Newbie wanna-be contributor setup

[viable-alternative]

I want to contribute as a coder. My PC has no GPU.

I have seemingly built CCCL with WSL and DevContainers. (Currently fails on test [10635/21696] .../thrust.omp.cuda.cpp17.test.copy.dir/copy.cu.o, which I suspect is due to absence of a GPU.)

Assertion: If one hypothetically has a change to submit, no mater how small, that person would want to run all tests, therefore needs a GPU?

If the above assertion is correct, what GPU would you recommend in order to be the most helpful to the project in terms of being able to work on the most needed features? My "volunteering budget" is under $1K. If compromises are to be made, I choose helpfulness to the project over build speed. The card will go into an ATX Windows box (AMD Ryzen 5 7600X). I will have no other use for that GPU, so CCCL contribution is the only deciding factor.

Secondly, is there a build setting to run "CPU only" tests, in order to defer the GPU procurement to the point when one has the actual code change, which would pass all CPU-only tests first?

Anything else I am missing in my setup?

Thank you for reading.

[miscco]

Hi @viable-alternative and thanks for your interest in contributing to CCCL.

In general it is beneficial to have a GPU to test locally, but that is by no means a necessary requirement. We have an incredibly broad support matrix both on the compiler and on the GPU side, so we do not expect a contributor to test all of them. However, it would be greatly appreciated if any changes would at least compile with one of the supported setups, as to not waste CI resources.

Also note that the requirements on a change depend on its scope. A simple typo or documentation fix doesnt need a full CI run. A change in a CUB algorithm does.

Regarding your test errors, it seems that the CUDA backend has been selected (which is the default) so the failure is expected. You can however configure Thrust to only build the c++ / OMP / TBB backend by setting the unavailable backends to false in the CMakePresets.json file. Those are

        "THRUST_MULTICONFIG_ENABLE_SYSTEM_CPP": true,
        "THRUST_MULTICONFIG_ENABLE_SYSTEM_CUDA": false,
        "THRUST_MULTICONFIG_ENABLE_SYSTEM_OMP": false,
        "THRUST_MULTICONFIG_ENABLE_SYSTEM_TBB": false,

[viable-alternative]

Thank you @miscco for the hints! Indeed, all-dev was the build target (due to the (still) limited understanding on my side).

Made the suggested config changes (there were two spots, all-dev and thrust-base). Let me go over the new build errors and try to decipher them. Is there a channel to ask questions about build errors?
(First error is seemingly when building thrust/benchmarks/bench/adjacent_difference/basic.cu.cpp "thread_operators.cuh:216:26: error: ‘using cub::CUB_300000_SM___CUDA_ARCH_LIST__::Sum = struct cuda::std::__4::plus’ is deprecated: use cuda::std::plus instead")

I suspect the easiest might be to "just" get a GPU and retry.

[bernhardmgruber]

Is there a channel to ask questions about build errors?

Feel free to open an issue on GitHub if you found a build error.

"thread_operators.cuh:216:26: error: ‘using cub::CUB_300000_SM___CUDA_ARCH_LIST__::Sum = struct cuda::std::__4::plus’ is deprecated: use cuda::std::plus instead"

This one apparently did not trigger anything in the CI. Here is the fix: #3918

[viable-alternative]

It looks like RTX 5070 Ti is the right choice, whenever they are back in stock? :D Just realized how crazy the GPU market now is!