using vkWaitForPresentKHR

[FrancoisBerard]

Hello,

I am a longtime OSG user, considering the switch to VSG for about a week, i.e. almost no experience with VSG/Vulkan.

I need to synchronise my program running on the CPU to exactly when the current frame is being presented on the display (as accurately as possible). I have 2 main purposes for this: reducing latency (i.e. wait as much as possible before sampling sensors and rendering) and controlling stereo shutter glasses "by hand" (i.e. I need to know when to shut the left eye and let the right eye be exposed when the right eye image is about to be presented).

I've found the "VK_GOOGLE_display_timing" vulkan extension but it is not available on any of my vulkan implementations.

So I am planning on using the "VK_KHR_present_wait" vulkan extensions (available on both MacOS/M4 and Linux/NVidia), but I am having troubles using them within the VSG workflow.

typedef VkResult (VKAPI_PTR *PFN_vkWaitForPresentKHR)(VkDevice device, VkSwapchainKHR swapchain, uint64_t presentId, uint64_t timeout);

PFN_vkWaitForPresentKHR needs a presentId which I don't know how to get after viewer->present(). I understand that I may have to replace viewer->present() with my own call to vkQueuePresentKHR, but this requires a VkQueue; which I don't know how to get.

Of course, I may be completely off the mark. Any help will be greatly appreciated.

[robertosfield]

I am not familiar with the extension a quick check on my Kubuntu 22.04 + Geforece 2060 shows that it's supported:

$ vsgdeviceselection --extensions | grep VK_KHR_present_wait
extensionName = VK_KHR_present_wait, spec = 1

A quick search online for details on the vkWaitForPresentKHR extension:

https://docs.vulkan.org/refpages/latest/refpages/source/vkWaitForPresentKHR.html

As it's extension it'll need to be enabled why setting up the vsg/vkInstance and vsg/vkDevice, then have the functions called in the appropriate way.

A subclass from vsg::Viewer::present() to re-implement the present would likely be the way you'd add the required extra calls, or get the details from the viewer and then call vkWaitForPresent there after.

It may that this functionality would be best rolled into the core VSG in some fashion.

What hardware are you using?

[FrancoisBerard]

MacOS 15.7.2 (M4 Mac integrated graphics), Lubuntu 24.04.3 (NVIDIA RTX 5000 Ada).
I am also working on a Lubuntu 22.04.5 (AMD Radeon Pro 7600) but it does not seem to have VK_KHR_present_wait; which may be due to 22.04 vs 24.04... or AMD vs NVIDIA.

I'll be looking into subclassing vsg::Viewer and/or vsg::Presentation.

Thanks!

[robertosfield]

What display hardware are you using?

[FrancoisBerard]

As displays, I'm using a fast/low input lag gaming monitor and a fast/low input lag projector (Barco), both doing stereo.

I think I made some progress: I subclassed vsg::Viewer and replaced the call to viewer->present() to my new method viewer->present_and_wait() (see below)

If I only enable the extension on the device (windowTraits->deviceExtensionNames.push_back("VK_KHR_present_wait");), then the call on waitForPresentKHR returns 2 (VK_TIMEOUT).
If I try to also enable the extension on the instance (windowTraits->instanceExtensionNames.push_back("VK_KHR_present_wait");), then the instance creation fails with VK_ERROR_EXTENSION_NOT_PRESENT.

but:

fberard% vulkaninfo | grep present_wait
	VK_KHR_present_wait                         : extension revision 1
	VK_KHR_present_wait2                        : extension revision 1

void gilvk::viewer_t::present_and_wait()
{
	static PFN_vkWaitForPresentKHR waitForPresentKHR = nullptr;

	if (presentations.size() != 1) {
		printf("ERROR: gilvk::viewer_t::present_and_wait has presentations.size=%d, should be 1.\n", int(presentations.size()));
		return;
	}

	vsg::Presentation& presentation = *(presentations.front());
	if (presentation.windows.size() != 1) {
		printf("ERROR: gilvk::viewer_t::present_and_wait has presentation.windows.size=%d, should be 1.\n", int(presentation.windows.size()));
		return;
	}

