Missing include on windows in targets\x64\nvtx3\nvtxDetail\nvtxInit.h

[h-vetinari]

In conda-forge/kaldi-feedstock#48, I'm adding CUDA 12.0 support, and after finding where nvToolsExt.h lives and adding it as a host dependence, I get:

%LIBRARY_INC%\targets\x64\nvtx3\nvtxDetail\nvtxInit.h(164): error C3861: '_wgetenv': identifier not found

I can probably get around this by including <stdlib.h> wherever relevant headers get included, but IWYU would dictate that this should actually be done correctly in nvtxInit.h itself. For reference, here are the MSVC docs for _wgetenv.

[h-vetinari]

@conda-forge/cuda-nvtx, could we carry a patch for NVIDIA/NVTX#102 that simply does:

--- a/c/include/nvtx3/nvtxDetail/nvtxImpl.h
+++ b/c/include/nvtx3/nvtxDetail/nvtxImpl.h
 #if defined(_WIN32)  
  
 #include <Windows.h> 
+#include <stdlib.h> 
  
 #else 

This should be completely benign.

[h-vetinari]

@conda-forge/cuda-nvtx @jakirkham
Does anyone mind if I create a branch here for CUDA 12.0? We need to backport NVIDIA/NVTX@909007c to be able to compile kaldi with CUDA 12.0

[jakirkham]

Upstream has this fix, as noted above, in issue ( NVIDIA/NVTX#102 ). Once that makes its way through the release process it will make its way here

For now, we opted to add this patch to the nvtx-c package with PR ( conda-forge/nvtx-c-feedstock#9 ). So users needing this fix are encouraged to use the nvtx-c package

[h-vetinari]

Do you want to leave this issue open in case anyone else hits it?

[jakirkham]

Yeah think that is ok

We probably also want to track when that fix makes it in here and note the version

So keeping it open can serve multiple purposes