Early Virgil 3D Results Show This Virtual GPU For QEMU Has Room For Improvement

[Jabberwocky]

Thanks for sharing the test results dagger 👍

[GizmoChicken]

Cannot wait to see Direct3D 11 level API support arrive in virtual drivers.

Hope this will happen in the not-too-distant future.

[airlied]

Guest info:


These tests were done on Fedora24 Beta + kraxel virgl repo - https://copr.fedorainfracloud.org/coprs/kraxel/virgl/

I'm pretty sure the numbers will be different once you pull everything from git.

Interesting I reproduced the PTS numbers yesterday, openarena running from packages runs at 60fps for me on the same box, but PTS must be setting something that triggers a slow path.

Dave.

[dagger]

Interesting I reproduced the PTS numbers yesterday, openarena running from packages runs at 60fps for me on the same box, but PTS must be setting something that triggers a slow path.

Dave.

Dave, did you have any random lockups while testing it? I experienced random lockups on some - but not all! - tests.

[airlied]

Dave, did you have any random lockups while testing it? I experienced random lockups on some - but not all! - tests.

I think the random lockups might be fixed, there was a fence waiting issue.

I've tracked down the problem with PTS openarena, the flare code in openarena does a bunch of 1x1 glReadPixels of the depth buffer on a bunch of frames, this kills virgl as each readpixel needs to go down to the hw on the host side, and pretty much syncs everything.

Not sure what I can do to mitigate it, might be worth migrating the whole surface, but for 8 pixels to be readback from a screen size surface, that seems like it might be overkill.

Dave.