Can X11 return an image with an NVIDIA GPU memory address?

Aaron Plattner aplattner at nvidia.com
Mon Aug 12 20:58:50 UTC 2019


On 8/3/19 10:09 AM, Suhail Doshi wrote:
> Hi there,
> 
> My goal is to try get the frame of a desktop to do low-latency remote 
> desktop. I am interested in using X11 as the window manager.
> 
> When I run nvidia-smi, I noticed that X11 is a process that interacts 
> the GPU:
> +-----------------------------------------------------------------------------+
> | Processes:                                                       GPU 
> Memory |
> |  GPU       PID   Type   Process name                             Usage 
>       |
> |=============================================================================|
> |    0      3255      G   /usr/lib/xorg/Xorg                           
>   57MiB |
> |    0      3286      G   /usr/bin/gnome-shell                         
>   81MiB |
> +-----------------------------------------------------------------------------+
> 
> My question is: Is there a way to get an image of the desktop that 
> returns a pointer that's GPU memory? For example, in Windows 10, there's 
> an API called the Desktop Duplication API which will allows you to do 
> this. Then, it lets you copy the data in that GPU memory block to 
> another one such that you can encode the frame with H264, for example. I 
> am using NVIDIA GPUs and utilizing their NVENC SDK.

The video memory used for the desktop is not directly accessible by the CPU.

> I am looking for an equivalent.
> 
> Right now, it appears that XShmGetImage returns the frame of the desktop 
> in host memory versus device memory which increases latency by 5-10ms. 
> For our purposes, a reduction of 5ms is meaningful.

CPU-mapped device memory is likely much slower to access than you think, 
so XShmGetImage is probably much faster than the CPU-mapped approach 
you're thinking of.

What it sounds like you're looking for is the NVIDIA Capture SDK, which 
provides an API that applications can use to capture frames into video 
memory and do on-GPU encoding to H.264 before streaming to system memory.

You can find more information at
https://developer.nvidia.com/capture-sdk

Sincerely,
Aaron

> Should I use kernel-level APIs like interacting with Direct Rendering 
> Manager and Kernel Mode Setting to accomplish this instead?
> 
> Just looking for options. Thanks!
> 
> Suhail
> 
> --
> 
> Founder


More information about the xorg mailing list