Improve object picking performance

[Aeva]

An interesting performance bottleneck was made evident by #189, wherein when static drawing a largish number of objects, performance would be about 60fps until a picking pass was invoked, at which the frame rate would drop very sharply.

Picking performance could be improved in a lot of ways:

• [ ] Picking should be rendered entirely with indirect buffers, which would then be rendering at a lower resolution
• [ ] The picking pass could be overriding by registering a render node for the object ID pass and the location pass respectively. The stock fragment shaders could use the MRT syntax described in #191 to just provide these automatically.
• [ ] When the picking pass is not provided with render nodes, it should make use of the MRT syntax described in #191 to combine both render passes when possible.
• [ ] Configure "pick location" per object rather than per graph (and maybe allowing for an optional global configuration for situations where it would be simpler), since non-MRT location picking is expensive.

[wijnen]

Also, I'm not sure how much overhead the location info costs, but given that it must be switched on, I take it that it's significant.

From a practical standpoint it seems to me that this should be defined per object, not as a global setting. For example, in a board game it may be useful to get coordinates when clicking on the board, but you don't care when clicking on one of the pieces.

So perhaps you should consider moving skip_location_info from graph.picking to every (leaf) node.

[Aeva]

The idea for having it per graph is that you likely would be using simplified geometry for the picking pass. so one would use a different graph for rendering vs picking.

Location info currently requires a 3rd pass on top of object picking, though now that I have a mechanism for multiple render targets in place, the two could be consolidated on systems that support it.

[wijnen]

Yes, I expected that a 3rd pass was required, that's why I suggested to make it object-based; then the 3rd pass can be skipped if the object doesn't need it.

If both passes are bundled, there is no need for that, of course.

[Aeva]

the reason why it draws the entire scene instead of just the individual objects needed for picking is because if a pickable object is occluded by another object (pickable or otherwise), then it should not be something you can select.

[wijnen]

For the third pass this is not relevant; it was established in the second pass that the click did land on the object. Rendering only that object may be wrong for other pixels, but you're only going to look at the same pixel as in the second pass, and that pixel is guaranteed to be visible.

This isn't what I was talking about though. I'll give an example to hopefully be more clear. I made a game with 16 pieces and a board. Players click on a piece to select it, or on the board to place it. When a piece is clicked, I do not care about the location of the click. If the board was clicked, I need to know which square was clicked.

In the current system, because the board needs to know the location, I must set graph.skip_location_info=false. But this means that the third pass is always activated on click, both for the board and for the pieces. In case of the pieces, this is a waste of resources.

That's why it seems to me that it would make sense to set skip_location_info on the object that has selectable=true instead of on the graph. Then the third pass would only be activated when the board is clicked, not when the pieces are clicked.

[Aeva]

Oh! I think I get what you're saying. Let me make sure I understand correctly, what you're proposing is that "skip location info" should be a per-object setting rather than a global setting? That seems sensible.

[wijnen]

Yes, exactly.

[Aeva]

Ah! yes that is a good idea. I will include it for this feature :D I updated the above checklist.