Louder for the people in the back!

Webp is in interesting one, you could probably use a consumer gpu discrete video asic to handle it. I expect it would be faster than a compute shader or CPU software library. There exist dedicated asics for other common image formats as well, but they're usually only present on embedded devices.

That's not to say it's a bad idea or can't be done. You'll just have to invent a new image format that is massively parallelized when coding it.

I experimented with this (for different reasons). The problem with decoding an image on the GPU is that most image formats use huffman encoding at some point in the chain which is inherently a single core operation, and thus badly suited for a GPU. Any performance you gain, you lose at that step.

Example: Wasm app A needs 8gigs of me on startup, frees it immediately, then chugs along using 100mb of mem. Wasm app B needs 9gigs on startup, frees it immediately, then chugs along using 100mb of mem. Your 16GB laptop will not be able to run these 2 apps at the same time. 🤡

What I meant to say is that a wasm app memory will be allocated up front (or grown later on), regardless if it’s ever written to or read from and will be kept allocated even after the app frees parts of it, as wasm doesn’t allow memory shrinkage. The virtual memory proposal tries to fix that a bit.

A more interesting proposal is wasm virtual memory. It’s strange that apps needing more than 4gibibyte of memory can allocate this without the VM being able to swap it out. Need 1TB for 1 sec? Sure, you’re now stuck with it for the remainder of your apps life cycle. 🤷‍♂️

Did he have a staff and a pointy hat?

There is but it fell somewhat out of fashion:

👈