Installation

BMLP-GPU is usable both from Python and SWI-Prolog.

Conda installation

Download conda (Anaconda) for your device. Then run in terminal, e.g.

bash Anaconda-latest-Linux-x86_64.sh
conda list

To confirm installation success, you should see the list of installed conda packages.

Dependency

BMLP-GPU requires the following packages in a conda environment:

• python
  • pytorch
  • python-graphblas
  • cudatoolkit
  • ipykernel (optional for running Jupyter notebook)

To create a conda environment with above packages:

conda env create -f environment.yml
conda activate bmlp

Examples

Two BMLP modules RMS and SMP [1] are in bmlp.py.

See the list of linear algebra operations supported by PyGraphBLAS in bmlp.py.

Calling from Python

Please see examples for calling RMS and SMP modules in examples.ipynb.

Calling from SWI-Prolog

As an example, the following clauses in the test/int_matrix.pl file describe rows in a boolean matrix, we show how to use BMLP-GPU to compute its transitive closure.

            0  1  2  3  4
a(2).   0|     1         |  0
a(4).   1|        1      |  1
a(8).   2|           1   |  2
a(1).   3|  1            |  3
a(16).  4|              1|  4
            0  1  2  3  4

On a terminal session, create a Python server for SWI-Prolog clients.

python -m bmlp.service.swipl_server

On another terminal, create a SWI-Prolog client (alternatively, you can consult swipl_client.pl in your source code).

swipl bmlp/service/swipl_client.pl

On the SWI-Prolog client side, request the conversion from integers in test/int_matrix.pl to a graphBLAS sparse matrix. The converted matrix is stored by the python variable 'a'.

?- run_python_command("import bmlp.core.utils", Res).
?- run_python_command("a = bmlp.core.utils.integers_to_boolean_matrix('test/int_matrix.pl')", Res).

Now, we call RMS to compute the transitive closure taking a single matrix argument. Here we input 'a'. The computed result is assigned to 'a' again.

?- run_python_command("import bmlp.core.matrix", Res).
?- run_python_command("a = bmlp.core.matrix.RMS(a)", Res).

The final step is to convert the graphBLAS sparse matrix back to SWI-Prolog. The first argument is a graphBLAS matrix, the second argument is the name of the new predicate, the third argument is the file location where the set of clauses will be saved into.

?- run_python_command("a = bmlp.core.utils.boolean_matrix_to_integers(a,'a','test/output.pl')", Res).

The output.pl file now contains clauses that represent the transitive closure boolean matrix.

            0  1  2  3  4
a(15).  0|  1  1  1  1   |  0
a(15).  1|  1  1  1  1   |  1
a(15).  2|  1  1  1  1   |  2
a(15).  3|  1  1  1  1   |  3
a(16).  4|              1|  4
            0  1  2  3  4

References

[1] Ai, Lun, and Stephen H. Muggleton. ‘Boolean Matrix Logic Programming’. arXiv, 19 August 2024. https://doi.org/10.48550/arXiv.2408.10369.