Reduce PagedLOD draw calls

[avilkinnick]

Hello, we are working on railway simulator and currently for static objects we are using matrixTransforms that are pointing on PagedLODs (same PagedLODs for same objects) but RenderDoc shows that this way every pagedLOD is rendered in different draw call, how can this be optimized so the same objects would be rendered in only one draw call? And how instancing could be realized here?
Here is code for creating pagedLOD:

auto pagedLOD = vsg::PagedLOD::create();
pagedLOD->bound = vsg::dsphere(vsg::dvec3(0.0, 0.0, 0.0), settings.view_distance);
pagedLOD->children[0] = vsg::PagedLOD::Child{0.1, {}};
pagedLOD->filename = model_filename_path;
pagedLOD->options = options;

And other source code can be viewed at https://github.com/maisvendoo/RRS/blob/v1.9.0-devel/viewer/src/RouteViewer.cpp

[robertosfield]

You explanation is confusing, makes me wonder if you might be misusing PagedLOD.

Normally I'd expect PagedLOD to be used within a hierarchy such as a quad tree, so they are used as internal nodes rather than ones placed nearer the leaves of the scene graph.

My best attempt and making sense of your explanation and the code kinda looks like you are instancing a subgraph many times placing it in different positions in the scene, with the subgraph containing a PagedLOD. This really isn't how I'd expect PagedLOD to be used.

What are you using PagedLOD for? What subgraphs are you putting underneath a PagedLOD?

[robertosfield]

What type of models are you trying to load?

How often are the models instanced in the scene?

[avilkinnick]

All models are loaded with PagedLODs, rails, trees, ground, power lines, buildings, vehicles, semaphores and so on.
And models like ground, rails and power lines could be instanced more than 200 times at once

[robertosfield]

Models that are rendered many times I'd just load at start up and then instance them. This could be done simply by using separate MatrixTransform above a shared subgraph, or using InstanceNode for GPU level instancing.

When GPU instancing you loose view frustum culling for below the InstanceNode, so it can be best to group instanced subgraphs into a local area and then place a CullNode above the InstanceNode so that that local area can be culled as one. This would be good woods, street lights, power lines etc.

I would only use PagedLOD for items that are unique to geographical areas rather than items that are shared.

[avilkinnick]

Thanks, if I will use first method with just separate MatrixTransforms for preload models, would they be rendered in one draw call?

[robertosfield]

The VSG doesn't attempt to batch draw calls at runtime, this is purely a scene graph creation time decision that loads can make.

For the case of multiple MatrixTransform with a shared subgraph, the subgrph may have multiple draw calls, and all these draw calls will be culled for MatrixTransform that isn't culled.

Putting a CullNode above a MatrixTransfrom would add culling of the instanced subgraphs that one screen.