cmake fails when installing OpenJij with CUDA/GPU

[italogs]

Describe the bug
I am unable to conclude the installation of OpenJij with CUDA/GPU.
The error log seems quite complicated for me to understand. It seems that cmake cannot compile openjij and jij-cimod (maybe due to external dependencies).
Conversely, installing OpenJij without CUDA/GPU works fine.

To Reproduce
Steps to reproduce the behavior, as in the README.md:

1. pip install -U cmake
2. pip install --no-binary=openjij,jij-cimod openjij

Expected behavior
Installation to conclude without errors.

Screenshots
There are two screenshots attached for better clarity.

Environment (please complete the following information):

OS: Ubuntu
Version: Python 3.7
CUDA: 11.7
cmake: 3.22.6
gcc: 9.4.0
OpenJij: 0.6.15 (latest)

Additional context
If anyone has it correctly installed, please, show your current environment so I can go towards a promising avenue :)

Thanks.

[29rou]

Thank you for reporting the build error, we will take care of it.
I would like to ask you to try if possible, is it possible to Git Clone and build the package?
Sincerely.

[29rou]

A corrected version of MANIFEST.in has been released.
We believe that in the previous version, the required source code was not included in the package and an error occurred.

[j-i-k-o]

Please note that in the current version of OpenJij. It is recommended to use OpenJij with CPU mode since the performance is not so changed compared to GPU.

[dvnoble]

Thank you for your prompt response, j-i-k-o. Regarding Issue #360, I've found that the version I tested already includes the fix in the MANIFEST of commit a9753d9. However, I'm still experiencing issues with the Cuda files, specifically an error message stating that "error: 'info' is not a constant expression." If it would be helpful, I can provide the full log for your review. Thank you again, and best regards.

[italogs]

Thanks, @29rou and @j-i-k-o, for the quick answers.
The purpose of installing all OpenJij is to test if GPU-based code can surpass CPU-based one. However, as indicated by @j-i-k-o, GPU-based one has not much advantage yet.

After this new round of updates, I am still having problems with cmake.
If any of you has a working version of GPU-based, could you export all needed environment information and libraries (e.g., cmake version, CUDA version, SO, GCC version, and so on)?
This could help us (as users) a lot in investigating how to properly compile the code on our side.
Thanks in advance!

[hardy-science]

I too failed in my attempt to install Openjij with CUDA support. An error is reported when compiling kernel.cu (error: ‘info’ is not a constant expression). I would like to try to Git clone and build the package, but my question is, would that install Openjij with CUDA support?
Many Thanks for your answer in advance.

[czachow]

Hi! I also ran into this problem when trying to install openjij with cuda support. Any news about this error?

[j-i-k-o]

Thanks for the comments.
We recently removed cuda support from OpenJij due to the following reasons:

• Current Cuda support is only for specific types of graphs, which is not suitable for practical applications.
• Current CPU implementation usually outperforms the GPU implementation.
  delete GPU implementation #387

Please try the installation process using the latest main branch.
Many thanks.

[czachow]

Dear Jiko, thanks a lot for your reply. I will test the new version of openjij. Best, Christopher Von: jiko ***@***.***> Antworten an: OpenJij/OpenJij ***@***.***> Datum: Mittwoch, 9. August 2023 um 15:45 An: OpenJij/OpenJij ***@***.***> Cc: "Zachow, Christopher" ***@***.***>, Comment ***@***.***> Betreff: Re: [OpenJij/OpenJij] cmake fails when installing OpenJij with CUDA/GPU (Issue #360) Thanks for the comments. We recently removed cuda support from OpenJij due to the following reasons: * Current Cuda support is only for specific types of graphs, which is not suitable for practical applications. * Current CPU implementation usually outperforms the GPU implementation. #387<#387> Please try the installation process using the latest main branch. Many thanks. — Reply to this email directly, view it on GitHub<#360 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ADBLIGG3WA5XNEXZA2Z2AU3XUOH63ANCNFSM6AAAAAAV3YZCEU>. You are receiving this because you commented.Message ID: ***@***.***>