Intent to Experiment: WebAssembly SIMD

340 views
Skip to first unread message

Deepti Gandluri

unread,
Mar 27, 2020, 10:22:33 AM3/27/20
to blink-dev, v8-...@googlegroups.com, Adam Klein, Zhi An Ng, Thomas Lively, Thomas Nattestad

Contact emails

gde...@chromium.org, natt...@chromium.org, zh...@chromium.org


Spec

Overview document: https://github.com/WebAssembly/simd/blob/master/proposals/simd/SIMD.md

Proposal directory: https://github.com/WebAssembly/simd/tree/master/proposals/simd

Tag review: https://github.com/w3ctag/design-reviews/issues/487


Summary

The WebAssembly SIMD proposal defines a portable, performant subset of SIMD operations that map to vector instructions that are available across most modern hardware platforms.


This is purely a WebAssembly performance feature that adds WebAssembly operations and does not affect web API behavior, but is still useful for developers to be aware of as it can change performance characteristics of applications using WebAssembly.


Link to “Intent to Prototype” blink-dev discussion: 

This links to the Intent to Implement discussion for Simd.js, which was the basis for the WebAssembly SIMD. The Spec section above has updated links for WebAssembly SIMD, as well as the Tag review. 


Goals for experimentation

Performance: 

The goal for this experiment is to determine the performance benefit of having WebAssembly SIMD enabled applications in production code, and determine real world performance numbers. We currently have data from isolated benchmarks, and demos, but these do not always reflect real world usage. 


API Shape: 

The SIMD proposal currently consists of a set of operations that are well supported across most modern hardware platforms. The set of instructions available on hardware is large and varied, and we are looking for feedback on whether this portable subset is sufficient for a cross section of applications. 


Experimental timeline

M84-M89


Any risks when the experiment finishes?

The risk is a drop in performance for sites that were enhancing performance using WebAssembly SIMD operations. 


Ongoing technical constraints

None.


Debuggability

WebAssembly SIMD needs debugging support in devtools for 128-bit values, as well as baseline compiler support in V8. Both of these efforts are ongoing. (Tracking bugs: crbug.com/v8/10347, crbug.com/v8/9909)


Will this feature be supported on all five Blink platforms supported by Origin Trials (Windows, Mac, Linux, Chrome OS, and Android)?

Yes.


Link to entry on the feature dashboard

https://www.chromestatus.com/feature/6533147810332672


Mounir Lamouri

unread,
Mar 27, 2020, 4:29:09 PM3/27/20
to Deepti Gandluri, blink-dev, v8-...@googlegroups.com, Adam Klein, Zhi An Ng, Thomas Lively, Thomas Nattestad
Hi Deepti,

This Intent to Experiment looks good, though, I'm wondering why this is expected to run for 5 releases, which is roughly 6 months. Is there something specific we are concerned about and that this timeline aims to solve?

-- Mounir

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.
To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/e4010247-bb4c-47a4-9b82-ea8188c3b232%40chromium.org.

Deepti Gandluri

unread,
Mar 27, 2020, 6:02:36 PM3/27/20
to blink-dev, gde...@chromium.org, v8-...@googlegroups.com, ad...@chromium.org, zh...@chromium.org, tli...@chromium.org, natt...@chromium.org
Hi Mounir, 

The reason for a longer trial here is that this proposal is somewhat large, and while enabling auto-vectorization can give some performance benefits, to really experiment with this applications may have to rewrite their existing code that uses native SIMD intrinsics to use portable Wasm intrinsics (example: Emscripten SIMD intrinsics header), or builtins that map to Wasm operations. 

The intent is to build in enough time for new applications to experiment with this in parallel to ongoing standardization effort in the WebAssembly CG. The ballpark estimate I've used here is from previous applications experimenting with SIMD, with some buffer to gather data. That said, I do understand that this is a longer trial than is usual, so would be open to adjusting the timeline based on concerns. 

Thanks,
Deepti
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

Yoav Weiss

unread,
Apr 2, 2020, 5:12:35 PM4/2/20
to Deepti Gandluri, blink-dev, v8-...@googlegroups.com, Adam Klein, zh...@chromium.org, tli...@chromium.org, Thomas Nattestad
I'm wondering if there's something preventing partners from implementing SIMD versions of their applications with the behind-the-flag implementation?

