We are developing a Vulkan application targeting Windows and ran into a very strange driver issue where performance slowly degrades (in one test, every 93rd frame took WAY too long, and as the app kept running, it kept taking longer and longer) and eventually the application hangs forever in the driver.
The issue ended up being that we called
vkResetCommandPool(device, commandPool, VK_COMMAND_POOL_RESET_RELEASE_RESOURCES_BIT);
every frame. If this was replaced with
vkResetCommandPool(device, commandPool, 0);
then the issue stopped happening. This is a good change and it should have always been 0, but I still think it's pretty bad that the driver ends up killing performance and freezing in an spec-compliant vulkan application.
System Information:
Windows 10: Version 10.0.19045 Build 19045
Intel Graphics Driver: Version 31.0.101.4575
CPU/GPU: i9-11900H
Repro Insructions:
Download attached vktext.cpp file, which is based on https://github.com/Overv/VulkanTutorial/blob/main/code/15_hello_triangle.cpp with a single-line change on line 658, as well as removing vsync which makes it reproduce faster.
You will need to setup GLFW and Vulkan SDK to be able to build it. Instructions at https://vulkan-tutorial.com/Development_environment
Run the app for 10 minutes or so, with some way to monitor Frametimes (I use MSI afterburner)
After a while, frame times look like this:
Things of note:
1. the spacing between the lag spikes is always the same
2. size of spikes increases over time
If you look at line 656, you can see the original reset command in there. If you swap which line is active, then it will work smoothly.
Obviously, I expect some performance hit from reallocating the underlying command buffer memory every frame, however I would not expect that cost to increase to hundreds of milliseconds in an extremely simple application.
If you continue running the program for a few hours, it will eventually totally hang (I don't remember which call it hangs on and don't really want to run it again, but if that would be helpful I'm happy to)
Link Copied
