Bug 978338 changed the tiles to be a fixed ratio (and size) per bug 975208 design. This caused TART regression for bug 990212. And in that bug 990212 comment 7, I guessed that thumbnails regress TART more than directory tiles because history thumbnails are captured at higher resolutions than the directoryLinks' imageURISpec data uri which has the correct size of 243x150. Not sure if these thumbnails are used elsewhere that would require higher resolutions. Or if so, we could specially cache the correctly sized thumbnails for newtab.

(In reply to Ed Lee :Mardak from comment #0) > This caused TART regression for bug 990212. Sorry, I meant when Directory Tiles appeared with bug 975228 landing, things regressed as now newtab actually had content to show instead of empty.

Tim, do you know if we need to capture at 2x the size so that they will not look pixelated on HiDPI displays?

Yeah, we need to have sizes specific to the resolution of the default output device. We could nowadays in theory have static image sizes because tiles are static, however we have other thumbnail consumers. In that case it's not really easy to say which size we should capture in a just a slight deviation from the actual size would make us scale the image, not just putting it 1:1 into the back buffer.

We could always capture at the HiDPI resolution, and scaling to 1/2 size is pretty quick.

That is only true for about:newtab then. Other consumers display thumbnails in other dimensions and that might impact quality or speed there, at least a little. Also, I remember from some tests we did at the newtab perf work week that images captured at HiDPI resolution looked quite ugly. The default mechanism to scale down images doesn't seem to do too well without some blur on some platforms.