GPU_ACCELERATION.md

GPU Acceleration

GPU acceleration of diagrams is still experimental. The following classes are exposed in yarrow.cupy:

• yarrow.cupy.CupyDiagram
• yarrow.cupy.CupyFiniteFunction
• yarrow.cupy.CupyBipartiteMultigraph

Here's a usage example:

import yarrow.cupy as yarrow
# Create a GPU-backed identity diagram with 10,000 objects, all labeled '0'.
# The type of operations Σ₁ is the singleton set `{●}`.
f = CupyDiagram.identity(CupyFiniteFunction.terminal(10000), CupyFiniteFunction.initial(1))

CuPy dependencies

To use the GPU backend, you will need some additional dependencies:

• cupy for "GPU-accelerated numpy"
• pylibcugraph for GPU-accelerated connected components.

At time of writing, you will need to use CUDA 11, since CUDA 12 packages for cugraph are not yet available. See this issue

The following incantation seems to work for installing dependencies:

# create a conda environment called "cuda11" with python 3.10
conda create -n cuda11 python=3.10

# enter the environment
conda activate cuda11

# Install cugraph into the environment (which comes with pylibcugraph)
conda install -c rapidsai -c numba -c conda-forge -c nvidia cugraph cudatoolkit=11.8 pylibraft rmm