Speakers
Description
Last year, we introduced how to enable graphic virtio native context and GPU passthrough based on Xen hypervisor[1]. However, graphic usage is not enough for modern requirement of GPU virtualization. Because AI and compute are more and more popular recently, and the community also wants to have a way to support compute virtualization on general GPU devices.
AMD provided an open source stack named ROCm[2] to support machine learning, coupled with being inspired by the idea of virtio native context. Then we dig out a solution to use Thunk (libhsakmt) API forwarding [3] via virtio-gpu to implement ROCm support on virtual machine and add virtio-gpu support into ROCm runtime as another GPU backend. With this way, we can have a prototype to implement GPU compute virtualization with virtio.
OpenCL is the first step that we step in ROCm virtualization with virtio-gpu. This talk will provide an overview of AMDs experiences enabling compute virtualization with OpenCL over ROCm native context approach.
Reference:
[1] https://indico.freedesktop.org/event/4/contributions/216/
[2] https://github.com/ROCm
[3] https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1370
In-person or virtual presentation | In-person |
---|---|
Code of Conduct | Yes |