Accessing the hermetic downloaded CUDA root path from another CC toolchain?

[cemlyn007]

Greetings 💯

I have a little project, where I am writing a CC toolchain that will allow me to use Intel OneAPI SYCL, and would like to enable NVIDIA GPU support, and this repository is quite convenient because it can download the NVIDIA GPU dependencies. However does anybody know how to access the CUDA root path from within a custom CC toolchain? For example I could imagine it would be pretty neat if I could have one hermetic toolchain for SYCL, this rules_cuda hermetic toolchain, and blend the two together to enrich the SYCL toolchain.

I felt like the hacky solution would have been to load "find_cuda_toolkit" from "@rules_cuda//cuda/private:toolchain.bzl", however this yielded a cyclic error:

cemlyn@master:~/Development/hello-sycl$ bazel run main:hello-world
ERROR: /home/cemlyn/.cache/bazel/_bazel_cemlyn/ee030b46204fede619a1b165eb323152/external/bazel_tools/tools/cpp/BUILD:27:10: in alias rule @@bazel_tools//tools/cpp:current_cc_toolchain: cycle in dependency graph:
    //main:hello-world (a807b43c07fd2ad95f991746772b57f35d2bcb822e73c58dd9594bd80ce4ee6c)
    //main:hello-world (3723722276aac1314320ca9a9fb2a278d80e4eb3b422da8939e984231a537159)
.-> @@bazel_tools//tools/cpp:current_cc_toolchain (3723722276aac1314320ca9a9fb2a278d80e4eb3b422da8939e984231a537159)
|   @@rules_cc+//cc:current_cc_toolchain (3723722276aac1314320ca9a9fb2a278d80e4eb3b422da8939e984231a537159)
|   //toolchain:linux_x86_64_toolchain (3723722276aac1314320ca9a9fb2a278d80e4eb3b422da8939e984231a537159)
|   //toolchain:linux_x86_64_toolchain_config (3723722276aac1314320ca9a9fb2a278d80e4eb3b422da8939e984231a537159)
|   @@rules_cuda++toolchain+local_cuda//toolchain:nvcc-local (3723722276aac1314320ca9a9fb2a278d80e4eb3b422da8939e984231a537159)
`-- @@bazel_tools//tools/cpp:current_cc_toolchain (3723722276aac1314320ca9a9fb2a278d80e4eb3b422da8939e984231a537159)

Any ideas would be immensely appreciated! I'll keep digging on my side but otherwise maybe I should just copy the code that downloaded CUDA 🤷🏻‍♂️

[cloudhan]

to access the CUDA root path from within a custom CC toolchain?

Maybe use Label("@local_cuda_<component>//").workspace_root. This may not work well if you are using Module based project.

[cemlyn007]

Thank you for your reply, I got a bit stuck in the end though (I am still a Bazel novice lol) and am currently just letting sycl detect the CUDA library at runtime which seems to be working alright for now. My dream would be to have a hermatic sycl toolchain, for now I have a non-hermatic sycl toolchain, ideally I'll figure out how to nicely plugin the CUDA hermatic toolchain in the future as well! Hope you have a good week!

[cemlyn007]

(btw if you think this is the right answer then I am happy to mark it as answered)

[cloudhan]

I think this is still an open question, pls leave it open