FP16 Extension Document #2647

jzm-intel · 2022-03-07T07:34:45Z

Define the GPUFeatureName fp16-in-shader-and-buffer and describe this extension in WebGPU spec. In WGSL spec, add f16 extension identifier and description, add f16 scalar type and corresponding vector, matrix types and describe their memory layouts, add type constructor expressions, zero value expressions, conversion expressions, reinterpretation of representation expressions, arithmetic expressions, and comparison expressions for f16, and describe accuracy requirement for expressions for f16.

Issue: #2512

EDIT:
This PR currently has a complete markdown document that describe the whole extension.

Define the GPUFeatureName `fp16-in-shader-and-buffer` and describe this extension in WebGPU spec. In WGSL spec, add `f16` extension identifier and description, add `f16` scalar type and corresponding vector, matrix types and describe their memory layouts, add type constructor expressions, zero value expressions, conversion expressions, reinterpretation of representation expressions, arithmetic expressions, and comparison expressions for f16, and describe accuracy requirement for expressions for f16. Issue: gpuweb#2512

jzm-intel · 2022-03-07T07:42:35Z

Following the guidance of #2301, this PR add a markdown file describing FP16 extension in the extensions folder. This extension is aimed at being merged into the spec eventually.
This PR contains only a part of the full extension, the remain part will be updated in next PR(s). The full extension draft can be found at https://hackmd.io/912xyyVDR3y-r0NsgglRJw.

github-actions · 2022-03-07T07:43:25Z

Previews, as seen when this build job started (d0ba8df):
WebGPU | IDL
WGSL
Explainer

jzm-intel · 2022-03-07T07:45:22Z

Please take a look, thanks!

kdashg · 2022-03-09T18:50:11Z

WGSL meeting minutes 2022-03-08

DN: On our side, we wonder if this should be a branch. Like, how should we add this? Branch to be folded in later?
AB/KG: Write the PR that adds the feature to the main spec.
AB: There’s value in writing separate high level doc to describe the feature, for initial review. But we’re past that now.
DN: The PR can be incomplete, and then get added onto. Wait to land it until the whole feature is described in that PR, or “complete”.
KG: If a separate doc is useful, then use the separate document. For this one think this one is past that stage.
JB: I don’t see a way to write a f16 literal. Deliberate?
AB: Conversion operators. We waved them off of trying to add suffixes.
MM: So why have suffixes at all?
KG: I think we had more of a reason for them historically, though we have less pressure to do so now. Do you want to file an issue?
MM: I probably should file an issue.
KG to write issue for suffix discussion

kainino0x · 2022-03-10T01:13:50Z

DN: On our side, we wonder if this should be a branch. Like, how should we add this? Branch to be folded in later?

I prefer maintaining documents like this in the extensions/ directory, over a branch, until they're ready. But if it's time to spec this in the actual specs, we can go straight to PRs against the specs.

kainino0x · 2022-03-10T01:17:58Z

extensions/fp16-in-shader-and-buffer.md

+
+# FP16 usages support (PR0)
+
+**Roadmap:** This draft extension is **on the standards track**. It is currently intended to be used in Web-compliant contexts, any user with a capable device will be able to use this extension to acquire the FP16 ability. 


Anything that's not described in the actual specs must not be exposed to the Web. So if we land this document, this should say something like:

Suggested change

**Roadmap:** This draft extension is **on the standards track**. It is currently intended to be used in Web-compliant contexts, any user with a capable device will be able to use this extension to acquire the FP16 ability.

**Roadmap:** This draft extension is **on the standards track**, but is a work in progress. User agents must not implement/expose these features until they are merged into the main specs.

github-actions · 2022-03-10T01:35:10Z

Previews, as seen when this build job started (56446bc):
WebGPU | IDL
WGSL
Explainer

jzm-intel · 2022-03-10T04:47:22Z

I prefer maintaining documents like this in the extensions/ directory, over a branch, until they're ready. But if it's time to spec this in the actual specs, we can go straight to PRs against the specs.

For this FP16 extension, if we going to maintain a markdown document, I will simply move the content from the draft in HackMD into this document with minor fixes and ask for review. After all content are moved to the markdown document here, another PR against the main spec shall be opened (assumedly by me) and go for another rounds of reviews.

kdashg · 2022-03-14T06:03:28Z

I prefer maintaining documents like this in the extensions/ directory, over a branch, until they're ready. But if it's time to spec this in the actual specs, we can go straight to PRs against the specs.

For this FP16 extension, if we going to maintain a markdown document, I will simply move the content from the draft in HackMD into this document with minor fixes and ask for review. After all content are moved to the markdown document here, another PR against the main spec shall be opened (assumedly by me) and go for another rounds of reviews.

