I don't think we're imagining the same things here. Think less '3D landscape with text floating in it for some reason' (e.g. SecondLife), and more '3D landscape with labels applied to it' (e.g. Google Earth). Nothing from the "VR" layer obscures the "labels" layer; the labels float above everything, as if part of the GUI, but following the same scale as the things they label.
> PDFs
I'm actually unsure why a (non-interactive) PDF page, displayed at 100% zoom, can't be baked into a single extremely-high-resolution tile—applied as a texture to a single GL rectangle with plain-old lanzcos downsampling. That'd work for the overwhelming majority of the frames.
You'd only need to re-render the tile when you zoom in enough that it'd look bad; effectively, you could think of the PDF page as coming from a mipmapped texture, but where you only have the current size and will just-in-time rerender when the zoom-factor changes. And then, when the zoom factor makes the PDF not fit on the screen, the "tile" would be the screen size—so you'd re-render sibling tiles when the user moves.
Or, in other words: PDFs can be rendered exactly the way the image tiles on a mapping website are rendered, can't they? Doesn't that approach win over re-rendering all the text from a font-atlas every frame? And if not, why do browser rendering engines use tiles? (This isn't a rhetorical question to make my point; I'm not a graphics dev and I honestly don't see why.)
Sure, but that uses a lot of memory, and initial pageload is going to be really slow. Especially when you consider you need to generate mipmaps if it's really high resolution, or else blow out your GPU's L1/L2 cache in FS execution.
> Doesn't that approach win over re-rendering all the text from a font-atlas every frame?
Not significantly, in my experience. Assume, in the simple case, that you're just rendering a bunch of text with no overlap. In that case, the only thing that tiles buy you over rerendering all the text from an atlas is decreased vertex count. You're still touching the same number of pixels in the FS and ROP units either way: you just do so with more vertices in the texture atlas case and fewer vertices in the tiling case. Now consider that creating and maintaining the tiles has costs over maintaining the glyph atlas (which you have to do either way): you have the memory of the tiles, the overhead of creating and switching FBOs, lots of little textures to keep around (which in naive implementations results in tons of state changes), and extra draw calls to render the tiles after rendering the content.
> why do browser rendering engines use tiles?
Mostly because it was an easy way to fit GPU-accelerated panning and zooming into the existing, originally CPU-based, rendering architectures that browser engines used (and still largely use). Tiling was popularized by Mobile Safari on the iPhone in 2007 as an easy way to avoid repainting the entire page on the CPU every frame when the user performed touch gestures. I don't think it's necessarily a globally optimal decision.
I think this is actually a harmful approach in most cases--all it does is reduce the number of vertices and slightly reduces overdraw in exchange for a lot more memory consumption and a lot more draw calls. Overdraw can be reduced in other ways, such as sorting front to back and taking advantage of early Z for opaque content. Web pages and PDFs are so totally not bound on either ROPs or VS, and are very sensitive to draw call count, that caching tiles to minimize overdraw is almost never worth it.
But this is somewhat off topic.
> I'm not sure if you need the medium-level (font atlasing) if you have both lower-level (bezier-curve atlasing), and higher-level (pre-composited "tiles.")
You definitely do. You're sacrificing a lot of performance for the rendering of each tile if you don't do that.