FBGEMM_GPU: Support for Clang on CUDA?

[BwL1289]

Is there a plan or ETA to support FBGEMM_GPU compiling with clang?

See: #2248 (comment)

[q10]

We do support FBGEMM_GPU compilation with clang now - see "C/C++ Compiler (Clang)" section in https://docs.pytorch.org/FBGEMM/fbgemm_gpu/development/BuildInstructions.html.

The instructions are dependent on having an isolated conda environment inside a container, and might be incomplete, but the build scripts hosted in https://github.com/pytorch/FBGEMM/blob/main/.github/scripts are the most up to date and should be working, as our CI runs depend on being able to build FBGEMM_GPU using clang, As of current, the release packages are still still built using gcc.

[BwL1289]

Thanks. Is this section still relevant for builds with CUDA?

From the build scripts it looks that way. Is there a plan to support pure clang builds for CUDA?

Note that for CUDA code compilation, even though nvcc supports Clang as the host compiler, only libstd++ (GCC’s implementation of the C++ standard library) is supported for any host compiler being used by nvcc.

This means that GCC is a required dependency for CUDA variant of FBGEMM_GPU, regardless of whether it is built with Clang or not. In this scenario, it is recommended to first install the GCC toolchain before installing the Clang toolchain in this scenario; see C/C++ Compiler (GCC) for instructions.

[q10]

Thanks. Is this section still relevant for builds with CUDA?

From the build scripts it looks that way. Is there a plan to support pure clang builds for CUDA?

Note that for CUDA code compilation, even though nvcc supports Clang as the host compiler, only libstd++ (GCC’s implementation of the C++ standard library) is supported for any host compiler being used by nvcc.
This means that GCC is a required dependency for CUDA variant of FBGEMM_GPU, regardless of whether it is built with Clang or not. In this scenario, it is recommended to first install the GCC toolchain before installing the Clang toolchain in this scenario; see C/C++ Compiler (GCC) for instructions.

Yes, this section is still relevant. We've found many strange compilation errors that ultimately come from the fact that libc++ is not quite the same as libstdc++. I recall that Nvidia has placed the disclaimer somewhere in its docs or forum responses that it is possible to use clang to compile, but libstdc++ is still needed.

[BwL1289]

Got it. So it's not that there's specific libstdc++ code in FBGEMM_GPU, but this is an overarching comment based on this comment from the nvidia docs.

FWIW, we've been able to compile cuda code for nvidia libraries (i.e. nccl, cudnn_frontend) using pure clang (and no libstdc++). That said, it's not trivial typically.

[q10]

Got it. So it's not that there's specific libstdc++ code in FBGEMM_GPU, but this is an overarching comment based on this comment from the nvidia docs.

FWIW, we've been able to compile cuda code for nvidia libraries (i.e. nccl, cudnn_frontend) using pure clang (and no libstdc++). That said, it's not trivial typically.

That is correct (thanks for linking to the Nvidia docs). If you have docs and/or patches for building purely in clang without libstdc++, please don't hesitate to send us docs and pull requests, respectively, though I think the endeavor might be difficult, considering that our biggest dependency, libtorch, might also be using libstdc++.

[BwL1289]

Thanks. We believe it's possible. Will report back.

While I have you, any chance there's an update for #2074?