Bug 293284 – In WebGPU when calling copyExternalImageToTexture, the results are different if called in the 'load' event vs after

• Status: REOPENED
• Component: WebGPU

[Gregg Tavares]

Created attachment 475317 [details]
image via decode vs addEventListener

In WebGPU these two snippets of code should load the same data

```js
const img = new Image();
img.src = 'noise.png'
await img.decode();
device.queue.copyExternalImageToTexture(...)
```

vs

```js
const img = new Image();
img.src = 'noise.png'
img.addEventListener('load', () => {
  device.queue.copyExternalImageToTexture(...)
});
```

They don't. Note, if you run the first one in first in the same page then the 2nd one will get the same results

Repo attached

[Mike Wyrzykowski]

Likely a bug in the external texture import path

[Radar WebKit Bug Importer]

<rdar://problem/151696944>

[Mike Wyrzykowski]

I think this is above WebGPU in the general webkit image loading path, confirming

[Mike Wyrzykowski]

CGImageGetAlphaInfo(...) is returning a different value in both cases.

When the image is darker (matching Chrome), CGImageGetAlphaInfo(...) returns kCGImageAlphaPremultipliedFirst (0x2), /* For example, premultiplied ARGB */

When the image is brighter (problem case), CGImageGetAlphaInfo(...) returns kCGImageAlphaLast (0x3),               /* For example, non-premultiplied RGBA */

[Mike Wyrzykowski]

This is indeed at least partly a WebGPU issue, we are not handling when WebCore gives us back a premultiplied image (which appears to differ based on explicit call to decode).

After the image is cached, we get back the same image.

Though it also seems odd WebCore is giving us a premultiplied image in some cases when the source image in this repro is not premultiplied

[Mike Wyrzykowski]

Change to match between the two versions and also match Chrome is quite simple, just checking the CTS to ensure zero regressions.

[Mike Wyrzykowski]

Pull request: https://github.com/WebKit/WebKit/pull/45744

[EWS]

Committed 295690@main (b8888bb0a749): <https://commits.webkit.org/295690@main>

Reviewed commits have been landed. Closing PR #45744 and removing active labels.

[Mike Wyrzykowski]

Reopened Bugzilla.
[WebGPU] bloooooom.vercel.app appears different in Safari/FireFox than Chrome, tracking revert in https://bugs.webkit.org/show_bug.cgi?id=295357.