How to avoid rendundent scene renders and how to work with the render pipeline

[TirushOne]

Hi all

I am having a go at making a virtual texture system for Bevy. For virtual texturing to work you need to render the scene with a prepass that records which parts of objects are visible, and hence, which parts of which textures are needed to render the scene in full. The thing about this though, is that a neive implimenation results in the scene being rendered twice, which is a huge waste. Ideally, you would rasterise the scene once and record all data needed for the rest of the render pipeline, to avoid having to rasterise multipal times (so basicly deffered rendering). This issue compounds with more advanced rendering with features like occulsion culling, as this adds to the duplicated work.

So my question is, how do I avoid the issue of rasterising the scene multipal times, while keeping my code modular and avoiding creating my own wholey new render pipeline? I am willing to get my hands dirty if necessary, but im not even really sure where to tinkering with things.

It's part of Bevy's desigh philosophy that code should be modular and easy to switch out, from what I have seen of the rendering pipeline though, it doesn't quite seem that easy.

Any help is greatly appreciated!