[Build] Failed to build labor sampling related code on windows_cu118

[Rhett-Ying]

🐛 Bug

To Reproduce

Builder: vs2019

Expected behavior

Environment

• DGL Version (e.g., 1.0): 1.1.x, master
• Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3):
• OS (e.g., Linux): Windows
• How you installed DGL (conda, pip, source):
• Build command you used (if compiling from source): vs2019
• Python version:
• CUDA/cuDNN version (if applicable): cuda11.8
• GPU models and configuration (e.g. V100):
• Any other relevant information:

Additional context

[Rhett-Ying]

@mfbalin Hi, have you ever tried to build on windows with cuda? Do you have any ideas on the build error I posted above?

[mfbalin]

Hi @Rhett-Ying, that line has a static_assert, static_assert with no error message in 2nd argument is available only in c++17. We can fix it by adding a second argument such as "alignment requirement is not met!".
https://en.cppreference.com/w/cpp/language/static_assert

[Rhett-Ying]

let me try this

[Rhett-Ying]

@mfbalin errors thrown with thrust still exist even resolve the static_assert. Maybe build with c++17 works.

[mfbalin]

I agree, it might help. We can attempt to upgrade to c++17 again.

[Rhett-Ying]

let me try to update thrust.

[Rhett-Ying]

after update thurst in #6139, many errors are gone.

those lines are added in #5648

[Rhett-Ying]

probably cuda driver needs to be updated as DeviceCopy was added 4 months ago.

any ideas on this? @mfbalin

[Rhett-Ying]

latest status: #6139 (comment)

[Rhett-Ying]

As a conclusion, let's wait the CCCL release to see if could resolve this issue. For now CCCL is setting up windows CI: NVIDIA/cccl#248.

[mfbalin]

I opened up an issue at the CCCL repository. NVIDIA/cccl#327

[mfbalin]

@Rhett-Ying The fix they made at the CCCL repository seems to have solved the issue. I can compile with the new CCCL branch/2.2.x branch now.

[mfbalin]

I will open a PR updating to the new version, we don't have to merge it now. We can merge it when they officially announce the new release of CCCL.

[Rhett-Ying]

@mfbalin really appreciate your help on this.

[github-actions]

This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you

[mfbalin]

@Rhett-Ying CCCL finally has a Windows CI, we can either wait for a new release, I don't know when that is though, or we can depend on the current head on the main branch that has Windows CI to fix the windows cuda compilation problems.

[Rhett-Ying]

Let's wait for a new release for now. If when we decide to cut a new release of DGL which incorporate this issue, let's depend on the head on the main branch to fix this issue.

[Rhett-Ying]

As mentioned in NVIDIA/cccl#327 (comment), CCCL 2.2 is ready. I will continue on this work item.

[mfbalin]

You need to use this branch, their official release doesn't contain the fix for the build error we encountered.
It is here: https://github.com/NVIDIA/cccl/tree/branch/2.2.x

[github-actions]

This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you