createImageBitmap(videoFrame) Semantics

[chcunningham]

The plan is to make VideoFrame a CanvasImageSource (#158) and then (quickly) deprecate our videoFrame.createImageBitmap() in favor of window.createImageBitmap(videoFrame). This issue explores how window.createImageBitmap(videoFrame) would be spec'ed (in HTML).

Full credit to @sandersdan for the following. I'm just turning his excellent doc into a bug. I support his proposals. @domenic, are you the right html spec reviewer? Thoughts on the proposals below?

---

Purpose

There are a variety of things that createImageBitmap() for VideoFrame could optionally do, at a performance cost. This document tries to clarify what those are, whether they should be done, and whether there is or should be a configuration parameter.

Background

For a long time, it was assumed that transferToImageBitmap() would be the most efficient way to use VideoFrame, and a lot of focus was put on making ImageBitmap creation efficient (zero-copy). It’s now clear that other APIs (texImage2D(), drawImage(), and WEBGL_webcodecs_video_frame) will offer better performance, so our assumptions should be re-evaluated.

Properties

Lifetime

In some cases we can create an ImageBitmap that references the VideoFrame contents to avoid making a copy. This could require that the ImageBitmap is closed to avoid stalling the decoder, which is atypical for Web APIs.

• Proposal: use zero-copy only when it cannot cause a decoder stall. Apps can rely on GC for ImageBitmaps, as they typically already do.

Crop

It is unclear whether users only need the visible region or will sometimes want the whole coded region.

• Proposal: simplify usage and provide only the visible region. Recommend lower-level APIs for coded region access.

Scale

Scaling the frame to its natural size may prevent zero-copy. It’s also unclear whether API users would prefer raw pixels or 1:1 pixels.

• ImageBitmapOptions has resizeWidth and resizeHeight to control this behavior.
• ImageBitmaps are specified to be measured in CSS pixels, which implies a 1:1 PAR.
• VR360 videos may have a PAR that causes the natural width to exceed the maximum WebGL texture size. We can recommend those users to specify a natural size in the VideoDecoderConfig that does not result in scaling.
• Proposal: simplify usage and scale to the natural size. Recommend lower-level APIs for raw access.

Rotation

Rotating the frame to match the requested orientation may prevent zero-copy. Most API users will prefer that orientation is handled for them, unless they are intentionally trying to pass through the raw stream.

ImageBitmapOptions has imageOrientation, but it is for flipY only.

• Proposal: simplify usage and always apply image orientation. Recommend lower-level APIs for raw access.

Color Conversion

VideoFrames can be created in a wider variety of color spaces than ImageBitmaps. Converting to an ImageBitmap color space is likely to be lossy, but ImageBitmaps do not carry their own color space information.

• ImageBitmapOptions has colorSpaceConversion to disable this behavior.
• Proposal: default to applying color space conversion (to sRGB). Recommend lower-level APIs for raw access.

[domenic]

@domenic, are you the right html spec reviewer? Thoughts on the proposals below?

I don't feel like I have the ImageBitmap or VideoFrame domain expertise necessary to make strong judgment calls here. (@annevk might?) In such scenarios my preferred working mode is to delegate to the multi-implementer consensus, while doing my best to ensure local consistency at a high level. At a quick glance, your suggestions seem to meet that criteria. So the natural next question is, do you have multi-implementer input on these questions?

[chcunningham]

So the natural next question is, do you have multi-implementer input on these questions?

Firming that up now :). @aboba @padenot, any objection to the proposals in comment 1?

[chcunningham]

Discussed in editors call. Generally no objections, but @padenot is taking a closer look.

[chrisn]

On colour conversion, I want to point to discussion happening in the Color on the Web CG, which is looking at how to enable HDR support in HTMLCanvasElement. This includes potential changes to ImageBitmap to support compositing of HDR images with other content in the canvas. See slide deck (slides 33 to 41) presented by @ccameron-chromium at a recent meeting (notes).

[sandersdan]

Yep, we can't expose HDR (except by readback) until <canvas> supports it. Regardless of the final form, <canvas> will support fewer color spaces than VideoFrame, and even within the ones that are supported they will be less configurable (VideoFrame will likely allow independent control of primaries, transfer function, and color matrix).

Given multiple choices for color spaces within an ImageBitmap does mean that we will need to pick the "right" one to convert to, but I expect that createImageBitmap() needs to solve that for non-WebCodecs cases anyway.

[annevk]

whatwg/html#6562 adds display-p3 support to ImageBitmap so you'll have to account for that. I'm not sure I understand the Lifetime proposal. Would an actual copy not result in a stall? Or would you throw if that were required?

[chcunningham]

Would an actual copy not result in a stall?

That's correct. We would copy from a buffer that is effectively "owned" by the decoder (they only own a handful) to a buffer that is owned by the ImageBitmap. This copy would be made silently (no throwing).

[chcunningham]

FYI, updates for createImageBitmap() are now proposed in whatwg/html#6589

Semantics match the what we proposed above.

[chcunningham]

Edit: on second thought, still launch blocking. While we now have other better paths to render VideoFrame (e.g. drawImage()), we still need to define the behavior for createImageBitmap() since this takes a VideoFrame (CanvasImageSource) and changing the semantics may be observable.