I'm assuming you already have some partners lined up for the experiment. Would they be able to develop using the flagged implementation? That would enable us to start the origin trial when they're ready, and as a result, have a shorter trial which purpose is to gather real life usage metrics.

To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.
To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/e9f7ebf3-495e-46b7-82bd-a0f963f1625d%40chromium.org.

Mounir Lamouri

unread,
Apr 3, 2020, 2:05:32 PM4/3/20
to Yoav Weiss, Deepti Gandluri, blink-dev, v8-...@googlegroups.com, Adam Klein, Zhi An Ng, Thomas Lively, Thomas Nattestad
Deepti, what did you mean by "ongoing standardization effort"? Is the API shape still being discussed?

Deepti Gandluri

unread,
Apr 3, 2020, 4:21:51 PM4/3/20
to blink-dev, yo...@yoav.ws, gde...@chromium.org, v8-...@googlegroups.com, ad...@chromium.org, zh...@chromium.org, tli...@chromium.org, natt...@chromium.org
The partners we have lined up, currently have a (mostly) full SIMD version of their applications ready. So they would be able to enable this right away. 

There are a couple of issues I was hoping to address with a longer trial, for the partners that currently have an in progress implementation, as well as for partners where to get feedback, there is a level of indirection involved when it comes to how users can take advantage of this. To clarify what I mean by the second use case - for partners like Halide, and Tensorflow.js, which have WebAssembly backends with experimental SIMD support, to get feedback their users will have to sign up for an origin trial, as they serve as backends/libraries and are not end users themselves. So some of that buffer was to ensure we can reach a larger number of users - I'm somewhat unsure how to account for that type of indirection as I haven't found many other examples where this is the case, so feedback here is welcome. 

With respect to "ongoing standardization effort", I should have been clearer there. The proposal is currently at Phase 3, and the standardization effort here refers to moving to Phase 4, the requirement for which is that there needs to be one other Web VM that implements this. At this point, the shape of the API is quite stable, we have an implementation in both V8, as well as the WebAssembly LLVM toolchain, and we are waiting on more engine implementations to move the proposal forward. 
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

Mounir Lamouri

unread,
Apr 7, 2020, 7:40:18 PM4/7/20
to Deepti Gandluri, blink-dev, Yoav Weiss, v8-...@googlegroups.com, Adam Klein, Zhi An Ng, Thomas Lively, Thomas Nattestad
Thank you for the details, Deepti. I'm not an owner but I would recommend going for a 3-release Origin Trial and assess the situation then and decide whether extending makes sense. If there isn't enough usage by then, extending would probably be an easy call. It seems better to me to assess it then than preemptively set the experiment length to 5 releases assuming the usage and feedback will take some time to ramp up.

-- Mounir

To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.
To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/722401d6-1424-47c5-913f-33092ae4b6d0%40chromium.org.

Deepti Gandluri

unread,
Apr 8, 2020, 2:36:42 PM4/8/20
to blink-dev, gde...@chromium.org, yo...@yoav.ws, v8-...@googlegroups.com, ad...@chromium.org, zh...@chromium.org, tli...@chromium.org, natt...@chromium.org
Thanks Mounir, adjusting the experimental timeline to the recommended 3-release trial, and then optionally extending based on feedback sounds reasonable to me. 

To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscribe@chromium.org.

Yoav Weiss

unread,
Apr 9, 2020, 3:50:27 AM4/9/20
to Deepti Gandluri, Jason Chase, blink-dev, v8-...@googlegroups.com, Adam Klein, zh...@chromium.org, tli...@chromium.org, Thomas Nattestad
Thanks Mounir! :)

LGTM to experiment M84-M86

+Jason Chase on the subject of "indirect" Origin Trials. That's something we've seen as a hurdle on other fronts as well (e.g. third party vendors who want to experiment, but need their customers to sign up).

To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.

Jason Chase

unread,
Apr 9, 2020, 5:16:35 PM4/9/20
to Yoav Weiss, Deepti Gandluri, Jason Chase, blink-dev, v8-...@googlegroups.com, Adam Klein, zh...@chromium.org, tli...@chromium.org, Thomas Nattestad
Yes, I'm working on a proposal to address trial participation for third-party, or "indirect", origins. I hope to have something soon. I've created crbug.com/1069623 to track the issue. Folks that are interested are welcome to follow or comment on that issue.

Thanks,
Jason
Reply all
Reply to author
Forward
0 new messages