Speaker
Description
XR compositors rely on timely execution of GPU workloads to apply a late-stage framebuffer update shortly before scanout. This operation often contends with other concurrent rendering workloads occupying the GPU, and its completion might get delayed past its deadline, resulting in undesired artefacts.
A widely used solution to this problem is to apply some form of GPU-level preemption. However, preemption latency still varies with the type of workload to preempt, a risk of excessive delays of the time-sensitive workload remains.
In this talk we discuss a different approach: provided we have multiple independent execution GPU cores available, can we reserve a core for the time-sensitive workload to execute independently of everything else? We present a brief case study using the ARM Mali G610 GPU, a four-core configuration, driven by the PanVK/panthor open-source Vulkan driver.
| Code of Conduct | Yes |
|---|---|
| GSoC, EVoC or Outreachy | No |
| In-person or virtual presentation | In-person |
