CP-CALS

Software for computing the Canonical Polyadic Decomposition (CPD), also known as PARAllel FACtors (PARAFAC), using the Concurrent Alternating Least Squares Algorithm (CALS).

Requirements

Mandatory

• CMake 3.17.5 or higher. A bash script is provided, to help install CMake 3.17.5 (for Linux) in the extern folder and updates the PATH environment variable to point to it. Try using it by navigating to the directory where you cloned CALS and running: source scripts/environment_setup.sh. Then, using the same terminal session, running cmake --version should return version 3.17.5.
• OpenMP
• BLAS/LAPACK (not required for MATLAB MEX generation). Either of the following libraries has been tested to work:
  • Intel MKL
  • OpenBLAS

Optional

• CUDA 11
• MATLAB 2019b

CALS has been tested to compile with g++-8, g++-10 and clang-10.

Compilation

Clone the CP-CALS repo using:

git clone https://github.com/HPAC/CP-CALS.git

Use the following commands to compile the MKL version.

cd CP-CALS/build
cmake  \
-DCMAKE_BUILD_TYPE=Release \
-DWITH_MKL=ON ..

make -j 8 all

or the following commands to compile the OpenBLAS version.

cd CP-CALS/build
cmake  \
-DCMAKE_BUILD_TYPE=Release \
-DWITH_OPENBLAS=ON ..

make -j 8 all

Use the following commands to compile the MKL version with CUDA enabled.

cd CP-CALS/build
cmake  \
-DCMAKE_BUILD_TYPE=Release \
-DWITH_MKL=ON \
-DWITH_CUBLAS=ON ..

make -j 8 all

To compile the MEX files for Matlab use:

cd CP-CALS/build
cmake  \
-DCMAKE_BUILD_TYPE=Release \
-DWITH_MATLAB=ON \
-DMATLAB_PATH=/path/to/Matlab ..

make -j 8 all

Executing example code

src/examples/driver contains a demonstration of how to use CALS (The -h flag is supported for a look at possible input arguments).

MATLAB examples

After compiling the MATLAB MEX, one can execute file matlab/matlab_src/TTB_vs_CALS.m in MATLAB. This executable performs a comparisson of TensorToolbox and CALS. The user first needs to point MATLAB to the CALS MEX and the Tensor Toolbox source code by editing the first two lines of the file.

Related Publications

• Algorithm 1026: Concurrent Alternating Least Squares for multiple simultaneous Canonical Polyadic Decompositions

• Accelerating jackknife resampling for the Canonical Polyadic Decomposition