Skip to content

Fix semantics of VK_KHR_present_wait.#2693

Open
kleinerm wants to merge 1 commit intoKhronosGroup:mainfrom
kleinerm:VK_KHR_present_wait_fix
Open

Fix semantics of VK_KHR_present_wait.#2693
kleinerm wants to merge 1 commit intoKhronosGroup:mainfrom
kleinerm:VK_KHR_present_wait_fix

Conversation

@kleinerm
Copy link
Contributor

The VK_KHR_present_wait spec states that the wait of vkWaitForPresent() should end as close as possible to the point in time when the presented image for a presentID becomes visible to the user ("first pixel visible").

This can be achived by calling _swapchain->notifyPresentComplete() from MVKPresentableSwapchainImage::endPresentation(), ie from the Metal drawables presentCompleted handler, after Metal signals present completion.

Right now, _swapchain->notifyPresentComplete() is called from the Metal command buffers completedHandler(), ie. when the command buffer scheduling the present completes. This typically happens multiple milliseconds before actual present, and in case of presents scheduled via VK_GOOGLE_DISPLAY_TIMING far in the future, there can be a long time between completedHandler() and actual present completion.

This commit moves the _swapchain->notifyPresentComplete() to the more appropriate place, and behind _swapchain->endPresentation(), so if a client returns from a call to vkWaitForPresent(), potentially queried VK_GOOGLE_DISPLAY_TIMING presentation timestamps are already available for the client to fetch.

The VK_KHR_present_wait spec states that the wait of vkWaitForPresent() should end
as close as possible to the point in time when the presented image for a presentID
becomes visible to the user ("first pixel visible").

This can be achived by calling _swapchain->notifyPresentComplete() from
MVKPresentableSwapchainImage::endPresentation(), ie from the Metal drawables
presentCompleted handler, after Metal signals present completion.

Right now, _swapchain->notifyPresentComplete() is called from the Metal command
buffers completedHandler(), ie. when the command buffer scheduling the present
completes. This typically happens multiple milliseconds before actual present, and
in case of presents scheduled via VK_GOOGLE_DISPLAY_TIMING far in the future,
there can be a long time between completedHandler() and actual present completion.

This commit moves the _swapchain->notifyPresentComplete() to the more appropriate
place, and behind _swapchain->endPresentation(), so if a client returns from a
call to vkWaitForPresent(), potentially queried VK_GOOGLE_DISPLAY_TIMING
presentation timestamps are already available for the client to fetch.
@CLAassistant
Copy link

CLAassistant commented Feb 22, 2026

CLA assistant check
All committers have signed the CLA.

@squidbus
Copy link
Collaborator

squidbus commented Feb 23, 2026

This was explicitly moved in #2543 because the place where you have it now has significant unpredictable latency and/or missed callbacks compared to actual present rate, and as a result using present wait even with a few frames acceptable delay would significantly drop FPS.

@SRSaunders
Copy link
Contributor

I would like to second what @squidbus is saying here. Using the internal Optik profiler within the RBDoom3BFG game, I see the following differences when observing VK_GOOGLE_display_timing presentTiming.actualPresentTime.

Look at the horizontal green Present trace, and also the green Present function chart in the lower left corner. You will see that the released code shows rock-solid present timing (in this case 16.771ms or 60 Hz), while this PR causes highly variable, unpredictable recording of present times.

Current MoltenVK 1.4.1 release from the 1.4.341.0 SDK:
Screenshot 2026-02-24 at 1 10 53 PM

Head of stream MoltenVK 1.4.2 with this PR applied:
Screenshot 2026-02-24 at 1 06 28 PM

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants