hyper-v host GPU encoding, slicing resources

[totaam]

This ticket may well end up moved to another project if it proves beneficial to more than just xpra.

Hyper-v:

• Download Windows Server 2019
• Hyper-V GPU Passthrough: An Essential Guide for Beginners lists some pretty significant limitations:
  • GPU Passthrough is available only on the highest Windows editions, such as Windows Server 2019 Datacenter
  • Clustering features such as High Availability and VM live migration are not supported
  • Save and restore with VM checkpoints are not supported_
• GPU partitioning
  • Partition and assign GPUs to a virtual machine
  • Easy-GPU-PV: GPU-PV allows you to partition your systems dedicated or integrated GPU and assign it to several Hyper-V VMs. It's the same technology that is used in WSL2, and Windows Sandbox.

Important: most passthrough and partitioning solutions don't allow guests to access to the hardware encoders, only opengl / vulcan.
Another (smaller) downside is that they require modifications to the OS configuration that cannot easily be managed using a package or script.
So, perhaps we should expose a memory mapped service to the guests so that they can hand over pixel buffers to the host which can take care of managing the encoder chip(s) and send back encoded frames.

For now, this is just an easy place to dump references and links:

• libvf.io: LibVF.IO is a vendor neutral GPU multiplexing tool driven by YAML & VFIO sounds good, but looks outdated - lots of dead links and no updates since last year...
• GVT - has now been abandoned by Intel
• GPU Virtual Machine GPU Virtual Machine for IOMMU-capable computers such as x86 and ARM looks abandoned, openmdev.io no longer exists.
• MxGPU-Virtualization - Linux kernel module for AMD SR-IOV based HW Virtualization, also abandoned
• vgpu_unlock - clearly the nvidia drivers could enable vgpu on all cards... but choose not to

I CAN (edit) find the ivshmem equivalent for hyper-v: see #3666 and VMBUS: VMbus enables efficient, low-latency communication between the host and the guest VM

---

ivshmem links for the Linux implementation: #3810

[totaam]

For OpenGL acceleration of WSL2 guests, see https://github.com/Xpra-org/xpra/blob/master/docs/Usage/WSL.md

[totaam]

As for GPU encoding, AMD AMF includes some code samples:
https://github.com/GPUOpen-LibrariesAndSDKs/AMF/blob/master/amf/public/samples/
It's not yet clear to me if we can use DirectX11 contexts when running in Windows Server context.
https://github.com/GPUOpen-LibrariesAndSDKs/AMF/blob/master/amf/public/samples/SamplesC/SimpleEncoderC/SimpleEncoderC.c : this sample encodes NV12 frames using AMF Encoder and writes them to H.264 elementary stream

[totaam]

Linux Kernel vmbus documentation: https://docs.kernel.org/virt/hyperv/vmbus.html

Microsoft Device Virtualization API: https://learn.microsoft.com/en-us/virtualization/api/hcs/reference/hdv/devicevirtualization

[totaam]

I have added some AMF findings here: #4418 (comment)
Most disappointing is that all encoders use YUV subsampling: All codecs are 4:2:0

On the plus side, apart from some driver compatibility issues, AMD GPUs can be used with Easy-GPU-PV