	if (waitForPresentKHR == nullptr) {
		waitForPresentKHR = reinterpret_cast<PFN_vkWaitForPresentKHR>(vkGetDeviceProcAddr(presentation.windows.front()->getDevice()->vk(), "vkWaitForPresentKHR"));
		if (waitForPresentKHR == nullptr) {
			printf("ERROR: gilvk::viewer_t::present_and_wait does not have PFN_vkWaitForPresentKHR\n");
			return;
		}
	}

	std::vector<VkSemaphore> vk_semaphores;
	for (auto& semaphore : presentation.waitSemaphores)
		vk_semaphores.emplace_back(*(semaphore));

	std::vector<VkSwapchainKHR> vk_swapchains;
	std::vector<uint32_t> indices;
	for (auto& window : presentation.windows)
	{
		size_t imageIndex = window->imageIndex();
		if (window->visible() && imageIndex < window->numFrames())
		{
			vk_swapchains.emplace_back(*(window->getOrCreateSwapchain()));
			indices.emplace_back(static_cast<uint32_t>(imageIndex));
		}
	}

	if (vk_swapchains.empty())
		// nothing to present so return early
		return;

	VkPresentIdKHR presentIdInfo{};
	presentIdInfo.sType = VK_STRUCTURE_TYPE_PRESENT_ID_KHR;
	presentIdInfo.swapchainCount = 1;
	presentIdInfo.pPresentIds = &presentIdCounter;

	VkPresentInfoKHR presentInfo = {};
	presentInfo.sType		= VK_STRUCTURE_TYPE_PRESENT_INFO_KHR;
	presentInfo.waitSemaphoreCount	= static_cast<uint32_t>(vk_semaphores.size());
	presentInfo.pWaitSemaphores	= vk_semaphores.data();
	presentInfo.swapchainCount	= 1;
	presentInfo.pSwapchains		= vk_swapchains.data();
	presentInfo.pImageIndices	= indices.data();
	presentInfo.pNext		= &presentIdInfo;

	auto res = presentation.queue->present(presentInfo);
	if (res != VK_SUCCESS) {
		printf("ERROR: gilvk::viewer_t::present_and_wait vkQueuePresentKHR returned %d\n", res);
		return;
	}

	res = (*waitForPresentKHR)(presentation.windows.front()->getDevice()->vk(), vk_swapchains.front(), presentIdCounter, UINT64_MAX);
	if (res != VK_SUCCESS)
		printf("ERROR: gilvk::viewer_t::present_and_wait vkWaitForPresentKHR returned %d\n", res);

	presentIdCounter++;
}

[FrancoisBerard]

Even more progress.

I still don't enable the extension on the instance, as this causes the instance creation to fail. But now the call for waitForPresentKHR returns VK_SUCCESS. Only on the Mac: when tested on Linux/NVIDIA, the call for waitForPresentKHR causes a segmentation fault, even though I check that waitForPresentKHR, the vk device and vk swap chains are not null.

So, I seem to have a successful execution on the mac, now I need to diagnose if it actually blocks until an accurate presentation time.

[FrancoisBerard]

Ok, I think I can confirm that it works now, both on Apple Silicon and Linux/NVidia.

The present_id extension needs to be activated on the device in addition to present_wait, that was what made the Linux/NVidia crash (I guess the Apple stack is more forgiving).

windowTraits->deviceExtensionNames.push_back("VK_KHR_present_wait");
windowTraits->deviceExtensionNames.push_back("VK_KHR_present_id");

I measured the execution time of waitForPresentKHR. On NVidia, it is almost 16.6 ms with a 60Hz refresh rate. My scene is very simple, it is quickly rendered and "wait for present" has to wait for almost a full refresh cycle. But if I let the program sleep for 14ms after a presentation, then wait for present only wait for about 2ms... which seems coherent with a synchronization to the next display refresh.

On Apple the timings are less regular, between 1 to 2 ms on a similar 60Hz refresh rate with no sleeping between frames. But this is still coherent with a synchronization.