Compilers

In addition to having access to the provided software modules, you can always compile and use your codes or other open-source codes on Midway. You can check out the source files from GitHub (via git clone) or copy from your local machine to your own space on Midway (e.g. under /home or /project).

Depending on the requirements of the codes, you can load the compilers and libraries that are provided as modules.

GNU GCC, Intel and AMD compilers are provided through modules on Midway2 and Midway3. The table below lists details about each of the module-provided compilers.

Midway2Midway3

Vendor	Module	Language	Compiler
GNU	`gcc`	C C++ Fortran	`gcc` `g++` `gfortran`
Intel	`intel`	C C++ Fortran	`icc` `icpc` `ifort`

Vendor	Module	Language	Compiler
GNU	`gcc`	C C++ Fortran	`gcc` `g++` `gfortran`
Intel	`intel`	C C++ Fortran	`icc` `icpc` `ifort`
AMD	`aocc`	C C++	`clang` `clang++`
NVIDIA	`nvhpc`	C C++ Fortran	`nvc` `nvc++` `nvfortran`

Note

AMD compilers are available on the Midway3 AMD cluster and with module use /software/modulefiles-amd.

Each module may have different versions. The default version is always loaded if you do not specify explicitly with module load. You should check with module avail to see what versions are available. For example, on Midway3 module avail gcc will return

---------------------------- /software/modulefiles -----------------------------
gcc/7.4.0  gcc/10.2.0(default)

and module avail intel

---------------------------- /software/modulefiles -----------------------------
intel/16.0               
intel/18.0.5             
intel/19.1.1(default)
intel/2022.0

You can check the changes to the environment variables made by a particular module by the module show command.

Multithreading with OpenMP

To compile your code or software packages that support multithreading with OpenMP, you just need to add to the compiling flags -fopenmp for GCC and AMD compilers and -qopenmp for Intel compilers.

Message-passing interface (MPI)

To compile your code or software packages with MPI, you need to load the MPI modules that are available. The list of libraries and software suites that provide MPI libraries is given as below.

Midway2Midway3

Library	Module	Language	Wrapper
OpenMPI	`openmpi`	C C++ Fortran	`mpicc` `mpicxx` `mpif90`
MPICH	`mpich`	C C++ Fortran	`mpicc` `mpicxx` `mpif90`
Intel	`intelmpi` `oneaapi`	C C++ Fortran	`mpiicc` `mpiicpc` `mpiifort`

Library	Module	Language	Wrapper
OpenMPI	`openmpi`	C C++ Fortran	`mpicc` `mpicxx` `mpif90`
MPICH	`mpich`	C C++ Fortran	`mpicc` `mpicxx` `mpif90`
Intel	`intelmpi`	C C++ Fortran	`mpiicc` `mpiicpc` `mpiifort`
NVIDIA	`nvhpc`	C C++ Fortran	`nvc` `nvc++` `nvfortran`

Note

For AMD C/C++ compilers clang and clang++ (available with module load aocc), you need to load a MPI module (e.g. openmpi or intelmpi) to compile MPI codes.

Note

Experienced users can build the MPI libraries of their preferences in their own space using the provided compilers above.

GPU codes

To compile GPU codes, you can use NVIDIA CUDA Toolkit and HPC SDK, or Intel OneAPI.

NVIDIA toolsets

CUDA Toolkit

There are several NVIDIA CUDA toolkit versions on Midway2 and Midway3. You can check the version provided with module avail cuda. On Midway3 there are several CUDA versions:

--------------------------- /software/modulefiles -----------------------------
cuda/10.2  cuda/11.2  cuda/11.3  cuda/11.5 cuda/11.7 cuda/12.0

The current version of the GPU driver on the GPU nodes supports all the above CUDA toolkit versions. You can check the GPU driver version via the nvidia-smi command after loading a cuda module.

Note

Although you can compile your CUDA code on the login node with the CUDA toolkit module loaded, running the generated binary on the login node will fail because there is no GPU on the login node.

NVIDIA HPC SDK

NVIDIA HPC SDK provides another toolset for compiling GPU-enabled C/C++/Fortran codes via OpenACC. NVIDIA HPC SDK also provides a set of GPU-optimized tools and math libraries. These compilers are available through the nvhpc module.

module load nvhpc

and check for the location of the compilers

which nvc++
which nvfortran

Intel oneAPI

The DPC++ Compiler compiles C++ and SYCL* source files with code for both CPU and a wide range of compute accelerators such as GPU and FPGA. You can load the oneapi module on Midway2 to get access to this compiler:

module load oneapi

and remember to source the shell script to setup the necessary environment variables:

source /software/oneapi-2021.beta7-el7-x86_64/inteloneapi/setvars.sh

and check for the location of the compilers

which dpcpp

Go

Go is an open-source compiled programming language that gain an rapidly increasing interest and usage by the industry. Although Go is not a traditional compiler, we include it here for convenience. You can load Go as a module on Midway3.

Java

Java is available as modules on Midway2 and Midway3. You can check the available modules via module avail java.