Skip to content

[kitsune][cudaABI] Remove section names from fatbinary globals#56

Closed
jsarrao wants to merge 1 commit intolanl:dev/18.xfrom
jsarrao:cudaABI-change
Closed

[kitsune][cudaABI] Remove section names from fatbinary globals#56
jsarrao wants to merge 1 commit intolanl:dev/18.xfrom
jsarrao:cudaABI-change

Conversation

@jsarrao
Copy link
Collaborator

@jsarrao jsarrao commented Sep 12, 2024

Section names caused error with numba-mlir. They do not seem to be necessary with cuda/12.3. Tested with raytracer.

Section names caused error with numba-mlir. They do not seem to be necessary with cuda/12.3. Tested with raytracer.
@pmccormick
Copy link
Collaborator

pmccormick commented Sep 24, 2024

@jsarrao and @tarunprabhu -- I'm not sure this is a good idea. I think some of the command-line tools Nvidia provides as part of cuda might need section this to work (but I have not verified that). So, it isn't a matter of our codegen but more related to working well in the overall cuda ecosystem...

Can you provide more details about what was happening w/ numba-mlir?

@jsarrao
Copy link
Collaborator Author

jsarrao commented Sep 24, 2024

This was related to the relocation errors we were seeing if you recall those. But removing the section name stuff seems to fix the problem.

Here's the full error from simple vecadd:
python: /vast/home/josephsarrao/kitsune_t/llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp:299: void llvm::RuntimeDyldELF::resolveX86_64Relocation(const llvm::SectionEntry&, uint64_t, uint64_t, uint32_t, int64_t, uint64_t): Assertion `(Type == ELF::R_X86_64_32 && (Value <= UINT32_MAX)) || (Type == ELF::R_X86_64_32S && ((int64_t)Value <= INT32_MAX && (int64_t)Value >= INT32_MIN))' failed.

@tarunprabhu
Copy link
Collaborator

Not planned

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants