Skip to content

Conversation

@kdobrowo
Copy link
Contributor

No description provided.

@kdobrowo
Copy link
Contributor Author

This code change fixes WaitForVbl WHQL test (d3d9).

@misyltoad
Copy link
Collaborator

I imagine it is due to the test being broken that it fixes the test?

It is not possible for us to implement this properly without present timing maybe (or doing something hacky with present wait).

@kdobrowo
Copy link
Contributor Author

The test:

  1. calls present
  2. calls WaitForVBlank which should "Suspend execution of the calling thread until the next vertical blank signal." (https://docs.microsoft.com/en-us/windows/win32/api/d3d9/nf-d3d9-idirect3ddevice9ex-waitforvblank)
  3. measures time between presents

This is why SyncFrameLatency was added there.

Isn't it good enough solution?

@misyltoad
Copy link
Collaborator

misyltoad commented Sep 15, 2022

No for two reasons:

  • Our frame latency implementation does not use present wait right now. (It is triggered on the present thread when we call queuepresent, not on vblank or TTL)
  • The call you are using waits for up to the Max frame latency, not 1.

@doitsujin
Copy link
Owner

doitsujin commented Sep 15, 2022

If the docs are anything to go by, this function doesn't really interact with presents anyway, so even present_wait wouldn't really help. We can't really implement this correctly.

The main problem with this particular implementation is that we'll return immediately if

  • the app is CPU-bound
  • there is any sort of GPU->CPU sync point between frames (e.g. occlusion queries)
  • no presents are queued between calls

I think there's a way to come somewhat close to expected behaviour if we store a CPU-side timer somewhere and do something like this (pseudocode):

auto now = dxvk::high_resolution_clock::now();
auto refreshPeriod = getDisplayRefreshPeriod();
auto vblankCount = (now - m_startTime) / refreshPeriod;
auto target = m_startTime + (vblankCount + 1) * refreshPeriod;
sleep_until(target);

This is going to need some preliminary work though, and may behave weirdly in VRR scenarios.

@doitsujin
Copy link
Owner

doitsujin commented Sep 15, 2022

I made a proof-of-concept implementation in this branch, this can be ported to d3d9 fairly easily.

@Blisto91
Copy link
Collaborator

Blisto91 commented Mar 9, 2023

@kdobrowo Friendly ping. What is the status of this?

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

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants