Core dump in EDDL 0.8a using pyEDDL 0.10.1

[diegobenedicto]

Describe the bug When I execute the python code attached (based on https://github.com/deephealthproject/use_case_pipeline/tree/3rd_hackathon/python) I get a core dump

To Reproduce DeepHealth libraries

• EDDL 0.8.a
• pyEDDL 0.10.1
• ECVL 0.3.0
• pyECVL 0.6.0

Use the file: code.zip

• Create a conda environment using the file attached environment.yml:

conda env create -f environment.yml

• Execute the python file attached using the conda environment:

  source activate <environment_name>
  python msseg08a.py

• A core dump happens: (p138-dh2-env-kdheddl08a) jovyan@jupyter-dhgpu3:~/projects/p138-dh2$ python msseg08a.py Illegal instruction (core dumped)

If I execute the python code using gdb (https://wiki.python.org/moin/DebuggingWithGdb) the error is: gdb -ex r --args python msseg08a.py

Thread 1 "python" received signal SIGILL, Illegal instruction. 0x00007fffd93d07a0 in Tensor::updateData(float, void, bool) () from /usr/local/lib/libeddl.so

This error comes from line in python code: in_ = eddl.Input([n_channels, size[0], size[1]])

Screenshots

[simleo]

Hi Diego,

I've seen the Illegal instruction problem several times recently while testing the latest Docker images. Every time the problem was that EDDL was being compiled with the new BUILD_HPC CMake flag, which is ON by default. That enables extra compiler flags like -march=native -mtune=native, which break the library if it's compiled on a machine and run on another. That's typically the case with Docker, where you build the image once and then run it somewhere else.

To avoid this problem, compile EDDL with -D BUILD_HPC=OFF.

This worked for me and hopefully should fix your issue as well.

[diegobenedicto]

I have compiled EDDL 0.8a with -D BUILD_HPC=OFF and now it works. Thanks Simone!

[jonandergomez]

Thanks Diego and Simone,

I understand it is not necessary yet I check it with C++.

Anyway, do not hesitate to ask Roberto, Salva or me any other problem you find.

Thanks and regards,

Jon

[salvacarrion]

Since the EDDL is intended for C++ users, we've set the -D BUILD_HPC flag to ON by default (this is due to the substantial increase in performance we get). Nevertheless, if you intend to compile the library for a different CPU or docker, you should disable it. To aid with the troubleshooting I've added a warning:

  [WARNING] The 'BUILD_HPC' flag is enabled and might not work on a different
  CPU.

  If you intend to compile the library for different CPUs (or for a docker
  image), you should disable it: '-D BUILD_HPC=OFF'.  This happens because it
  makes use of native compiling flags.

Ps.: If it gives users a lot of headaches, we'll disable it by default