Software libraries (generic)

The MeluXina User Software environment offers many software libraries that can be used in your C/C++, Fortran, Python and other applications. This page showcases how you can use them when building your application.

First you will need to activate the corresponding environment module with module load <library/version>.

The modules populate your shell's environment with several variables, one of which is called EBROOT<LIBRARYNAME>.

You may use env | grep EBROOT to see all the variables that appear once a module is loaded, note that any dependencies of the module will also create this type of entries.

The EBROOT* variable contains the library's installation path, and can be used in your commands and scripts ($EBROOT<NAME>) to point to specific folders available underneath and link your application to the library (e.g. $EBROOT<NAME>/lib, $EBROOT<NAME>/include, etc.).

It is preferred to use the variable as the reference to the installation path, instead of hard-coding the full path, which may change.

We provide two different examples for data management and GPU-accelerated libraries:

HDF5 for file format and data model management
cuDNN is the NVIDIA Deep Neural Network for forward and backward convolution, pooling and normalization.

Example using HDF5

EasyBuild module description

HDF5 is a data model, library, and file format for storing and managing data. It supports an unlimited variety of data-types, and is designed for flexible and efficient I/O and for high volume and complex data.

Available HDF5 versions

Check the available versions on MeluXina with module command:

module avail HDF5

Terminal output example

------------------------- /apps/USE/easybuild/release/latest/modules/all --------------------------
HDF5/1.10.7-gompi-2020b     HDF5/1.12.0-gompi-2020b
HDF5/1.10.7-gompic-2020b    HDF5/1.12.1-gompi-2020b (L,D)

Where:
L:  Module is loaded
D:  Default Module

For example, in the MeluXina 2021.2 software stack you will find:

HDF5 - serial	MPI(OpenMPI) - parallel
	HDF5/1.10.7-gompi-2020b
	HDF5/1.10.7-gompic-2020b (CUDA)
	HDF5/1.12.0-gompi-2020b
	HDF5/1.12.1-gompi-2020b

Compilation of code using HDF5 library

Load a serial or MPI-parallel build of HDF5 in order to compile your program:

module load HDF5/1.12.1-gompi-2020b

The EBROOTHDF5 and HDF5_DIR environment variables are then populated in your shell's environment and point to the HDF5 installation dir:

echo $EBROOTHDF5

Terminal output example

/apps/USE/easybuild/release/2021.2/software/HDF5/1.12.1-gompi-2020b/

You can now compile your C/C++/Fortran code (e.g. writedata.cpp) step by step, using the MPI C++ compiler wrapper:

mpicxx -c -I/$EBROOTHDF5/include writedata.cpp

and then link HDF5 library:

mpicxx -o writedata writedata.o -L$EBROOTHDF5/lib -lhdf5_cpp -lhdf5

Note that HDF5 contains helper utilities that can be used to compile in a single step:

h5c++ -o writedata writedata.cpp

Example using cuDNN

EasyBuild module description

 The NVIDIA CUDA Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks.

Available HDF5 versions

Check the available versions on MeluXina with module command:

module avail cuDNN

Terminal output example

-------------------------- /apps/USE/easybuild/release/latest/modules/all --------------------------
   cuDNN/8.0.4.30-CUDA-11.1.1    cuDNN/8.2.1.32-CUDA-11.3.1
   cuDNN/8.1.1.33-CUDA-11.2.2    cuDNN/8.2.2.26-CUDA-11.4.1 (L,D)

   Where:
    L:  Module is loaded
    D:  Default Module

For example, in the MeluXina 2021.2 software stack you will find:

cuDNN/8.0.4.30-CUDA-11.1.1 
cuDNN/8.1.1.33-CUDA-11.2.2    
cuDNN/8.2.1.32-CUDA-11.3.1
cuDNN/8.2.2.26-CUDA-11.4.1

Compilation of code using cuDNN library

Load one of the available version in order to compile your program:

module load cuDNN/8.2.2.26-CUDA-11.4.1

The EBROOTCUDNN environment variables are then populated in your shell's environment and point to the cuDNN installation dir:

echo $EBROOTCUDNN

Terminal output example

/apps/USE/easybuild/release/2021.2/software/cuDNN/8.2.2.26-CUDA-11.4.1

You can now compile your cuda code (e.g. pooling.cu) step by step, using CUDA compiler:

nvcc -c -I/$EBROOTCUDNN/include pooling.cpp

and then link cuDNN library:

nvcc -o pooling pooling.o -L$EBROOTCUDNN/lib -lcudnn -lcudnn_static