Is there a way to cmake-use thrust library

[correaa]

I know Thurst is a header only library, however it is convenient for it still to work with Cmake.
In turn, to cmake-install Thrust it requires CUDA (CUDA compiler).

That is, if I do mkdir build; cd build; cmake ..; it will complain that CUDA is not present.

Is there a way to install Thrust without CUDA?

This is of course to use Thrust in a system without CUDA, but where the idea is to use Thrust-OMP or Thrust-TBB. (It is also a Mac).

[bernhardmgruber]

Hi! Thrust is part of CCCL so you can and should only install the entire CCCL library (Thrust needs libcu++ and CUB internally for example).

if I do mkdir build; cd build; cmake ..; it will complain that CUDA is not present.

I would agree that installing Thrust without a CUDA compiler would be convenient. Maybe @alliepiper can help whether this is possible and if not, how hard it would be to make it possible.

[correaa]

Yes, it is a kind of inconsistent; Thrust is header-only and has backends that are not related to CUDA.

I managed to do this partially, with the archived version, of standalone Thrust, by disabling (through the cmake command line) all tests and examples.

I have to see if I can do it with CCCL, it looks harder.

[alliepiper]

Try using the installation instructions here: https://github.com/NVIDIA/cccl?tab=readme-ov-file#installation

If you still have issues with those, please open an issue and ping me on it.

[correaa]

Thank you @alliepiper , I will try that.

In the meantime this workflow is working with Old Thrust:

git clone --recursive https://github.com/NVIDIA/thrust.git    
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$HOME -DTHRUST_ENABLE_HEADER_TESTING=OFF -DTHRUST_ENABLE_TESTING=OFF -DTHRUST_ENABLE_EXAMPLES=OFF
make install
...
use CMAKE_PREFIX_PATH=$HOME/lib/cmake/thrust  # setting Thrust_DIR didn't work

This allows me to test Thrust with OMP in Mac (not that it is my main platform, but it helps).
(it is a pity that Thrust is not part of brew or vcpkg either)

[bernhardmgruber]

For CCCL (which is what Thrust+CUB+libcu++ have become) we generally recommend to either take it directly from the CTK (every CUDA toolkit bundles CCCL), or download the latest version from GitHub (using CPM, or CMake fetch_content, or whatever).

I was a heavy user of vcpkg during my PhD and it helped me a lot getting third-party dependencies installed on various clusters. So if you say it would help you tremendously, we can consider adding CCCL to vcpkg. I have added a bunch of libraries myself and for header-only ones with no dependencies, it's almost trivial.

[correaa]

I use vcpkg as a last resort, and on Github Windows CI.
In this particular case, what it would have been convenient really was to get a version Thrust from Mac OS brew. (not sure who maintains brew, I am new to Mac)

brew install thrust

in the same way to Ubuntu/Debian have package for standalone Thrust.

apt install libthrust-devel

I am not sure what is the future of the standalone Thrust package now that is is part of CCCL. (apt install libcccl?)

FetchContent is more or less working for me now

[correaa]

Thank you again @alliepiper ,

The instructions at https://github.com/NVIDIA/cccl?tab=readme-ov-file#installation worked!

CCCL is a bit more download than Thrust, but other than that it worked (on Mac).
No need to set up variables.

git clone https://github.com/NVIDIA/cccl.git
cd cccl
cmake --preset install -DCMAKE_INSTALL_PREFIX=$HOME
cd build/install
ninja install
...
use CMAKE_PREFIX_PATH=$HOME/lib/cmake/thrust  # setting Thrust_DIR didn't work

[alliepiper]

use CMAKE_PREFIX_PATH=$HOME/lib/cmake/thrust # setting Thrust_DIR didn't work

Try setting CCCL_DIR to $HOME/lib/cmake/cccl, that should work IIRC.

[correaa]

Just wanted to add that I needed to add some more things to my CML to make it work.
Also that I can't compile in Debug mode on Mac #5186

namely,

find_package(Thrust REQUIRED CONFIG)
thrust_create_target(Thrust)

    thrust_create_target(ThrustOMP DEVICE OMP)

    add_executable(omp.cpp.x omp.cpp)

    target_link_libraries(omp.cpp.x PRIVATE multi Boost::boost OpenMP::OpenMP_CXX ThrustOMP)

[pauleonix]

https://github.com/NVIDIA/cccl/blob/main/examples/thrust_flexible_device_system/CMakeLists.txt might be of interest.

[correaa]

Further improvement,

Since I compile my project with warnings to the max (clang-tidy), I had do target_include_directories( ... SYSTEM ${_THRUST_INCLUDE_DIR}). (ThrustOMP_INCLUDE_DIR didn't work)

find_package(Thrust REQUIRED CONFIG)
thrust_create_target(Thrust)

thrust_create_target(ThrustOMP DEVICE OMP)

add_executable(omp.cpp.x omp.cpp)

target_include_directories(omp.cpp.x SYSTEM PRIVATE ${_THRUST_INCLUDE_DIR})
target_link_libraries(omp.cpp.x PRIVATE multi Boost::boost OpenMP::OpenMP_CXX ThrustOMP)

[alliepiper]

It shouldn't be necessary to use target_include_directories at all, the include paths will be set up when target_link_libraries is used to link the ThrustOMP target into your executable. You should be able to omit that line entirely.

Also be aware that the old Thrust repo is very very out of date at this point and will not be updated in the future. It only exists for legacy purposes, and all new projects should be using CCCL.