[CUDA 12] clang error: no template named 'texture'

[lahwaacz]

For a trivial foo.cu:

int main() {}

Running clang++ foo.cu results in:

clang-15: warning: CUDA version is newer than the latest supported version 11.5 [-Wunknown-cuda-version]
In file included from <built-in>:1:
In file included from /usr/lib/clang/15.0.7/include/__clang_cuda_runtime_wrapper.h:365:
/usr/lib/clang/15.0.7/include/__clang_cuda_texture_intrinsics.h:696:13: error: no template named 'texture'
            texture<__DataT, __TexT, cudaReadModeNormalizedFloat> __handle,
            ^
/usr/lib/clang/15.0.7/include/__clang_cuda_texture_intrinsics.h:709:13: error: no template named 'texture'
            texture<__DataT, __TexT, cudaReadModeElementType> __handle,
            ^
2 errors generated when compiling for sm_35.

The platform is Arch Linux:

$ clang++ --version
clang version 15.0.7
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Tue_Feb__7_19:32:13_PST_2023
Cuda compilation tools, release 12.1, V12.1.66
Build cuda_12.1.r12.1/compiler.32415258_0

[Artem-B]

I believe that's already been fixed in recent clang: 8173405

[lahwaacz]

@Artem-B Thanks for the info. Any chance of backporting it to clang 15, or do I have to wait for version 16 to be released? What is the general policy?

[Artem-B]

It's a low-risk fix which should be OK to cherry-pick, though I don't know the current state of the release process.

[Artem-B]

/cherry-pick 8173405

[llvmbot]

/cherry-pick 8173405

Error: Command failed due to missing milestone.

[Artem-B]

/cherry-pick 8173405

Now, with the 16.0.1 milestone. @tstellar - FYI.

[llvmbot]

Failed to cherry-pick: 8173405

https://github.com/llvm/llvm-project/actions/runs/4409411801

Please manually backport the fix and push it to your github fork. Once this is done, please add a comment like this:

/branch <user>/<repo>/<branch>

[llvmbot]

Failed to cherry-pick: 8173405

https://github.com/llvm/llvm-project/actions/runs/4409418263

Please manually backport the fix and push it to your github fork. Once this is done, please add a comment like this:

/branch <user>/<repo>/<branch>

[Artem-B]

Oh, we don't need any cherry-picks. The change is already in 16.x.

[lahwaacz]

@Artem-B What about 15.x? Or is 16.0 expected to come out soon?

[Artem-B]

16.0 should be released soon, though I don't know the exact date.
I'm not sure it's worth bothering with 15.x, cuda-12 needs other changes that would not be suitable for backporting two major releases back.

[milthorpe]

I see the same error with Clang 17.0.6 and CUDA 12.4:

~/source/chapel-2.0$ clang++ --cuda-gpu-arch=sm_60 foo.cu
clang++: warning: CUDA version is newer than the latest partially supported version 12.1 [-Wunknown-cuda-version]
In file included from <built-in>:1:
In file included from /usr/lib/llvm-17/lib/clang/17/include/__clang_cuda_runtime_wrapper.h:365:
/usr/lib/llvm-17/lib/clang/17/include/__clang_cuda_texture_intrinsics.h:697:13: error: no template named 'texture'
  697 |             texture<__DataT, __TexT, cudaReadModeNormalizedFloat> __handle,
      |             ^
/usr/lib/llvm-17/lib/clang/17/include/__clang_cuda_texture_intrinsics.h:710:13: error: no template named 'texture'
  710 |             texture<__DataT, __TexT, cudaReadModeElementType> __handle,
      |             ^
2 errors generated when compiling for sm_60.
~/source/chapel-2.0$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2024 NVIDIA Corporation
Built on Thu_Mar_28_02:18:24_PDT_2024
Cuda compilation tools, release 12.4, V12.4.131
Build cuda_12.4.r12.4/compiler.34097967_0
~/source/chapel-2.0$ clang++ --version
Ubuntu clang version 17.0.6 (++20231208085846+6009708b4367-1~exp1~20231208085949.74)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

[Artem-B]

In file included from /usr/lib/llvm-17/lib/clang/17/include/__clang_cuda_runtime_wrapper.h:365:
/usr/lib/llvm-17/lib/clang/17/include/__clang_cuda_texture_intrinsics.h:697:13: error: no template named 'texture'
697 | texture<__DataT, __TexT, cudaReadModeNormalizedFloat> __handle,
| ^
/usr/lib/llvm-17/lib/clang/17/include/__clang_cuda_texture_intrinsics.h:710:13: error: no template named 'texture'
710 | texture<__DataT, __TexT, cudaReadModeElementType> __handle,
| ^

That code is guarded with #if CUDA_VERSION < 12000 and should not be seen by the comiler:

llvm-project/clang/lib/Headers/__clang_cuda_texture_intrinsics.h

Line 669 in 026a29e

#if CUDA_VERSION < 12000

My guess is that something is broken w/ CUDA installation in your setup and clang picks up wrong/inconsistent CUDA headers somehow.

Considering that you didn't specify --cuda-path explicitly, you're at the mercy of what ever CUDA packages you have installed on your system.

I would suggest running compilation with -v and check the CUDA include path clang ends up using.
You may also want to check preprocessed output and see where CUDA_VERSION comes from (compile with -dD -E --cuda-host-only)

[milthorpe]

Thanks Artem! Your suggestion to compile with -dD -E --cuda-host-only was very helpful: it showed that my system had a rogue symlink from /usr/include/cuda.h to a CUDA 11 version of the header. I removed this, after which the CUDA 12 version was included as expected, and I no longer see the error.