Sample trace

Macrobenchmark FrameTimingMetric makes a couple assumptions that I'm not seeing hold true on the trace above:

that on API 31+, frame ID is constant between:

actual <nr>
expected <nr>
Choreographer#doFrame <nr>
DrawFrame <nr>

Note: we observe instead that frame numbers for actual/expected get out of sync, see screenshot.

end of the actual frames align with the end of the frame

Note: we observe instead that the actual frame is wrapping a separate renderthread render

The problem

Together, these mean that macrobenchmarks drop many frames, and don't report them as part of metrics. In this scenario (MotionLayoutBenchmark#messageJson), it's roughly ~100 frames dropped out of ~140.

See also b/210605870