2/5 missing presentation frames (very weird) about vulkansamples

[ianelliottus]

Hi @Radamanthe,

I just applied your "cube.c" changes and tried it on my Ubuntu system, also with NVIDIA graphics (sorry, I don't know which card), and saw some anomalies. Normally, cube looks pretty steady for me, but it was good to slow it down, per your changes. What I saw was:

• Occasional "stutter". That is, most of the time the cube rotated in a certain direction, about 2FPS. Sometimes, it would go back-and-forth very quickly (i.e. go forwards and backwards at a must faster frame rate) for one of the frames.
• A few times, I didn't see a new frame drawn (like what you reported, only much more occasionally).

I then ported this to Windows, and ran it on a laptop with an Intel Skylake GPU. I saw no visual anomalies at all. My guess is that this is a driver bug.

[Axel512]

Yes, by default (without modification), the cube.c sample is quite steady for me too. But that's not surprising: in IMMEDIATE presentation mode, it runs at several thousands FPS! In other words, the presentation is made several times per V-Blank (maybe like ~20 or more times). So we cannot see that some of them are missing. It's a little more obvious when simply changing the presentation mode to FIFO, logically limiting the number of presented frames to 1 per V-Blank.

I made this patch on cube.c because that behavior I described happened to my current Vulkan engine (WIP) and I wanted to make sure I did not do something wrong. Your description is quite different: it seems more random (still erratic though). Mine is always exactly as I described it (and that to me is even more weird... looks like a deliberate behavior).

Anyway, I'm quite sure it's a driver bug. BTW, it's not very surprising that the GNU/Linux version of the drivers is more bugged than Windows one. So I can only guess it will be fixed at some point. I think I'll take some time soon to report it properly to NVidia directly. Thanks for sharing your experience.

Even though it definitely doesn't look like a LunarG issue, I won't close it for now. Again, I would not take it wrong if LunarG team decides to close it promptly. I just hope some more people will try it on their GNU/Linux rig and share their experience.

[damienleone]

Hi,

Thanks for reporting this issue. I do confirm that this is an NVIDIA driver bug. It will be fixed in our next 367 release.

[Axel512]

Glad to see NVIDIA catched this one. I'm eager to test the 367 release.... thanks!