Hi Marc-André,
On 7/2/2024 3:29 AM, Marc-André Lureau wrote:
> Hi
>
> On Fri, Jun 21, 2024 at 3:20 AM <dongwon.kim@intel.com
> <mailto:dongwon.kim@intel.com>> wrote:
>
> From: Dongwon Kim <dongwon.kim@intel.com <mailto:dongwon.kim@intel.com>>
>
> Introducing new virtio-gpu param, 'render_sync' when guest scanout blob
> is used (blob=true). The new param is used to specify when to start
> rendering a guest scanout frame.
>
> By default (and so far) rendering of the guest frame is started in
> the draw event to make sure guest display update is sychronized with
> host's vsync. But this method inevitably brings some extra wait because
> most of time, the draw event is not happening right after the guest
> scanout frame is flushed.
>
>
> This delay often makes the guest stuck at certain frame for too long and
> causes general performance degradation of graphic workloads on the
> guest's
> side especially when the display update rate is high. This unwanted perf
> drop can be reduced if the guest scanout frame is rendered as soon
> as it is
> flushed through 'VIRTIO_GPU_CMD_RESOURCE_FLUSH' msg. The gl display
> pipeline can be unblocked a lot earlier in this case so that guest can
> move to the next display frame right away.
>
> However, this "asynchrounous" render mode may cause some waste of
> resources
> as the guest could produce more frames than what are actually displayed
> on the host display. This is similar as running rendering apps with
> no vblank
> or vsync option. This is why this feature should stay as optional.
>
>
> Indeed, I don't see much point in doing so, it's pure waste. If you want
> to benchmark something perhaps. But then, why not simply run a benchmark
> offscreen?
Benchmark score is not the primary reason. The problem we want to avoid
is the laggy display update due to too many asynchronous plane updates
happening in the guest in certain situations like when moving SW mouse
cursor rigorously on the guest. This is because the fence (as well as
response for the resource-flush cmd) is signaled in the draw event.
Presumably, you won't get more frames displayed (perhaps even less due to extra load), so why is it laggy? Is it because the guest is doing too much buffering? Because the mouse event queue isn't being drained between scanouts?
>
>
> The param 'render_sync' is set to 'true' by default and this is in line
> with traditional way while setting it to 'false' is basically enabling
> this asynchronouse mode.
>
>
> As it stands now, the option should actually be on the display backend
> (gtk/gtk-egl), it's not implemented for other backends.
We can help to deploy this concept to other backends if needed.
Why not make it a gtk display option instead?
>
> I am not convinced this is generally useful to be an extra option though.
I initially thought this should be the default because the virtio-gpu
guest should not be hold by the host for too long in any cases. And this
new approach is comparable to the default mode with blob=false where the
guest is released as soon as the resource flush is done. Only difference
is the resource synchronization using the fence.
virgl should be blocking rendering too
Could you detail your testing setup ?