Random crash after running for a long time

[gaetanaxs]

We found a random crash issue after running encoder for a long time (e.g. 40 hours). The crash happens randomly when we run command like below:

ffmpeg -y -stream_loop -1 -c:v h264 -i 1.mp4 -an -c:v libsvt_hevc -rc 1 -preset 9 -r 25 -flags +cgop -g 8 -keyint_min 8 -forced-idr 1 -hielevel 0 -pred_struct 0 -bf 0 -la_depth -1 -sc_detection 0 -vf scale=w=1280:h=720 -b:v 500000 -maxrate:v 1000000 -bufsize:v 1500000 -f segment -segment_time 0.32 ./seq1_%d.mp4

With gdb, we captured two crash core dumps as follows:
1st time crash --

2nd time crash --

Both of them point to the function of EncodeQuantizedCoefficients_SSE2 in https://github.com/OpenVisualCloud/SVT-HEVC/blob/master/Source/Lib/Codec/EbEntropyCoding.c

It seems that sometimes the while(1) loop cannot be ended up successfully, leading to the final crash.

Could you help us have a look on it? Thanks a lot!

[tianjunwork]

Hi @gaetanaxs, what is the hw setup? Could you use svt-hevc debug library and attach the information it prints before starting encode? Thanks.

[gaetanaxs]

Thanks! @tianjunwork
CPU: AMD EPYC Milan
memory: 128G
OS: Linux 4.14.105-19-0012 x86_64
GCC: gcc version 4.8.5
SVT CommitID: b09984c

[tianjunwork]

Hi @gaetanaxs, Thanks providing the snapshot. I will try to reproduce on my side. How often did you get random crash? Every time app crashes it is around 40 hours?
Your GCC is a very old version. For better performance, newer GCC is recommended. You can switch GCC versions: https://askubuntu.com/a/26502.

[gaetanaxs]

Hi @tianjunwork, thanks for your suggestions for switching newer GCC.
This random crash appears after running for a quite long time, sometimes it is around 40 hours and sometimes 110 hours. You can launch multiple commands (e.g., 16) on the machine to increase the probability of getting this random crash.

[gaetanaxs]

@tianjunwork
In the captured two crashes, we found that in the function of EncodeQuantizedCoefficients_SSE2,
the while(1) loop cannot be smoothly ended up by triggering "break" sentence (as shown in the figure below), leading to the continuous increasing of "subSetIndex".
Hope this may help. Thank you!