When performing device only compilation, there was an issue where
cubin outputs were being renamed to cubin despite the user's name.
This is required in a normal compilation flow as the Nvidia tools only
understand specific filenames instead of checking magic bytes for some
unknown reason. We do not want to perform this transformation when the
user is performing device only compilation.
Details
- Reviewers
tra - Commits
- rG3b52341116b7: [CUDA] Fix output name being replaced in device-only mode
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
clang/test/Driver/cuda-bindings.cu | ||
---|---|---|
148 | Ideally we want to test that the output file actually survives the compilation. It's traky to do for in-tree tests as we don't have ptxas. However, we should be able to add a script which would just touch whatever is specified with -o argument and ignore the rest and then actually do run clang with --ptxas-path=/path/to/dummy-ptxas and then verify file existence. |
clang/test/Driver/cuda-bindings.cu | ||
---|---|---|
148 | Yeah, I wasn't sure if there was a good way. I don't have any CUDA machines to test so I just went with the simplest version. |
clang/test/Driver/cuda-bindings.cu | ||
---|---|---|
148 | This is a more general problem, not specific to this test -- we have other tests where we assume that if something is printed by cc -###, then it must be real. This patch is good to go as is. |
clang/test/Driver/cuda-bindings.cu | ||
---|---|---|
148 | Okay, I'll land it soon. |
Ideally we want to test that the output file actually survives the compilation. It's traky to do for in-tree tests as we don't have ptxas. However, we should be able to add a script which would just touch whatever is specified with -o argument and ignore the rest and then actually do run clang with --ptxas-path=/path/to/dummy-ptxas and then verify file existence.