Several months back I was tasked with building an interactive instagram scroller that was to be slotted into large display via a Safari 5.1 “widget”.
The project was generally fairly simple, 3 columns scrolling infinitely in alternate directions. The individual pictures, or “chips” as I called them were to do a 3D rotate transition to keep the visual appeal up.
I covered all the usual suspects to maximize performance, I used as many CSS transitions as possible
to harness the GPU, keept everything asynchronous as possible, kept the DOM cleaned up, use
However the biggest problem I encountered was getting the scroll smooth.
Initially the page was fully dynamic and my anmiation code used the
scrollTop property (while
inserting new chips as needed to prevent stoppage). That proved problematic and terrible performance-wise.
Next I tried inserting chips in the top of their column container with a negative margin and slowly animating the first chip to a margin of 0 (effecitvely constantly pushing all other chips down). This also posed problems and I couldn’t approach what looked like a fluid animation. It would always do a slight, maybe 2-3ms stutter when a new chip was introduced and there was nothing I could do to eliminate it. I tried optimizing my DOM interaction to use DOM DocumentFragment, dropping all jQuery code and going full-native.
Finally I moved to using absolutely positioned chips that I tracked one by one in my animation loop. However, even this was insufficient and I was still getting a stutter.
After much forehead-meets-desk time, I remembered a conversation I had with Tim Caswell a long while back (revolving around the HP TouchPad) where he told me about an interesting quirk of WebKit that turned out to be the cause of my problems:
To make my display flexible I was using the
getBoundingClientRect method to calculate when a chip rolls off-screen
so I can either remove it from the DOM and insert a new chip at the beginning (as well as to calculate initial position).
The problem is WebKit likes to recalculate the layout of the dom pretty much every single time you use something
getBoundingClientRect. (Even getting the window’s
innerWidth will force a recalculation).
The stutter was coming from WebKit constantly recalculating the layout!
So, a quick refactor to use static, precalculated dimensions for the chips (calculated once on window resize/document load) and BOOM, performance issues gone. So, lesson learned?
All calls to get any calculated dimension from the DOM should be cached or avoided.