[SYCL][CUDA][BUILD] cuPTI error while building the compiler with CUDA backend

[TejaX-Alaghari]

Bug Description
Build failure while compiling unified-runtime-src\source\adapters\cuda\tracing.cpp

To Reproduce
python .\llvm\buildbot\configure.py --cuda -o /path/to/install/dir
python .\llvm\buildbot\compile.py -o /path/to/install/dir

Environment:

• OS: Windows 10
• Target device and vendor: Nvidia GPU
• Dependencies version: CUDA Toolkit - 12.0 or 12.3

Additional Context
Complete error log: LLVM_CUDA_win10_build_log.txt

[rafbiels]

The issue is all the function pointer typedefs here:
https://github.com/oneapi-src/unified-runtime/blob/main/source/adapters/cuda/tracing.cpp
are of the format:

using foo_fn = CUPTIAPI CUptiResult (*)(args);

but should be of the format:

using foo_fn = CUptiResult (CUPTIAPI *)(args);

because CUPTIAPI expands to __stdcall on Windows and this needs to be placed after the return type. There is an example of a function pointer typedef in the __stdcall docs.

Here's a proof the other syntax should work: https://godbolt.org/z/Pd3Ess79G

I don't currently have a Windows machine to test the fix in a full build, so I'd prefer someone who can test it makes a PR. Tagging @kbenzie for UR.

[kbenzie]

Tagging @npmiller who's team owns the CUDA adapter.

[TejaX-Alaghari]

I have tested the suggestion on Win 10 with 12.0 toolkit and I can confirm it works.

Thanks @rafbiels