random stuttering after swapchain recreation with VK_PRESENT_MODE_FIFO_KHR

[Lewa-dev]

Hi!
I've found a possible synchronisation issue causing stuttering with regards to the swapchain acquisition/presentation in the FIFO_KHR mode.
I've noticed this issue in my own vulkan renderer but after checking out the vulkan samples in this repository i can reproduce the exact same behaviour in those samples too which is why i'm posting this here in the hopes that anyone else has an idea as to what is going on. (I'm suspecting a driver issue at this point as i can't find the error anywhere.)

Specs:

• Windows 11
• RTX 5070 TI

Issue:
Consistent stuttering/choppy visual output (even though the reported framerate is stable at 60) with VK_PRESENT_MODE_FIFO_KHR in windowed mode. Can be triggered (and fixed) randomly by recreating the swapchain.

---

I've made a quick video demonstrating this issue in the vulkan samples:

https://www.youtube.com/watch?v=f9PB445C-ms

To make it quick, for debugging purposes that's the changes i've made on the codebase of the samples repository:

1. force FIFO mode in VulkanSwapchain.cpp:
   VkPresentModeKHR swapchainPresentMode = VK_PRESENT_MODE_FIFO_KHR;
2. in prepareFrame() of vulkanExamplesBase.cpp i added an output command to print the previous and currently acquired swapchain index to the debug output window (which can be also seen in the video above.)

	auto prev = currentImageIndex;
	VkResult result = swapChain.acquireNextImage(presentCompleteSemaphores[currentBuffer], currentImageIndex);
	std::string out = "SC: " + std::to_string(prev) + " > " + std::to_string(currentImageIndex)+"\n";
	OutputDebugString(out.c_str());

In the video you can see that the sample runs fine and smooth at 60 FPS with a consistent swapchain index rotation. (The debug/console output is on the right side of the screen.)
Example (the format is "SC: [previous index] > [new index]")

SC: 0 > 1
SC: 1 > 0
SC: 0 > 1
SC: 1 > 0
SC: 0 > 1
SC: 1 > 0
SC: 0 > 1
SC: 1 > 0
...

but sometimes after resizing/repositioning the window (triggering a swapchain recreation) the visual output starts to stutter visually, even though the framerate is still reported as 60FPS. (You can see this in the linked video at the timestamp 0:18 )

However if you look at the debug output, you may notice that vkAcquireNextImageKHR often acquires the same swapchain index that was used by the previous frame (possibly causing stalls in the renderpipeline?) which may lead to to the stuttery visual output.

SC: 1 > 1 <<< same index
SC: 1 > 0
SC: 0 > 0 <<< same index
SC: 0 > 1
SC: 1 > 1 <<< same index
SC: 1 > 0
SC: 0 > 0 <<< same index
SC: 0 > 1
SC: 1 > 1 <<< same index
SC: 1 > 0
SC: 0 > 0 <<< same index
SC: 0 > 1
SC: 1 > 1 <<< same index
etc...

This can be (randomly) fixed by recreating the swapchain again.

Is anyone else experiencing this issue?

[SaschaWillems]

Is this actually an issue only related to my samples or to the GPU/driver you use? If not directly related to my samples, this is prob. not the right place to ask general Vulkan related questions. Better places would be the Vulkan reddit or Discord.

[Lewa-dev]

Is this actually an issue only related to my samples or to the GPU/driver you use? If not directly related to my samples, this is prob. not the right place to ask general Vulkan related questions. Better places would be the Vulkan reddit or Discord.

Yes, this issue is related to the samples. You can see in the linked video that i tested this with your samples (specifically the deferred rendering one) and the issue described above is occuring there.

This is not about my rendering implementation.
I only mentioned this to describe how i encountered this behaviour.
(I based my renderer on your initial vulkan tutorial. Thank you greatly for that).

This bug/issue/behaviour exists/can be encountered in your samples, hence why this issue was opened.

I'm not sure if this is a driver related bug with Nvidia or if the way the swapchain synchronisation is implemented is incorrect. Hence why i asked/posted this issue here to see if others are encountering it too (specifically with the samples).
I would like to get to the bottom of this.

[Lewa-dev]

Just as a quick update in case anyone else stumbles upon this in the future:
I've found a sample app from nvidia that has a vulkan reference implementation with swapchain sync implemented:
https://github.com/nvpro-samples/vk_minimal_latest

The nvidia sample does not suffer from any kind of stuttering (at least i'm not able to reproduce it there).
Briefly looked through the code and at first galance i noticed that the nvidia sample is using timeline semaphores for the GPU-CPU sync instead of fences. Tried implementing the timeline semaphore in the samples from this repo. This helped a bit (or at least subjectively it's smoother overall, not running into the stutter quite as often) but doesn't fix it entirely. After a few swapchain recreations it still ends up acquiring the same swapchain index, leading to a choppy visual output.

I'm currently comparing/debugging both repos to see where exactly the synchronisation differences lie to see if i can find the cause of this issue.
Will post an update if I find anything meaningful.