Otherwise if you are very sure about this issue cannot possibly originated via javacpp then I will report this to tensorflow java again.
I do think it's a problem with TF Java. Please read this thread, you may find it enlightening: deepjavalibrary/djl#690
But for us as long as we remove the ZGC option it all works for us. So I am only reporting this here in the hope of ironing out some potential issues around the libraries which might having memory leaks when ZGC is used. I will not have more time and energy to help much with further debug or getting this issue eventually solved.
And that's great that you were able to debug this, but it's not a good idea to rely on GC anyway. It cannot be relied upon.
Issues like that will continue to happen, until you simply stop to try and do things with GC!
It's never ever always going to work correctly. You will encounter issues like that over and over again.
I mean... If I am writing a Java application, the GC is always there, isn't it?
And as I have mentioned that, setting this
org.bytedeco.javacpp.nopointergc=true has no effect, the ZGC flag still gives error.
Maybe we have misunderstanding of what
ZGC flag does? I think it only replace a Java GC algorithm, from the default G1GC to something else.
It is not we wanted to turn on some GC in our application. It was just because ZGC seemed to be a better GC than the default GC so we used that all the time until we had this out of memory bug.
And that's great that you were able to debug this, but it's not a good idea to rely on GC anyway. It cannot be relied upon.
I don't understand this. In Java there is always GC, I cannot turn it off and we were not rely on anything. -XX:+UseZGC is simply telling Java "use a different (experimenting but hopefully better) GC policy".
Sorry I will have to leave this conversation now. I have posted the issue on tf java as well. It seems we don't have a directly answer why this happens. Probably not many people are using ZGC anyway so this is not some major issue anyway. It's only when people are using ZGC and encounter this issue I hope it would be easier for them to find out that it might be related to ZGC.
@liufuyang I'm talking about GC in relation to JavaCPP and native code in general. It was never designed to be used to manage native resources, and should never be used for anything critical. If you're having issues with GC and JavaCPP, we can disable it by setting the "org.bytedeco.javacpp.nopointergc" system property to "true". If you're still having issues even when doing this, then the issue is probably unrelated to JavaCPP. So, thank you for posting this tensorflow/java#315, and let's continue the discussion there.
Has anyone seen presets that will sometimes produce
java.lang.NoClassDefFoundError exceptions? I see it as Could not initialize class org.bytedeco.ffmpeg.global.swscale, and random suggestions are usually around static initialisers. It mostly works for me locally, and dies in the github CI during test, which is ugly to debug. Any suggestions appreciated. Will see what I can find in any case.
Debug: Loading /home/runner/.javacpp/cache/ffmpeg-4.3.1-1.5.4-linux-x86_64.jar/org/bytedeco/ffmpeg/linux-x86_64/libswscale.so.5
Debug: Creating symbolic link /home/runner/.javacpp/cache/ffmpeg-4.3.1-1.5.4-linux-x86_64.jar/org/bytedeco/ffmpeg/linux-x86_64/libswscale.so
Debug: Locking /home/runner/.javacpp/cache before extracting
Debug: Failed to extract for jniswscale: java.lang.UnsatisfiedLinkError: java.nio.channels.FileLockInterruptionException
I was trying to understand what might be making that message. Looks like https://github.com/bytedeco/javacpp/blob/master/src/main/java/org/bytedeco/javacpp/Loader.java#L1758
Why its sometimes getting
interrupt()ed is perhaps a race between that loading sequence and the calling code?
Hello:)
I saw that the issue like
[ERROR] Failed to execute goal org.bytedeco:javacpp:1.5.6-SNAPSHOT:parse (javacpp-parser) on project cocoa: Failed to execute JavaCPP Builder: /home/doktor/code/javacpp-presets/cocoa/cppbuild/linux-x86_64/include/CoCoA-0.99712/TmpMorseElement.H:551:Unexpected token 'EOF' ->Occurred several times to couple of you. I saw one suggestion to build the presets on Centos here. However, when I tried I got an error:
autoreconf: running: /usr/bin/autoconf --force
autoreconf: configure.ac: not using Autoheader
autoreconf: running: automake --add-missing --copy --force-missing
configure.ac:15: error: require Automake 1.14.1, but have 1.13.4
autoreconf: automake failed with exit status: 1I also saw that the EOF error can be solved otherway (like bytedeco/javacpp-presets#889). However, the parsed file is quite simple in point of view of C++ macros...
What should I do to be added to contributors of a project? I'd be able to push my feature branch to share more details about my issue.
Start by going through https://github.com/bytedeco/javacpp/wiki/Mapping-Recipes, but if you already have something, feel free to show what you have, we can help from there. Thanks!
Can you let me to do push to the presets repository?
Parser has problem with this file
I forked your repo and introduced my changes to symbollic-algebra branch: https://github.com/doktor-ziel/javacpp-presets/tree/symbollic-algebra
It would be great if somebody looks at it:)
and hi :)
Hi, if you're using the Maven plugin, that's the includePaths here in the pom.xml file:
https://github.com/bytedeco/javacpp-presets/blob/master/pom.xml#L188
https://github.com/bytedeco/javacpp-presets/blob/master/pom.xml#L188
i've made some pretty decent progress today. :) failing now trying to compile the generated JNI classes, iiuc. i'll get there :)
As part of trying to upgrade to the new GPL/LPGL ffmpeg preset, I'm seeing something weird with the licenses. Tools like cyclonedx seem to think the LGPL version is GPL. I see something similar at https://mvnrepository.com/artifact/org.bytedeco/ffmpeg/4.3.2-1.5.5
Not sure where it is getting the license info from - doesn't seem to be in the
.pom.
Probably from the parent pom.xml file here: https://github.com/bytedeco/javacpp-presets/blob/master/pom.xml#L14-L30