Crash on Windows

[maleadt]

Using Julia 1.11, NVTX#master:

$ julia --project -e 'using NVTX; NVTX.activate()'

Please submit a bug report with steps to reproduce this fault, and any error messages that follow (in their entirety). Thanks.
Exception: EXCEPTION_ACCESS_VIOLATION at 0x6be45c40 -- nvtxGlobals_v3 at C:\Users\Tim\.julia\artifacts\b4eeaf094ffb6aacf1b20ee5d2ac9aa1818fc732\bin\libnvToolsExt.dll (unknown line)
in expression starting at none:1
nvtxGlobals_v3 at C:\Users\Tim\.julia\artifacts\b4eeaf094ffb6aacf1b20ee5d2ac9aa1818fc732\bin\libnvToolsExt.dll (unknown line)
Allocations: 2905 (Pool: 2896; Big: 9); GC: 0

[maleadt]

It's the initialization that crashes, so I guess there's something wrong with the JLL:

$ julia --project -e 'using NVTX_jll; ccall((:nvtxInitialize, libnvToolsExt), Cvoid, (Ptr{Cvoid},), C_NULL)'

Please submit a bug report with steps to reproduce this fault, and any error messages that follow (in their entirety). Thanks.
Exception: EXCEPTION_ACCESS_VIOLATION at 0x6be45c40 -- nvtxGlobals_v3 at C:\Users\Tim\.julia\artifacts\b4eeaf094ffb6aacf1b20ee5d2ac9aa1818fc732\bin\libnvToolsExt.dll (unknown line)
in expression starting at none:1
nvtxGlobals_v3 at C:\Users\Tim\.julia\artifacts\b4eeaf094ffb6aacf1b20ee5d2ac9aa1818fc732\bin\libnvToolsExt.dll (unknown line)
Allocations: 2905 (Pool: 2896; Big: 9); GC: 0

[simonbyrne]

Unfortunately I don't have a Windows machine to test it out.

[maleadt]

I'm not really familiar with Windows either, i.e., how to best debug this. The backtrace is sparse, on GCC 8+ it's absent (just a ReadOnlyError), compiling with -g gives a DLL that Windows refuses to load, and compiling with -gcoff + WinDBG doesn't give any debug info at all...
I'll just disable NVTX support in CUDA.jl for now. I've also filed a bug with NVIDIA.

[mkitti]

I have a Windows computer, and I have some ability to help debug this sort of thing. I don't really know what NVTX is though. It looks like reproduction is pretty straightforward though.

[maleadt]

I have a Windows system too, I'm just entirely unfamiliar with how to debug crashes like this on Windows. If you'd know how: this is how we build nvToolsExt.dll, https://github.com/JuliaPackaging/Yggdrasil/blob/master/N/NVTX/build_tarballs.jl, and doing a simple ccall to initialize the library crashes. The sources of the library can be found at https://github.com/NVIDIA/NVTX/. I'm guessing there may be some incompatibility between the library and mingw (it does pretty low-level things).

[simonbyrne]

I don't really know what NVTX is though.

It's an instrumentation library, which the profiler then intercepts. In C it is used as a header-only library, but we build it as a dynamic library, which may be where the issue is.

The source is here: https://github.com/NVIDIA/NVTX/

[huiyuxie]

Looks like only one simple hotfix has been done - is anyone already working on this issue/planning to take over this issue?

[maleadt]

I've reproduced this in C, and filed an issue with NVIDIA. I'll let you know when I hear back.

[huiyuxie]

Thanks 👍