Skip to content

WGSL 2024‐08‐27 Minutes

Jump to bottom
Kelsey Gilbert edited this page Aug 27, 2024 · 1 revision

WGSL 2024-08-27 Minutes

🪑 Chair: KG

⌨️🙏 Scribes: DN, KG

🗺 Location: meet.google.com

⌚ Time: Tuesday **11am-noon **Americas/Los_Angeles (Atlantic-timed)

Specification: https://webgpu.dev/wgsl

Meeting Issues: Marked Issues

Open Issues: Untriaged, M0, M1

Todos doc: WGSL TODOs

Previous: 2024-07-30 WGSL - Agenda / Minutes

Note: These are the minutes taken in real-time. The official minutes can be found on the WebGPU wiki.

If you didn't receive a meet.google.com invitation and plan on participating, please send dneto a Google Apps enabled address and he'll add you.

📋 Attendance

WIP, the list of all the people invited to the meeting. In bold, the people that have been seen in the meeting:

  • Apple
    • Dan Glastonbury
    • **Mike Wyrzykowski **
    • Myles C. Maxfield
  • Cocos
    • Huabin Ling
    • Zeqiang Li
    • Zhenglong Zhou
  • Connecting Matrix
    • Muhammad Abeer
  • Google
    • Alan Baker
    • Antonio Maiorano
    • Ben Clayton
    • Brandon Jones
    • Corentin Wallez
    • dan sinclair
    • David Neto
    • Ekaterina Ignasheva
    • James Price
    • Kai Ninomiya
    • Natalie Chouinard
    • Rahul Garg
    • Ryan Harrison
    • Stephen White
    • Peter McNeely
  • Intel
    • Hao Li
    • Jia A Chen
    • Jiajia Qin
    • Jiawei Shao
    • Narifumi Iwamoto
    • Shaobo Yan
    • Yang Gu
    • Yunchao He
    • Zhaoming Jiang
  • Kings Distributed Systems
    • Daniel Desjardins
    • Hamada Gasmallah
    • Wes Garland
  • Microsoft
    • Damyan Pepper
    • Greg Roth
    • Michael Dougherty
    • Rafael Cintron
    • Tex Riddell
  • Mozilla
    • Ashley Hale
    • Erich Gubler
    • Jim Blandy
    • Kelsey Gilbert
    • Teodor Tanasoaia
  • UC Santa Cruz
    • Reese Levine
    • Tyler Sorensen
  • Unity
    • Brendan Duncan
  • Dominic Cerisano
  • Dzmitry Malyshau
  • Eduardo H.P. Souza
  • Jeremy Sachs
  • Joshua Groves
  • Lukasz Pasek
  • Matijs Toonen
  • Mehmet Oguz Derin
  • Michael Shannon
  • Pelle Johnsen
  • Robin Morisset
  • Timo de Kort
  • Tyler Larson
  • Jason Erb

📢 Announcements/Meta

Office Hour

FYIs and Notable Offline Merges

⏳ Timeboxes (until XX:20)

Clamp overload for scalar min and max with vector value · Issue #4815

  • JB: Mozilla’s thinking:these sound like good changes, with real improvements. Have a lot of places in the spec where it makes sense. Going through them piecemeal seems not optimal. Should review systematically. At that point it looks like a bigger project. The language is usable as-is. We feel it makes sense to treat this as M2. For now focus on shipping what we have right now. When we come to M2 and address all these issue they’ll be very happy with the improvement.
  • AB: We basically agree. We did do one like this, for mix. Do we want splatting in general, or in special cases. Look at it in a batch.
  • JB: For this and the next two items, generally.

Support for mixed type >>= operation · Issue #4816

  • As discussed more generally in #4815, M2 for now, but a nice little quality-of-life feature to add down the road.

Support for mat / scalar · Issue #4817

  • DN: Additionally for this, 1. Our sense is linear algebra has add and mul, but less so division. 2. Which division will it be: teh full precision / or the more approximate reciprocal. If we choose one then the other community will soon call for the other, which ends up being ugly.
  • JB: Kinda surprised you view division here as different enough conceptually. Might as well “scale down” as “scale up”.
  • DN: Fine to drop my point 1. Not worth arguing.
  • JP: Our sense is M3+; same with the clamp and shift issues above.
  • KG: We can pick M3+ here and change this or other around later as needed.

bgra8unorm storage textures without bgra8unorm-storage, should generate WGSL error? #4711

⚖️ Discussions

textureSampleLevel arguably doesn't work on Intel Mac in compute shaders · Issue #4818

  • Moz: what a well-filed issue - thanks to Greg for the comparisons, numbers, and plots
  • JB: It’s so strange that it’s only in compute shaders and not graphics shaders; only certain platforms. It falls within our responsibility as implementors to provide portability. Agree with suggestion to implement workaround on those platforms.
  • KG: In the spec we commit to be better than our implementations are, and better than the CTS tests make us be. CTS should match the spec.
  • MM: Difficulty in characterizing. If spec is going to describe some are ok and some are not, that’s going to be tricky.
  • JB: Well the polyfill just samples on integer boundaries and lerps.
  • MM: Your proposal is the AMD mac behaviour is also not acceptable?
  • JB: Yes.
  • MW: Gregg says WebKit is getting better results on AMD. Let’s get to root cause here before we move.
  • JB: I’m concerned that it does work on graphics pipeline, but not in compute pipeline. If everyone is getting good results on graphics pipeline we should give good service.
  • JB: These are the operations that _don’t _do derivatives: you supply the level specifically.
  • MM: Since the author has to invoke these special functions that give explicit levels/graidents, these are going to be common in compute shaders.
  • MM: Seems the AMD case should good enough. Rather than paying twice the cost to do two fetches.
  • KG: There’s a sliding scale of acceptability. The AMD result is strictly better than the Intel result.
  • JB: put the decision in the hands of the implementor how far out of spec you’re comfortable with.
  • MW: Apple’s implementation on AMD precisely matches M1 mac. So would not require a polyfill. Maybe some difference in implementations. The AMD issue is solvable without a polyfill. The Intel UHD 630 issue; should see if that reproduces outside the 630.
  • KG: If the spec says it’s linear, CTS should verify that (within some tolerance).
  • AB: Gregg was creating the test when this came up.
  • MM: Also the function should be monotonic.
  • DN: 1. Get to root cause difference between Chrome and WebKit on AMD. 2. Find out if other intel drivers/devices are affected.
  • KG: Amusingly, in practice there appears to be a consensus constraint, but it is quite loose: e.g. 0->0, 0.5->0.5, and 1->1!

The value of the fragment builtin @position #4777

  • Corentin: Please confirm the milestone. If it is M0 then we would need some resolution.
  • KG: We specify the standard positions. The bug is to specify it more clearly. If someone wants to propose more clear language.
  • DN: Gregg’s concern was that OpenGL might not follow what was specified.
  • JB: It would only be a problem for compat.
  • MM: Seems the cost of a polyfill would be cheap: observe the deviation, then adjust shaders like that
  • KG: It actually occurred with a particular ANGLE configuration. We could either: don’t ship compat on that config; relax compat for that reason, or the implementation may decide to deviate from the spec.
  • Resolution: no update needed to the spec. (but if you disagree, please just make a PR to add the language that satisfies you!)

📆 Next Meeting Agenda Requests

  • Next meeting: (Atlantic-timed) Tuesday September 17 2024, 11a-noon (America/Los_Angeles)
    • KG: Maybe Sep24?
    • MW: That’s during TPAC.
    • KG: Sep 17 then
  • But email Kelsey or the list if you want to meet sooner!

Clone this wiki locally