Minutes 2022 12 14

GPU Web 2022-12-14/15 APAC-timed

Note that unless stated otherwise this is a GPU for the Web community group meeting and not a working group meeting.

Chair: KG

Scribe: KR,

Location: Google Meet

Administrivia
CTS Update
“Tacit Resolution” Queue
Full burndown!!!
- Passive Fingerprinting Surface #3101
- Permissions Policy for WebGPU and for unmaskHints #3483
- Should the API enforce that writable storage buffer bindings don't alias each other? #1842
Agenda for next meeting

Chromium's shipping plan - published on blink-dev@
- Intent to Ship: WebGPU
- Chrome has announced an intent-to-ship in Chrome 113, shipping in early May. Branches in March.
- Reasonably high possibility that'll slip to M114, but pretty confident we can complete our work by this milestone
KG: we're still figuring out our schedule. Hopefully can communicate something soon (next year).

KN: Not a whole lot more. Gyuyoung finishing up depth bias tests now, then texture sampling tests. Contract ends in a couple of weeks.
Our team's budgeted time to write tests between now and our shipping date. Working on code coverage for Chrome+Dawn, see what we're missing.

WGSL: allow bgra8unorm as texel format with extension bgra8unorm_storage #3640
- Relaying update from WGSL meeting:
- Google team believes it's possible to implement bgra8unorm storage on all platforms with a bit of injected shader code, as long as bgra8unorm is specified inside the shader.
- Agreed to tentatively go forward with this, Google to get implementation experience, and we'll schedule it for revisit in January once we can report back confidently on it (and Myles is back to advocate for it if necessary), and make a final keep/cut decision for V1.
KG: wanted to file an issue for an API for the hint. Which format's the best format?
- Might not be the same as the best format for canvas.
- KN: think we can do this. I’d do an options dictionary to getPreferredFormat, but there are other possibilities.
- KG: think easy for people to add sniffing for this.
Resolution from WGSL meeting: make this an unconditional feature that's available in core WebGPU.

KG: Sorry, still need to table this.
KG: e.g. you can figure out if you're running an NVIDIA card from the last ~4 years, but you can't figure out which one without running performance tests locally.
KG: we think software rendering's the only reliable fingerprinting resistance.

KN(pretyped): I wrote a proposal in the GitHub issue. It defines a single permission policy, as we had been discussing. Seeking tentative approval while I figure out where to get expert review.
- "webgpu-unrestricted"
- Mostly a hint, as written. Only thing you definitely can't do without this is trigger a permission prompt for requestAdapterInfo, if, in the future, there is such a permission prompt.
- BJ: bit confusing: in reality, this'll dictate both the top-level page and can dictate iframes underneath it, too. Normally it'd be granted to "self", and not granted by default to any page underneath it. Semantic issue, but need to be careful when writing the text. Don't discuss top-level and embedded pages separately in the writeup.
  - I like this general direction.

KN: we landed validation in the spec: goes through bound resources, makes sure you don't have 2 storage buffers aliasing same memory.
- Said, it won't be a validation error. Then, went to write it down - Myles not happy with idea we would have undefined results.
- Left at a point - confused whether we had consensus or not.
KG: think this is just needing to get us on the same page about what we can write.