Update cmake with native cuda support #4316
Conversation
@RAMitchell Yap. There were two concerns around this. Other than the one you mentioned above, newer version of CMake is rarely packaged in distributions. Here is a search result from Ubuntu 16.04LTS: https://packages.ubuntu.com/search?suite=xenial&searchon=names&keywords=cmake So we might need to think it through ... |
|
@RAMitchell I have a branch for a complete rewrite of CMake: https://github.com/trivialfis/xgboost/tree/cmake . I don't want to pull it in because it still uses old CUDA wrapper. But if we were to bump the CMake version, I would like to merge that branch here, and enable distributing C-APIs along the way. |
|
@trivialfis I think I am ok with Ubuntu 16.04LTS users having to upgrade their cmake. We have prepackaged binary releases including nccl now so I think building from source is less common. What do you want me to do with this PR? Should I finish it or would you prefer to do it yourself as a part of the cmake rewrite? |
|
@RAMitchell I would like to make it part of my rewrite. |
|
@RAMitchell Feel free to reopen. |
FindCuda.cmake is deprecated, making it a good time to switch.
Native cuda support in cmake allows me to enable compilation dynamic parallelism in future (which I couldn't do with old cuda cmake) as well as enables multithreaded compilation on Windows, saving me a lot of development time.
@trivialfis I think we discussed this in the past and I was hesitant to switch because of problems, but this time when I tried it everything went smoothly so I guess I just did something wrong last time.