Docker images for developing C++ and Fortran HPC programs

Following image tags are pushed from private GitLab CI to public GitHub Container Registry (ghcr.io):

• {year}.{month}.{patch} formatted tags, e.g. 20.10.0
  • Be sure that it is not a semantic versioning. Every release can be a breaking change. You should use containers with a fixed tag.
  • See CHANGELOG for detail about changes
• latest
  • Corresponds to latest branch. DO NOT USE unless you are watching all changes in the latest branch.

Named in allgebra/{GPU}/{Compiler}/{Math} format:

In addition, there are support tools in each containers.

• clang-format
• clang-tidy
• doxygen
• poetry

The OSS compilers in allgebra containers (gcc, gfortran and clang) are compiled with OpenMP and OpenACC supports. There are several examples in this repository, and they are also copied into the above containers.

The requirements of these examples are following:

• Use Linux
  • If you'd like to use WSL2, see Use allgebra on WSL
• Install Docker
• Install NVIDIA Container Toolkit

You can build and run e.g. the clang with OpenMP Offloading example as following:

$ docker run --rm -it --gpus=all ghcr.io/ricosjp/allgebra/cuda10_1/clang12/oss:21.06.0
root@41b65ab23aaf:/# cd /examples/clang_omp_offloading
root@41b65ab23aaf:/examples/clang_omp_offloading# make test
clang++ -fopenmp -fopenmp-targets=nvptx64 -Xopenmp-target -march=sm_70 -O3 -std=c++11 -lm omp_offloading.cpp -o omp_offloading.out
clang++ -fopenmp -fopenmp-targets=nvptx64 -Xopenmp-target -march=sm_70 -O3 -std=c++11 -lm omp_offloading_cublas.cpp -o omp_offloading_cublas.out -lcuda -lcublas -lcudart
clang++ -fopenmp -fopenmp-targets=nvptx64 -Xopenmp-target -march=sm_70 -O3 -std=c++11 -lm omp_offloading_math.cpp -o omp_offloading_math.out
./omp_offloading.out 1000000
dot = 2e+06
Pass!
./omp_offloading_cublas.out 1000000
dot = 2e+06
Pass!
./omp_offloading_math.out 1000000
ret = 909297
Pass!

allgebra_get_device_cc command is contained in the allgebra containers, and it detects the [compute capability](compute capability) of your GPU using CUDA API. On a system with NVIDIA TITAN V (compute capability 7.0), for example, it returns 70:

root@3f6b34672c01:/# allgebra_get_device_cc
70

This output is used to generate the flag -Xopenmp-target -march=sm_70 in above example.

Singularity is a container runtime focused on HPC and AI. Since singularity supports Docker and OCI container images, allgebra containers can be used as it is.

singularity run --nv docker:https://ghcr.io/ricosjp/allgebra/cuda10_1/clang12/mkl:latest

--nv is an option for using NVIDIA GPU in the container. See the official document for detail.

You can build a SIF (Singularity Image Format) file from an allgebra container:

singularity build allgebra_clang_mkl.sif docker:https://ghcr.io/ricosjp/allgebra/cuda10_1/clang12/mkl:latest

and run it:

singularity run --nv allgebra_clang_mkl.sif

--nv option is required for singularity run and not for singularity build since singularity build only download the container and converts it.

Be sure that this allgebra_clang_mkl.sif contains CUDA and MKL binaries. You have to accept the End User License Agreement of CUDA, and follow the Intel Simplified Software License.

In order to identify the CUDA and LLVM versions in container, following environment variables are defined:

See DEVELOPMENT.md

Copyright 2020 RICOS Co. Ltd.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

https://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Be sure that you need to accept end user license agreements of CUDA, and Intel Simplified Software License to use these containers. You can find patched source code of GPL applications derived from nvidia/cuda container at nvidia/OpenSource.