How to load this in a custom pipeline?

[chriswiggins]

We're trying to load this into our Python pipeline, however using LD_PRELOAD doesn't seem to have an effect. I've tried to load the plugin through the GstRegistry, but this has no effect?

Thanks for the help

[kirushyk]

On Linux, if you have gst-instruments installed to /usr/local/ prefix, you have two options:

1. GStreamer ABI call interception:

• set LD_PRELOAD environment variable to /usr/local/lib/libgstintercept.so.
• set GST_DEBUG_DUMP_TRACE_DIR to a folder where to put output trace file.
• set GST_DEBUG_DUMP_TRACE_FILENAME to a filename you would like to use for your trace.

2. Hooks with GStreamer plugin:

• set GST_PLUGIN_PATH to /usr/local/lib/gstreamer-1.0/ (prefix will be taken into account when choosing installation location for plugin)
• set GST_TRACERS to instruments (filename of plugin is libgstinstruments.so, but not libgstintercept.so).
• filename will be constructed from top-level bin and .gsttrace suffix.

[jjrugui]

@kirushyk I followed your instructions in the previous comment (using intercept) and I can't get the trace. I can confirm that I get a trace and that I'm able to get the visualization if I use it on gst-launch.

When using Python, does the pipeline need to be closed gracefully in a specific way in order to get the trace?

[yuyou]

@jjrugui I managed to make it work using the 2nd method:

GST_TRACERS=instruments python3 <your_GST_PY_Code> 

And gst-instruments was installed to the same GST plugin directory, e.g. "/usr/lib/x86_64-linux-gnu/gstreamer-1.0/libgstinstruments.so"

[jjrugui]

@yuyou awesome! Thanks for the input

I'll be able to test it in ~30 mins. Will report back :)

[jjrugui]

@yuyou @kirushyk If I use the plugin method I get a segmentation fault. I can see though from the debug that the instruments library is loaded correctly.

Any idea on this? I'll try to investigate the issue and report it back

[jjrugui]

Note that I'm working on a Nvidia Xavier

[yuyou]

I am not sure if it would be helpful but my GST was built from source (1.18.3) using meson with release=debug.
Is it required to have a GST in debug build or works with the release version?

[jjrugui]

@yuyou how complex is the pipeline you debugged? Might be a stupid question (I'm pretty new to gstreamer) but are all types of pads compatible with the profiler? @kirushyk

In my case I'm running a fairly complex pipeline with a couple appsink/appsrc so I don't know if some of the problem comes from there. When done with gst-launch I have no problem with getting the trace.

[yuyou]

@jjrugui Hard to define the exact "complex" level :-O) but my pipeline includes elements like udpsrc, glupload/gldownload, and glshader for GLSL shading.

I tested with "appsrc" and it also worked.

The step are as follows:

1. use a sample code from appsrc
2. The command: GST_PLUGIN_PATH=/usr/lib/x86_64-linux-gnu/gstreamer-1.0/ GST_TRACERS=instruments GST_DEBUG_DUMP_TRACE_DIR=. python3 run_appsrc.py
3. The tracing output is stored at "pipeline0.gsttrace" file and you can use "gst-report" to view it.

ELEMENT                              %CPU   %TIME   TIME
videoconvert0                         40.3   95.5    1.35 s
appsrc0                                1.0    2.3   33.0 ms
autovideosink0-actual-sink-xvimage     0.7    1.7   23.8 ms
queue0                                 0.2    0.5   7.26 ms
autovideosink0                         0.0    0.0      0 ns
pipeline0                              0.0    0.0      0 ns

OS: Ubuntu 18.04
GST: 1.18.3

[jjrugui]

@yuyou yeah, sorry, not the best phrasing with "complex". I meant as in custom functions in an appsink, so I wasn't sure if that could be an issue but it seems from your test that it's not.

I'll keep investigating and try your suggestions, so as soon as I have some insight I'll share it here.

Thank you very much for your time, highly appreciated!