There might already be a bug on this, please dupe if necessary.

This is a case that could be helped by image checker-boarding, and other 'async' rendering ideas.

I'd like to understand the use in AMP better.  This test loads each iframe directly in the visible view which makes them blocking.  Doesn't AMP's carousel pre-load iframes off screen (giving us a chance to pre-load this kind of image)?

AMP does preload iframes, which would help a bit. This is definitely a slightly more exaggerated case. However, you can probably see a user trying to swipe left-left-left and then hitting the preload limit.

Also, the preload still happens on the main thread, which means that they could contribute to jank.

Yeah, you're right, checkerboarding is probably the right final resting place for this type of problem.

The image decodes do take a while which we can help by preloading images or frankly using smaller images :) In one of my traces, I saw that these were 3264x2448 images, which just takes a while to decode. Image checkerboarding would help here, but then you'd just see a smooth animation but no image which will appear later (I think that might be better, but maybe there are cases where that would be more jarring)

This does seem to be performing relatively OK in that it isn't decode the image every frame, but just when the image first appears.

Also this trace doesn't look bottlenecked on raster, but rather GPU is swamped. Or am I misinterpreting the trace? It seems like the flush takes a while (70ms or so)

I defer to you on any judgement around what's happening. I begin to look pretty ignorant as soon as I exit third_party/WebKit and content/renderer :)

> Also this trace doesn't look bottlenecked on raster, but rather GPU is swamped.

The trace had GPU framebuffer snapshotting on, so it's bottlenecked by that.  With that removed I think GPU is fairly cheap; the two image decodes are part of the janks, but also the main thread work is ~100ms during that time as well.

Can this bug be closed given comment #6?

I think the GPU was swamped due to constant snapshotting, but the original issue of image decoding blocking for a long time remains.

vmpstr@, over to you for reducing image decoding to 0ms.

Looks like both images are jpeg. The async & checkerboarding will help the jank, like Vlad said. But the actual decode time will also improve with the move to SkJpegCodec.