does cuda/nvidia support kvm

Gaming and Visualization Technologies Video Processing & Optical Flow
1

Hello there,
I installed NVIDIA driver and CUDA on CentOS 7, compiled ffmpeg to do hardware codec, I added video hardware coding code to libspice-server, so that KVM loaded libspice-server.so when the virtual machine played video. I hope video playback has the ability to accelerate hardware encoding, but now libspice-server.so called by kvm always prompts CUDA initialization error(cuInit return err != CUDA_SUCCESS)

I also compiled ffmpeg and CUDA examples, all example applications worked correctly, so is CUDA/NVIDIA does not support KVM or any other settings I need to do?

[AVHWDeviceContext @ 0x55db9479d0c0] Loaded lib: libcuda.so.1
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuInit
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuDeviceGetCount
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuDeviceGet
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuDeviceGetName
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuDeviceComputeCapability
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuCtxCreate_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuCtxSetLimit
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuCtxPushCurrent_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuCtxPopCurrent_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuCtxDestroy_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuMemAlloc_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuMemFree_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuMemcpy2D_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuMemcpy2DAsync_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGetErrorName
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGetErrorString
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuStreamCreate
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuStreamQuery
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuStreamSynchronize
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuStreamDestroy_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuStreamAddCallback
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuEventCreate
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuEventDestroy_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuEventSynchronize
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuEventQuery
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuEventRecord
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGLGetDevices_v2
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGraphicsGLRegisterImage
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGraphicsUnregisterResource
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGraphicsMapResources
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGraphicsUnmapResources
[AVHWDeviceContext @ 0x55db9479d0c0] Loaded sym: cuGraphicsSubResourceGetMappedArray
[AVHWDeviceContext @ 0x55db9479d0c0] Could not initialize the CUDA driver API
Failed to create a CUDA device. Error code: Unknown error occurred
2

If I decoded your question correctly you are missing access to nvDEC inside virtual/guest machine.

If you are looking for virtual/guest accelerate video decoding/stream-forwarding with KVM/qemu/spice I suppose that is not possible (see https://lists.freedesktop.org/archives/spice-devel/2016-August/031140.html).

If you have CUDA problems with KVM in virtualization host try also CUDA forums https://devtalk.nvidia.com/default/board/58/cuda-setup-and-installation/ or https://devtalk.nvidia.com/default/board/57/cuda-programming-and-performance/.

3

Thank you very much for your answer.

We do not want to use NIVIDA’s GPU virtualization. Instead, KVM is treated as a normal process. KVM enables GPU computing by loading libspice-server.so. The ffmpeg library with hardware-accelerated ability is compiled to libspice-server.so. So, how can we make KVM use GPU hardware acceleration in this situation? Whether the KVM process can directly enable hardware acceleration?

4

Hi lzhujian,

Are you trying to run on a virtual GPU set up?

If yes, as mentioned in earlier comment, only some Virtual GPU profiles/cards allows CUDA in virtual machine (see https://docs.nvidia.com/grid/latest/grid-vgpu-user-guide/index.html#features-grid-vgpu). Please check your’s is one of these.

Thanks,
Ryan Park

5

Hi rypark, thanks for your answer.

Our graphics card model is tesla p4, but we are not using GPU virtualization. Instead, KVM directly calls CUDA via libspice-server.so (actually the ffmpeg library with hardware encoding capabilities) to obtain hardware encoding capabilities.

6

Hi lzhujian,

Could you provide us a sample code that reproduces the issue?

Thanks,
Ryan Park