AMD is working to provide a complete GPU para-virtualization solution on Xen. It includes both on graphics virtualization and compute virtualization. We are using a native context way to implement the API forwarding (drm and thunk) for graphics and compute (AMD ROCm). But still have some challenges that would like to discuss in the session, for example, how to support shared virtual memory (unified cpu address and gpu address space) in xen guest, how to support user pointer (cpu buffer to be mapped for gpu address) in xen, virtio gpu on PVH guest, etc.