About full understanding of message passing APIs

Questions
#1

Hi Everyone,
I am a Phd Student in South Korea and i am tasked to implement efficient message passing graph learning kernel for one of our lab project. from past few weeks i am reading DGL documents and reading the code. However, I I am facing difficulties understanding the src code in C for message passing APIs. I have also cloned the DGL github repository and is debugging the code line by line but i get stuck on the CAPI (C APIs). I would appreciate help for how to debug c code after the C wrappers called in python, Plus is there a way to run the c code in debug ? where currently it is called from dll? and any more helping material for understanding message passing will be really appreciated. Thanks in advance

#2

Hi @tariqaf, to debug C code warped, I recommend to use GDB as debugging tool , refer GDB for python for python side debugging. And you need to build DGL like this:

  1. Install DGL from source. And add CMAKE_BUILD_TYPE=DEBUG when building cmake files. E.g. CUDA debug build
cmake -DUSE_CUDA=ON -DCMAKE_BUILD_TYPE=DEBUG ..
  1. Set envs to use self-built lib:
export DGL_HOME=/path/to/your/dgl/clone
export DGL_LIBRARY_PATH=$DGL_HOME/build
export PYTHONPATH=$PYTHONPATH:$DGL_HOME/python
cd python
python setup.py build_ext --inplace

Details can be found here, refer Building and Testing section.

Below is a simple live debugging process, add it here for your reference.

python code → test.py

    ...
    dgl.load_graphs('graphs_t.bin')

C code
load_graphs call GetFileVersion() function in graph_serialize.cc

Steps

  1. Enter into gdb python mode
gdb python
  1. Set breakpoint
b GetFileVersion

or

b graph_serialize.cc:122
  1. run your script
r test.py
  1. Look up info, use bt to make sure you entered C code:
bt

And after that, you can use other gdb commands for debugging.

2 Likes
#3

This is very helpful. Thank you so much for the guidance. I will now try the same for setting the code and learn by debugging. Again Thank you so much.

1 Like
#4

Hi @a651694192,
sorry for inconvienience. I have followed all the steps for Install DGL from source. where i first forked the DGL and clone the forked repository as mentioned. After that i created build directory and call the same cmake command within build directory. After that using visual studio 17 2022 i compiled the dgl.sln and it builded sucessfully. After that when i call the setup.py install in python folder it gives warning that cython is not supported by windows, however it compiled but after that when i import dgl it says module not found? i think it is still not build and i dont know what i am doing wrong? maybe cython is must? or maybe python wrappers dont find the c build folder? i dont know what i am doing wrong here. i am surely doing some silly mistake over here.

#5

It looks your C lib has been built successfully but python installation failed, thus import DGL doesn’t work. For Cpython, you need to install it with the same compiler as python(Cpython install).
I suggest you uninstall everything firtst(DGL, Cpython or even python), and then reinstall them following official guide using same tools(VS2019 x64 Native tools recommend here)

1 Like
#6

Thanks it worked as you suggested. Previously i was running python 3.10 and vs 17 2022. Thanks again

1 Like
#7

Does this work for visual studio code in window if i installed gdb using minGW? i am trying but facing some issues? sorry again for any incovience.

#8

VS code now support minGW compile and gdb, so it should be supported. Refer Get Started with C++ and Mingw-w64 in Visual Studio Code for more information.

1 Like
#9

Thank You so much… I am following the link you mentioned.

#10

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.