This works for me. There's enough consensus in the group to move to making the actual PR to add this to the WGSL spec, though it should be more or less the whole extension, not just parts of it.

jzm-intel · 2022-03-14T06:09:45Z

If I get it right, I will make this PR a complete markdown document, and after this PR is landed I will open a PR against the main spec.

github-actions · 2022-03-15T07:00:52Z

Previews, as seen when this build job started (7a3343f):
WebGPU | IDL
WGSL
Explainer

github-actions · 2022-03-15T12:44:31Z

Previews, as seen when this build job started (e4b8a79):
WebGPU | IDL
WGSL
Explainer

kdashg · 2022-03-15T23:24:28Z

WGSL meeting minutes 2022-03-15

**KG: **intel wants to make a document, then a PR. We suggested that they just make a PR directly. Does anyone think there should not be a separate document at all, and they should just jump directly to a PR?
AB: Does the document get updated to change the contents of the PR? There’s a place for a document describing an overview of the feature. There’s less utility in a document that just describes the diff that would be the PR.
KG: I think the former. Then once it’s merged, we can get rid of the document
MM: Intel is doing a lot of work, and we should let them work in whatever way is most effective for them
AB: Making them write that twice is what we want to avoid, we should help them not do that.
KG: It seems like that (write it twice) is what they’ve chosen to do, so we should let them approach it how they want.

github-actions · 2022-03-24T03:03:45Z

Previews, as seen when this build job started (a9f9a18):
WebGPU | IDL
WGSL
Explainer

github-actions · 2022-03-25T07:50:37Z

Previews, as seen when this build job started (c158c38):
WebGPU | IDL
WGSL
Explainer

github-actions · 2022-03-25T08:35:15Z

Previews, as seen when this build job started (d40029c):
WebGPU | IDL
WGSL
Explainer

github-actions · 2022-03-27T14:05:55Z

Previews, as seen when this build job started (6d8a555):
WebGPU | IDL
WGSL
Explainer

jzm-intel · 2022-04-14T08:02:38Z

Close this PR, as the extension against spec PR #2696 has been merged.

jzm-intel requested review from bjjones, dneto0, kainino0x, kdashg, kvark and litherum March 7, 2022 07:44

kainino0x reviewed Mar 10, 2022

View reviewed changes

Fix the roadmap section

56446bc

kdashg mentioned this pull request Mar 14, 2022

Why have literal suffixes at all? #2664

Open

Make the extension markdown document complete

7a3343f

jzm-intel changed the title ~~FP16 Extension: Part 1~~ FP16 Extension Document Mar 15, 2022

Fixing format

e4b8a79

kvark mentioned this pull request Mar 20, 2022

[hlsl-out] Insert padding between struct members gfx-rs/naga#1786

Merged

Minor rephrase for matrix type

a9f9a18

Rephrase the bitcast type rules

c158c38

Rephrase Arithmetic Expressions section to match the updated spec

d40029c

Add WebGPU change for f16 pipeline-overridable constant

6d8a555

jzm-intel mentioned this pull request Mar 28, 2022

FP16 Extension against spec #2696

Merged

jzm-intel closed this Apr 14, 2022

jzm-intel mentioned this pull request Apr 14, 2022

Proposal of FP16 WGSL extension #2512

Closed

TimothyBergstrom mentioned this pull request May 14, 2022

Parse shaders with half-precision floats (float16) gfx-rs/naga#1149

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FP16 Extension Document #2647

FP16 Extension Document #2647

jzm-intel commented Mar 7, 2022 •

edited

Loading

jzm-intel commented Mar 7, 2022

github-actions bot commented Mar 7, 2022

jzm-intel commented Mar 7, 2022

kdashg commented Mar 9, 2022

kainino0x commented Mar 10, 2022

kainino0x Mar 10, 2022

jzm-intel Mar 10, 2022

github-actions bot commented Mar 10, 2022

jzm-intel commented Mar 10, 2022 •

edited

Loading

kdashg commented Mar 14, 2022

jzm-intel commented Mar 14, 2022

github-actions bot commented Mar 15, 2022

github-actions bot commented Mar 15, 2022

kdashg commented Mar 15, 2022

github-actions bot commented Mar 24, 2022

github-actions bot commented Mar 25, 2022

github-actions bot commented Mar 25, 2022

github-actions bot commented Mar 27, 2022

jzm-intel commented Apr 14, 2022


		# FP16 usages support (PR0)

		Roadmap: This draft extension is on the standards track. It is currently intended to be used in Web-compliant contexts, any user with a capable device will be able to use this extension to acquire the FP16 ability.

FP16 Extension Document #2647

FP16 Extension Document #2647

Conversation

jzm-intel commented Mar 7, 2022 • edited Loading

jzm-intel commented Mar 7, 2022

github-actions bot commented Mar 7, 2022

jzm-intel commented Mar 7, 2022

kdashg commented Mar 9, 2022

kainino0x commented Mar 10, 2022

kainino0x Mar 10, 2022

Choose a reason for hiding this comment

jzm-intel Mar 10, 2022

Choose a reason for hiding this comment

github-actions bot commented Mar 10, 2022

jzm-intel commented Mar 10, 2022 • edited Loading

kdashg commented Mar 14, 2022

jzm-intel commented Mar 14, 2022

github-actions bot commented Mar 15, 2022

github-actions bot commented Mar 15, 2022

kdashg commented Mar 15, 2022

github-actions bot commented Mar 24, 2022

github-actions bot commented Mar 25, 2022

github-actions bot commented Mar 25, 2022

github-actions bot commented Mar 27, 2022

jzm-intel commented Apr 14, 2022

jzm-intel commented Mar 7, 2022 •

edited

Loading

jzm-intel commented Mar 10, 2022 •

edited

Loading