Improve performance for remote backends

[almarklein]

A lot of ideas in #40.

In short here are the steps:

• Encoding the image on the GPU (optional)
  • The purpose is to reduce the image size (in bytes), so the next steps are faster.
  • simple ideas are dropping alpha channel and yuv420
  • more advanced is DCT encoding
• Downloading from GPU into RAM
  • we've made this async which does not make the download itself faster, but because the CPU can do other things, it improves the total performance a lot.
• Encoding the image on the CPU
  • Takes time, should strive to reduce this.
  • turbojpeg helps, and doing the encoding on the GPU would help more.
  • we can also encode the result in the download-logic to avoid one data copy!
• Sending over the network, io
  • kindof out of our control
  • should use a binary websocket for sure
  • but we could respond to the throughput by e.g. increasing compression.
• Presenting in the browser
  • now using an <img> tag
  • would be much better to push the data to the GPU and decode there

[Vipitis]

there is a feature coming up, allowing host image copy gfx-rs/wgpu#9053 which helps in the upload direction but not download(yet). Technically the GPU already has hardware for image (and video) encoding onboard. But I remember this being way into the future for webgpu. The more reasonable path way might be external texture or canvas target which lets the browser encode images using hardware. But that's again not really useful for the remote rendering usecase here.

I remember one research paper from HPG that was about game streaming, and they did a partial reconstruction by the receiver. Its fairly elaborate and targeted towards game applications: https://youtu.be/QGNLitKiMp8

Isn't remote desktop/VDI all video stream compression? Perhaps that's the way to go? Really depends on if you focus on latency or quality when bandwidth constraints.

Some of the modern image formats support progressive decoding, but that is likely a nightmare with support.