EGL_ANDROID_native_fence_sync: Clarification of what a "fence command" is about egl-registry

[critsec]

Fences signal when all work submitted before them is completed. Their primary use case is synchronization, not timing. And in general for synchronization you want to unblock the dependent work as early as possible.

So that leads to designs where if you already have a synchronization object marking the end of previous work, and you get a request to create a new EGLSync, you just make it refer to that existing synchronization object. When you do this, sometimes you naturally end up with fences that have a signal timestamp from before the EGLSync was created. For synchronization, that's perfectly fine and in fact usually what you want.

[ppaalanen]

Right, but the problem is the wording in the spec.

The spec says, as quoted above, that creating a fence will insert a fence command into the command stream. That means that new work is pending, created by creating the fence. The new work (fence command) must then be submitted to the GPU to be executed. The fence signals when this work has been completed.

I really cannot see any other way to interpret the spec wording.

[stonesthrow]

What critsec described is reality. Unfortunately you can't infer Android specific behavior from this interpretation. Sorry can't give you a more useful answer. Closing.

[ppaalanen]

Why not fix the spec then?

Remove the word "command" from the phrase "fence command", and add a Q&A that explains that fences are not executable commands but something else. Also, remove the wording that says "inserts a command into the command stream" because implementations obviously do not do that, and apparently cannot do that.

The point here is to figure out what the spec means and clarify the situation.

Please, re-open.

[ppaalanen]

As it apparently has been established that the current implementations are correct with respect to the spirit of the specification, then the wording of the specification needs to be corrected to not mislead application developers.

[critsec]

I agree the language could be a little better. Re-opening and assigning myself, though I'm pretty busy at the moment and might not get to it for a bit.

[ppaalanen]

Thank you very much. Looking forward to seeing a new wording while not holding my breath. ;-)

[emersion]

Hi, any news about this @critsec? Just got hit by the same misunderstanding.

[emersion]

Ping