Speaker
Description
Last two years of XDC, we introduce interim pinning buffer method for blob memory on Xen that's to enable virtio native context solution on Xen hypervisor, and ROCm support on virtio-gpu for compute machine learning applications. However, it's still not enough for automotive use scenario. So we introduce more advanced features like dynamic address mapping instead of pinning buffer which is specific on Xen for blob memory as formal solution, adding pps producer on amdgpu and renderstage with radv for perfetto to monitor GPU performance counter and trace events, GPU recovery virtualization mechanism to add robust context in weston compositor and qemu for virtio-gpu, and multiple contexts support on AMD KFD for ROCm native context to resolve one context per one process manner that can support multi-process in ROCm at guest VM, etc. In this talk, Ray will continue giving the introduction for latest GPU Para-Virtualization updates and future work on Xen.
| In-person or virtual presentation | In-person |
|---|---|
| Code of Conduct | Yes |
