Re: [Qemu-devel] Re: [PATCH] Implement a virtio GPU transport

[Anthony Liguori <anthony@codemonkey.ws>]

On 10/28/2010 09:24 AM, Avi Kivity wrote:
>  On 10/28/2010 01:54 PM, Ian Molton wrote:
>>> Well, I like to review an implementation against a spec.
>>
>>
>> True, but then all that would prove is that I can write a spec to 
>> match the code.
>
> It would also allow us to check that the spec matches the 
> requirements.  Those two steps are easier than checking that the code 
> matches the requirements.

I'm extremely sceptical of any GL passthrough proposal.  There have 
literally been half a dozen over the years and they never seem to leave 
proof-of-concept phase.  My (limited) understanding is that it's a 
fundamentally hard problem that no one has adequately solved yet.

A specifically matters an awful lot less than an explanation of how the 
problem is being solved in a robust fashion such that it can be reviewed 
by people with a deeper understanding of the problem space.

Regards,

Anthony Liguori

>> The code is proof of concept. the kernel bit is pretty simple, but 
>> I'd like to get some idea of whether the rest of the code will be 
>> accepted given that theres not much point in having any one (or two) 
>> of these components exist without the other.
>
> I guess some graphics people need to be involved.
>
>>
>>> Better, but still unsatisfying. If the server is busy, the caller would
>>> block. I guess it's expected since it's called from ->fsync(). I'm not
>>> sure whether that's the best interface, perhaps aio_writev is better.
>>
>> The caller is intended to block as the host must perform GL rendering 
>> before allowing the guests process to continue.
>
> Why is that?  Can't we pipeline the process?
>
>>
>> The only real bottleneck is that processes will block trying to 
>> submit data if another process is performing rendering, but that will 
>> only be solved when the renderer is made multithreaded. The same 
>> would happen on a real GPU if it had only one queue too.
>>
>> If you look at the host code, you can see that the data is already 
>> buffered per-process, in a pretty sensible way. if the renderer 
>> itself were made a seperate thread, then this problem magically 
>> disappears (the queuing code on the host is pretty fast).
>
> Well, this is out of my area of expertise.  I don't like it, but if 
> it's acceptable to the gpu people, okay.
>
>>
>> In testing, the overhead of this was pretty small anyway. Running a 
>> few dozen glxgears and a copy of ioquake3 simultaneously on an intel 
>> video card managed the same framerate with the same CPU utilisation, 
>> both with the old code and the version I just posted. Contention 
>> during rendering just isn't much of an issue.
>