Questions about GADEN compilation

[chalan0830]

Thank you for your contributions again！！！

Problem Overview: Does the libbsc version seem outdated?

After pulling the project using git clone --recurse-submodules git@github.com:MAPIRlab/gaden.git, I compiled the entire project using colcon build --symlink-install, encountering two problems:

1. When compiling third_party/libbsc (a dependency of gaden_core), the build system attempted to set the GPU architecture to native. However, nvcc (the CUDA compiler) does not recognize the string "native"; it requires a specific numeric code.

Solution: I changed the "native" part in the cmakelist file under the libbsc folder to "86" --> 3080

2.Then I immediately recompiled, but encountered a second problem:

2: The problem lies with the CUB library (CUDA Unbound library) called by libcubwt.cu. See the image for details.

I have tried many solutions to this problem without success. I hope you can provide an answer when it is convenient for you.

My system version is Ubuntu 22.04, and my CUDA version is 12.1. My compilation command is:
colcon build --symlink-install --packages-select gaden_common --cmake-args -DCMAKE_CUDA_ARCHITECTURES=86 -DCMAKE_C_COMPILER=/usr/bin/gcc-10 -DCMAKE_CXX_COMPILER=/usr/bin/g++-10 -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-10

Thank you so much！！

[chalan0830]

The problem has been successfully resolved.

Regarding Problem 2: The error message indicated that cub::DeviceCopy::Batched points to a library called CUB.

However, the libcubwt library (inside libbsc), which GADEN depends on, has outdated code. The API it uses has changed in the newer versions of the CUDA-included CUB library or does not yet support this calling method.

Therefore, I changed the cub::DeviceCopy::Batched function to a simple loop copy, although this may result in a performance decrease.

The specific modifications are as follows:

Perhaps you have other, better solutions.