Lazy State update regression

[rolandhill]

Hi Robert,

My application has a RenderGraph with multiple passes to render Opaque objects, Transparent, Text Labels etc. Each pass is done using it's own View with a different mask for each view. All the views traverse the same scenegraph with different nodes being traversed according to the mask. I have created a specific Switch that goes at the top of each branch of the graph (each branch is a type of user selectable data) that accepts a RecordTraversal and will decide whether to traverse the scene depending on the value of the traversalMask and user selectable state for visibility and alpha.

This all worked perfectly well until VSG commit c1528ac, lazy state. This changed:

That change has been further refined to now be

In my application, the user now turns Text Label visibility on, but nothing appears in the scene. I can trace program execution and see that the relevant nodes are being traversed, but nothing is rendered.

If I put breakpoints in most methods within StateStack for the label pass, I can see that push() is called (see image below) with pos = 0 and the value equals the same as in stack[0]. Both stack[0] and stack[1] are already filled with old values. During push(), pos is pre-incremented to 1 and the new value overwrites the existing stack[1]. The problem is, stack[1] now equals stack[0], so current->record() is not called during the StateStack->record() function. Therefore, nothing gets rendered.

Looking at the StateStack class, it appears to be designed so that stack[0] is initialised to nullptr, and then updated with the "previous" value so that unnecessary calls to record the same state are not made. In this case, it looks like stack[0] is not being reset to nullptr before my LabelText view starts the record traversal.

If the user adds in a new branch within the scene graph, suddenly the text labels from the first branch will appear, even though the two branches are completely separate. Presumably this is because stack[0] is left with a different old value before the labeltext view is recorded.

I think this is what's happening. It's hard to explain - I hope it makes sense.

Regards,
Roland

[robertosfield]

Sorry to hear this optimization has caused a regression for your usage case.

A missing reset of the state stack might be all that is required, but where to add it will be the crucial question.

To figure out the most appropriate fix it would be useful to have example program that illustrates the usage case and the failure. Are there any existing vsgExamples that use a similar scene graph set up to your application?

If not then would it be possible to create one by adapting one or creating a new example. This could then be used to recreate the bug and then used as a unit test for fixing the problem and there after as test & example.

[rolandhill]

I've submitted a PR to fix this issue #1502

vsganaglyphicstereo uses multiple Views attached to a RenderGraph, but doesn't show the problem. The problem only manifests when the record traversal from the previous view leaves the StateStack in a certain configuration.

The PR above simply calls state->reset() at the beginning of the record traversal for each view. Is there any reason why you would want state carried over from one view to the next? It does solve the issue for my use case.

Roland

[theodoregoetz]

I'd like to reopen this issue since the reset introduced in PR #1502 broke use-cases where we have nested views. The state stack needs to remain in-tact for that to work. The solution to this problem may be more along the lines of a more targeted state clean-up at the end of RecordTraversal::apply(const View& view)

[rolandhill]

See further discussion at #1538 (comment)