I agree that the browser should be more like a game engine, and less like an X11 buffered window system, in terms of just rasterizing faster enough to render 60fps. But it seems like to me that the CSS spec contains a number of features (rounded corners, soft shadows, etc) that by themselves can be rendered fast on a GPU, but cobbled together in a complex scene graph might result in stalls.
Hinting really isn't important on mobile or HiDPI. In fact, no version of OS X or iOS does it. (I don't know about Android, but I wouldn't be surprised if it doesn't on HiDPI screens.)
What's more interesting is antialiasing quality, especially subpixel AA. That's something I think needs more investigation. But I'm cautiously optimistic; there's been some really exciting work in GPU AA and vector graphics lately :)
> Wouldn't a more realistic example be of a UI that contains mainly different font weights and sizes, to better gauge the effects of all of the pipeline state changes.
Don't change state then. :)
State changes really shouldn't be necessary. If you set things up properly, you should be able to render arbitrarily many glyphs with one draw call, bound only by your texture atlas size.
> But it seems like to me that the CSS spec contains a number of features (rounded corners, soft shadows, etc) that by themselves can be rendered fast on a GPU, but cobbled together in a complex scene graph might result in stalls.
They typically don't. The key is to render critical resources like that to a texture atlas and to batch like resources together as you do so. (For example, batch all box shadow pieces together into one draw call, batch all rounded corners into another, etc.) This effectively places a small fixed upper bound on the number of draw calls you issue. Then you can rerender the page for each animation frame with a very small number of draw calls (frequently just one) and a simple blitting shader. If you get your draw call/state change overhead down, essentially nothing on the Web comes close to taxing a modern GPU; GPUs are incredibly fast at rendering batches.
My talk goes into more details, if you're curious: https://air.mozilla.org/bay-area-rust-meetup-february-2016/
What's the status of Servo BTW? I mean in terms of HTML/CSS spec completeness.
Edit: just saw the demo in the video, awesome job!
Optimally implemented, I believe there should be no difference between "rasterization" and "compositing". You're painting the same number of pixels, paying the same fragment shading and rop cost, either way. Caching fully rendered Web content in textures assuming that the cost of rasterizing that content is high isn't working very well, and I think it's not worth it.
[1]: http://wdobbie.com/post/gpu-text-rendering-with-vector-textu...