Recent benchmark output is showing a large number of frames that take no time (under 20 nanoseconds) and happen back to back right after one another.

Unfortunately, these frames are also being recorded in the event stream and the benchmark reports will average them in with the frames that do real work. The net effect is that a bunch of "(near) zeros" are averaged in and cut the "average rasterizer time" in half.

For example, here is a graph of the rasterizer frame durations in a recent run of the imagefiltered_transform_animation_perf__timeline_summary benchmark. The first half of the frame times are all essentially zero/ungraphably small...

The most likely cause is the engine change (flutter/engine#18901) which changed the way we dispatch the GPU rasterizer thread on layer trees. That engine change landed in the framework as part of an engine roll (413702b) on 2020/06/11.

Note that averaging the rasterizer times in the output while ignoring any frame under 50ns yields a number that is similar to the data we have on that benchmark from before that fix landed, so it seems that the sudden improvement in that benchmark (and likely others) was specifically due to including bad numbers in the averaging function.