionally relied on 'implicit synchronisation' to make this coherent, with the kernel handling scheduling. The Vulkan API is instead built on explicit synchronisation, with userspace responsible for its own scheduling dependencies; modern hardware goes even further to give userspace exacting control - making implicit synchronisation less and less viable. Discussion has been ongoing in the graphics community to build new API and semantics which embrace these new models. This talk will demystify dma_fence, sync_file, drm_syncobj, dma_resv, and the upcoming 'userspace fences', giving a understanding into the problem space, showing how we are planning to solve these issues for upcoming hardware and APIs, and an insight into what we actually do under the hood to get cat videos into your retinas.