Batched/Linear uploads

[Fuzzyzilla]

Replaces the allocate-as-needed logic with bulk linear allocations for both image create, image update, and mesh upload.

• Doubles the framerate over master when compared using a modified demo_app.rs with PresentMode::Mailbox and a few spare swapchain images, even using 1% less of my CPU while doing so. It also starts up a sixteenth of a second faster, clearly the largest benefit here :P

• Image uploads are now batched per-frame and only synchronize once. Should help with stutter on UIs that load/modify fonts or images frequently (Scrubbing the font size slider on the demo_app's FontBook is a good example).

  • Could further improve this by not synchronizing the host at all, leaving all sync to the device.

• Mesh uploads now only memmap the buffers twice-per-frame instead of twice-per-mesh.

  • With a dependency on bytemuck this could be further reduced to a single memmap per frame, but it would require a private dependency addition and a public egui feature flag change for a somewhat trivial perf boost, and I'm not sure the policy on that (could also be achieved with some fairly trivial unsafe ops).

• It should use substantially less memory during these operations, but I do not have a convenient way to measure. On current master, image creation/deltas involve 4 in-memory copies, reduced here to 2.

I have further optimizations in mind, namely deduplicating command buffer commands, as there are currently ~250 commands for 20 meshes which could be greatly reduced. Always unsure on whether I should bundle the changes into one PR since they're a bit disjoint -- please advise!

[hakolao]

Always unsure on whether I should bundle the changes into one PR since they're a bit disjoint -- please advise!

The smaller the PR the easier it is to review and less chance for mistakes. So Separate PRs 👍

[hakolao]

This one is also appreciated! Don't see anything blocking the merge other than lints & merge conflicts

[Fuzzyzilla]

Alright, merged and fixed up! Did a few reads over and I'm happy enough with it to open this up. I hope I did alright with the merge, still learning git.

Doubles the framerate over master when compared using a modified demo_app.rs with PresentMode::Mailbox and a few spare swapchain images,

I should not be allowed to benchmark/write PR messages at 2am. The improvements are much more modest at ~25% improvement in frametimes. The other improvements are still true - I just have no idea how i got that number lol.