<rdar://problem/120916864>

Here's a couple of repro cases: Plain JS repro: https://jsfiddle.net/yjbanov/0vurdt42/8/ Flutter Web repro: https://image-bitmap-stress-test.web.app/ Fixing this will help us make Flutter web apps run much smoother in Safari and consume less memory.

https://jsfiddle.net/yjbanov/0vurdt42/8/ seems fast but https://image-bitmap-stress-test.web.app/ is still slower in Safari than Chrome.

This is exciting! We'll kick the tires in the newest Safari. This variant is still drastically slower compared to Chromium-based browsers: https://jsfiddle.net/6e4dz57v/4/

The difference may be in transferToImageBitmap vs createImageBitmap

I'd like to contribute two additional demos that cover scenarios not currently represented in this ticket: 1. The efficiency of `HTMLCanvasElement.captureStream` implementation: https://github.com/mstyura/webkit-issues/blob/main/video-transform-pipeline/README.md#mirror-4k-canvas-to-video-element-via-htmlcanvaselementcapturestream 2. The efficiency of `VideoTrackGenerator` when produce frames via `new VideoFrame(offscreenCanvas)`: https://github.com/mstyura/webkit-issues/blob/main/video-transform-pipeline/README.md#produce-a-mediastreamtrack-via-videotrackgenerator-from-an-offscreencanvas-rendered-in-a-worker

One thing I've noticed in the WebKit source is that I believe any use of IOSurfaces looks to set kIOSurfaceCacheMode to write combined: https://github.com/WebKit/WebKit/blob/46f0a4c36be821652d1582cbfabe5e51cfee5176/Source/WebCore/platform/graphics/cocoa/IOSurface.mm#L200 This is the default in AVAssetWriter too and in a separate project I discovered reading those buffers on the CPU had a large performance penalty (10x or more). There's a comment in iOS SDK the IOSurfaceTypes.h header that says this: /* ** Note: Write-combined memory is optimized for resources that the CPU writes into, but never reads. ** On some implementations, writes may bypass caches, which avoids cache pollution. Read actions may perform very poorly. ** Applications should investigate changing the cache mode only if writing to normally cached buffers is known to cause ** performance issues due to cache pollution, as write-combined memory can have surprising performance pitfalls. */ Perhaps worth someone at Apple reviewing the reasoning for setting write combined here and if the benefit is worth the penalty for CPU access.