LLVM OpenMP GPU Offloading Docker Image

Question

I've been trying to get clang OpenMP GPU offloading working in the Docker Image and it builds fine at first (cmdline:

~/llvm_project/llvm/utils/docker:$ bash build_docker_image.sh 
--source nvidia-cuda 
--docker-repository clang-cuda --docker-tag "latest" 
-p clang -i stage2-install-clang -i stage2-install-clang-resource-headers 
-- 
-DLLVM_TARGETS_TO_BUILD="host;NVPTX" 
-DCMAKE_BUILD_TYPE=Release 
-DLLVM_ENABLE_RUNTIMES="openmp;offload" 
-DLIBOMPTARGET_DEVICE_ARCHITECHTURES="sm_86" 
-DBOOTSTRAP_CMAKE_BUILD_TYPE=Release 
-DCLANG_ENABLE_BOOTSTRAP=ON 
-DCLANG_BOOTSTRAP_TARGETS="install-clang;install-clang-resource-headers" 
-DLLVM_ENABLE_PROJECTS="clang;clang-tools;lld;lldb;openmp"

) but then when I try to compile an example: clang++ -fopenmp -fopenmp-targets=nvptx64 -O3 run.cpp:

clang++: error: cannot determine nvptx64 architecture: nvptx-arch: posix_spawn failed: No such file or directory; consider passing it via '--offload-arch'; environment variable CLANG_TOOLCHAIN_PROGRAM_TIMEOUT specifies the tool timeout (integer secs, <=0 is infinite)

And when I pass --offload-arch="sm_86":

clang++: error: no library 'libomptarget-nvptx.bc' found in the default clang lib directory or in LIBRARY_PATH; use '--libomptarget-nvptx-bc-path' to specify nvptx bitcode library

I tried the current master branch and tag 20.1.5. Also multiple machines and GPUs (sm_60, sm_75, sm_86), Debian, NixOS (all x86_64). The Docker container is run with runtime nvidia: docker run -it --runtime=nvidia --gpus all -v $PWD:/app clang-cuda:latest /bin/bash and nvidia-smi shows the correct result:

Mon Jun  2 09:39:09 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03              Driver Version: 560.35.03      CUDA Version: 12.6     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3080 ...    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   42C    P8             14W /  150W |     593MiB /  16384MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
+-----------------------------------------------------------------------------------------+

Florian Becker · Accepted Answer · 2025-06-14 14:03:53Z

0

As per this discussion on the LLVM Forum, the solution is to build with -DLLVM_TARGETS_TO_BUILD="host;NVPTX;AMDGPU" .

The Dockerfile can be found on GitHub

answered Jun 14 at 14:03

Florian Becker

12312 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

LLVM OpenMP GPU Offloading Docker Image

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related