feat: hermetic CUDA clang support #360
Conversation
9400206 to
8039d7b
Compare
| cccl_repo = components_mapping_compat.repo_str(repository_ctx.attr.components_mapping["cccl"]) | ||
|
|
||
| clang_cuda_path = repository_ctx.path("clang_compiler_deps") | ||
| repository_ctx.execute(["mkdir", "-p", "clang_compiler_deps"]) |
There was a problem hiding this comment.
I'm not happy that this is not hermetic. It uses native mkdir which will not work on windows. I'm going to try using rules_coreutils with toolchain_utils as suggested in the rules_coreutils README here. We can use hermetic mkdir and cp from there then.
@cloudhan would you be okay to adding these two dependencies (rules_coreutils and toolchain_utils)?
There was a problem hiding this comment.
I think it would be better to create a new repo call minimal_ctk with nvcc, cudart and cccl (and maybe curand if you need it). Duplicate all the files to minimal_ctk (instead of symlink) will solve you a lot of headache.
465f868 to
217a96b
Compare
…on.json and version.txt unnecessary Signed-off-by: udayaprakash <[email protected]>
217a96b to
b44e39d
Compare
| "%{clang_path}": clang_path_for_subst, # Will be empty if label is used | ||
| "%{clang_label}": clang_label_for_subst, # Will be empty if path is used | ||
| "%{cuda_path}": _to_forward_slash(cuda.path) if cuda.path else "cuda-not-found", | ||
| "%{cuda_path}": _to_forward_slash(cuda.path) if cuda.path else "external/rules_cuda++toolchain+cuda/clang_compiler_deps", # needs to be relative to work in RBE. |
There was a problem hiding this comment.
you cannot hard code +, the encoding is bazel version specific, they previously used ~ but find + to be much faster (suprise!) so they switched.
There was a problem hiding this comment.
You are warned to be prepared for the headache of apparent and canonical repo name, :)
Partially part of #283
This PR adds the missing piece of hermetic CUDA compilation with clang. The provided
compiler_depsrepo (alias cuda_nvcc) wasn't enough for clang to infer everything from. Typically, it's documented in multiple places ((1), (2)) that clang expects the entire CUDA toolkit and there is no recommended "minimal" CUDA setup. Nevertheless, I found thatcccl,nvcc,cudart,nvvmcomponents were enough.Surprisingly, I didn't need the
version.jsonorversion.txtfile anymore so I removed the logic to generate the rootversion.jsonfile.In
config_clangwe create a new folder calledclang_compiler_depswhere we collectbin,lib,includefiles from the required components and place in a single collected folder.We've been using this implementation for more than a month now with over 20+
cuda_libraryusages andcc_tests that run on NVIDIA GPU hermetically in RBE with CUDA 12.3 and LLVM 19.1.0