We found out that
int16_t x = test::FuzzDataHelper::ReadOrDefaultValue(0)
reads 4 bytes from the fuzzer input instead of 2. That means that
almost half the bits in the input data to audio_processing_fuzzer are
ignored. This change adds template arguments to force reading 2 bytes
when we only need 2.
We also add a small manually generated corpus. During local testing we
let the fuzzer run for a few hours on an empty corpus. Adding the
manually-generated files resulted in an immediate coverage increase by
~3%, and then by another 3% over the next few hours.
The manually generated corpus contains a short segment of speech with
real echo. We suspect that triggering Voice Activity Detection or echo
estimation filter convergence can be difficult for an automatic
fuzzer.
We remove the Level Controller config. We read 20 bytes extra after the
config to guard against future configuration changes.
Bug: webrtc:7820
Change-Id: If60c04f53b27c519c349a40bd13664eef7999368
Reviewed-on: https://webrtc-review.googlesource.com/58744
Reviewed-by: Sam Zackrisson <saza@webrtc.org>
Commit-Queue: Alex Loiko <aleloi@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#22269}
Going through the coverage of audio_processing_fuzzer, it was noticed
that it didn't cover AEC-m and AEC-2 code. Therefore this CL adds 2
fuzzer targets that only fuzz the previous generation echo cancellers.
To avoid code duplication, the APM running code was broken out in a
new GN target. We have also changed all fuzzing code to use the
FuzzDataHelper class to avoid manual pointer arithmetic.
Bug: webrtc:7820
Change-Id: Ifea3266e396b487952a736945577fccea15d0e01
Reviewed-on: https://webrtc-review.googlesource.com/36500
Reviewed-by: Henrik Lundin <henrik.lundin@webrtc.org>
Reviewed-by: Sam Zackrisson <saza@webrtc.org>
Commit-Queue: Alex Loiko <aleloi@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#21638}
