In GPU trace in NSight, I have noticed weird processing gaps during GPU processing. I have dispatched multiple draw calls on a single D3D12 direct queue, there are no resource state transitions between individual draw calls. The test runs on PC with single connected monitor, the app is in the topmost full-screen mode, VSync is enabled, Windows 10 is used, Debug layer is not enabled, SetStablePowerState is set to true… and the interruptions are still there.

I understand that GPU needs to do other work during my processing, but is there a way to eliminate the interruptions in a way, that GPU tracing data would not be polluted by these unrelated interruptions?

Have you checked if you are actually being preempted by work from another process? GPUView should show you if this is the case. If it is what's happening, I doubt you're going to be able to fully prevent preemptions over such a large period of time. 156ms is unusually long for a single dispatch.

Thanks a lot, I totally forget about GPUView and its output. I have there interference with OS command lists.

If I run 4x regular HBAOs with the same inputs and outputs, I am getting these NOPs in the middle of my command lists:

If I use only single HBAO, the result is basically the same as the first upper HBAO block (long Blur with gap inside).

Is there a way to prevent the gaps in this smaller example? (Higher priority of the command queues do not solve it)

Activity on the paging queue suggests that an app or your system overall is overcommitting video memory. Have you checked how much GPU memory you're allocating?

The allocated VRAM by the app is less than 500MB and the VRAM has plenty of unused space. I spent quite some time for possible causes, but it was without the further progress. MJP, thank you for helping and the narrowing down the problem.