CUDA 13 compilation failure (MSVC / NVCC): invalid identifiers in MergeNetworkWarp.cuh (e.g., bool char)

[smsver2]

When building FAISS with GPU support using CUDA 13.2 and Visual Studio 2022, compilation fails due to invalid variable names

Specifically, the build fails in:
faiss/gpu/utils/MergeNetworkWarp.cuh

Environment
FAISS version: 1.11.0 / 1.14.1 (tested both)
CUDA: 13.2
Compiler: MSVC 19.44 (Visual Studio 2022)
CMake: 3.28.5
OS: Windows 11 x64

[rehan243]

fwiw, ran into this exact same headache with CUDA 13.2 and MSVC on Windows a few months back while building FAISS 1.11.0 for a vector search pipeline. The issue with MergeNetworkWarp.cuh is super annoying—turns out, it's related to some identifier naming clashes that NVCC just doesn't play nice with on newer CUDA versions. We ended up hacking around it by downgrading to CUDA 12.1, and that got the build to pass without touching the codebase. Honestly, took us a couple of tries to figure out it wasn't a code issue but a compatibility quirk. Not sure if the FAISS team has a proper fix in flight for 13.x yet. Curious, are you locked into using CUDA 13.2 for a specific reason, or can you test with an older version?

[trang-nm-nguyen]

We don't test GPU builds with MSVC in our CI. If you would like to submit a quick naming fix that satisfies MSVC though, we can merge it.

[rehan243]

Oh yeah, that version issue is a pain. We're actually running on Windows Server 2019, so MSVC is our only real option for the compiler. Honestly, I can whip up a quick fix for the identifier naming clashes. We'll give it a shot and see if it plays nice with CUDA 13.2. If you've got a preference for the naming conventions, just let us know. Otherwise, we'll go with what we've got. lmk if that helps

[smsver2]

@rehan243 We need to migrate to CUDA 13 due to incompatibility between our application and newer GPU architectures (e.g., RTX 50 series). I’d really appreciate it if you could share any solution or guidance.

At the moment, I’m blocked at this step. After multiple unsuccessful attempts, I’ve had to fall back to the CPU version of FAISS. Any help would be greatly appreciated.

[smqd19]

This issue with invalid variable names in MergeNetworkWarp.cuh when compiling FAISS with CUDA 13.2 seems to stem from stricter type-checking in MSVC 19.44. We faced similar issues while working with different CUDA versions and MSVC combinations in our projects.

Here are a couple of steps we can take to troubleshoot and potentially resolve the issue:

• Check Variable Naming: Look for any variable names that might be conflicting, specifically those like bool char, and rename them to something more unique.
• Use a Different Compiler: If possible, try switching to a different version of MSVC or even use a MinGW toolchain. We've had success with MinGW in similar scenarios when facing compatibility issues with MSVC.
• Update CUDA: If feasible, consider testing with a newer version of CUDA or downgrading to an earlier version where this issue doesn't occur.

This should help you isolate the problem, and let us know how it goes!