Compiling applications in Mahti
General instructions
- Whenever possible, use the local disk on the login node for compiling software.
- Compiling on the local disk is much faster and shifts load from the shared file system.
- The local disk is cleaned frequently, so please move your files elsewhere after compiling.
Building MPI applications
C/C++ and Fortran applications can be built with GNU, AMD, or Intel compiler suites. The GNU compilers are loaded by default. AMD compilers can be loaded using the Modules system with the command:
module load aocc
Different applications function better with different suites, so the selection needs to be done on a case-by-case basis.
The MPI environment in Mahti is OpenMPI, and when building MPI
applications all compiler suites can be used with
the mpicc (C), mpicxx (C++), or mpif90 (Fortran) wrappers.
The compiler options for different suites are different. The recommended basic optimization flags are listed in the table below. It is recommended to start from the safe level and then move up to intermediate or even aggressive, while making sure the results are correct and the program's performance has improved.
| Optimisation level | GNU | Intel | AMD (clang) |
|---|---|---|---|
| Safe | -O2 -march=native | -O2 -fp-model precise | -O2 -march=native |
| Intermediate | -O3 -march=native | -O2 | -O3 -march=native |
| Aggressive | -O3 -march=native -ffast-math -funroll-loops | -O3 -fp-model fast=2 -no-prec-div -fimf-use-svml=true | -O3 -march=native -ffast-math -funroll-loops |
A detailed list of options for the Intel and GNU compilers can be found on the man
pages (man gcc/gfortran, man icc/ifort) when the corresponding programming
environment is loaded, or in the compiler manuals (see the links above).
List all available versions of the compiler suites:
module spider gcc
module spider aocc
Building OpenMP and hybrid applications
Additional compiler and linker flags are needed when building OpenMP or MPI/OpenMP hybrid applications:
| Compiler suite | OpenMP flag |
|---|---|
| GNU and AMD | -fopenmp |
Building serial applications
For building serial applications, one needs to use compiler suite specific compiler command:
| Compiler suite | C | C++ | Fortran |
|---|---|---|---|
| GNU | gcc | g++ | gfortran |
| AMD | clang | clang++ | flang |
Building GPU applications
The CUDA, OpenACC and OpenMP Offloading (for C++ codes) programming models are supported on Mahti. Specific modules have to be loaded in order to use them.
For example, to load the NVIDIA HPC SDK 22.3 environment:
module load .unsupported
module load nvhpc/22.3
For more detailed information about the available modules, please see module
spider nvhpc.
Compilers:
* The (nvc) is a C11 compiler that supports OpenACC for NVIDIA GPUs while OpenACC and OpenMP for multicore CPUs.
-
The compiler (
nvc++) is a C++17 compiler which supports GPU programming with C++17 parallel algorithms, OpenACC, and OpenMP Offloading on NVIDIA GPUs. It does not support yet C++ CUDA codes. -
The (
nvcc) is the CUDA C and CUDA C++ compiler driver for NVIDIA GPUs. -
The (
nvfortran) is the CUDA Fortran compiler driver for NVIDIA GPUs, it supports OpenACC as also multicore for OpenACC and OpenMP.
CUDA
To generate code for a given target device, tell the CUDA
compiler what compute capability the target device supports. On Mahti, the
GPUs (Ampere V100) support compute capability 8.0. Specify this using
-gencode arch=compute_80,code=sm_80.
For example, compiling a CUDA kernel (example.cu) on Puhti (for C or C++ codes):
nvcc -gencode arch=compute_80,code=sm_80 example.cu
Compile a CUDA Fortran code named example.cuf
nvfortran -gpu=cc80 example.cuf
OpenACC
To enable OpenACC support, one needs to give -acc flag to the compiler.
To generate code for a given target device, tell the compiler what compute capability the target device supports. On Puhti, the GPUs (Ampere A100) support compute capability 8.0.
For example, to compiling C code that uses OpenACC directives (example.c):
nvc -acc example.c .gpu=cc80
For information about what the compiler actually does with the OpenACC
directives, use -Minfo=all.
For Fortran code:
nvfortran -acc example.F90 -gpu=cc80
For C++ code:
nvc++ -acc example.cpp -gpu=cc80
OpenMP Offloading
To enable OpenMP Offloading, the options -mp=gpu is required
For example, compile a C code with OpenMP offloading:
nvc -mp=gpu example.c -gpu=cc80
For Fortran code:
nvfortran -mp=gpu example.F90 -gpu=cc80
For C++ code:
nvc++ -mp=gpu example.cpp -gpu=cc80
The nvc++ compiler supports codes that contain OpenACC, OpenMP Offloading and C++ parallel algorithms in the same code,
for such case you can compile with:
nvc++ -stdpar -acc -mp=gpu example.cpp -gpu=cc80
For MPI, load the module
module load openmpi/4.0.5
The use of the wrappers mpicc, mpic++, mpif90, executes the corresponding nvc,nvc++,nvfortran respectively.