All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=590998

Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?keys=agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDAr6Wg8AgM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDAr6v2sAoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDA7_z1rgoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDA78TPpQoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDAr7nasgoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDAr6vioQoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDAr7mwvAkM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDAr4K2pgoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgIDAr52mtAoM


Bot(s) for this bug's original alert(s):

webrtc-mac-large-tests
webrtc-win-large-tests

I would expect this patch to be more complex than before, but 600% is a little bit too much. I will look into it and report back.

It seems to me that the increase in complexity for mac is reasonable, since the VAD started running on the render thread. But in Windows it seems to be too much and it also leaks to the capture thread complexity, which points towards a threading problem in the IE (which we are already aware of). What is your view on it, Per? I will continue looking into it, but I suspect this extreme regression should disappear once we fix the threading model of the IE.

Please disregard my blaming of the threading model in #4. The actual issue is in the vectorisation of the fmaf function. A fix was uploaded here: https://codereview.webrtc.org/1755943002/

aluebs@ - once https://codereview.webrtc.org/1755943002/ lands and the affected perf graphs look good, please let us know if you want that fix to be considered for an M50 merge. The WebRTC 50 branch was cut at r11765 (aka https://codereview.webrtc.org/1739713002), so it currently includes this regression. 

No, I don't think a merge makes sense, since this component isn't enabled anywhere.

Moving to M51 and downgrading to P2 per #7. Thanks for the clarification!

I am marking this as fixed, since the CL has landed and the performance bots returned to reasonable levels. Please note that they are not back to the original levels, since the CL that introduced this regression added a VAD to the render side, so a increase in complexity is expected. But at least there are no 600% regressions anymore.