GPU accelerated image processing#
As we work often with three-dimensional image data, potentially over time, classical image processing takes quite some time.
Hence, we will also dive into image processing on graphics processing units (GPUs) using OpenCL, pyopencl and pyclesperanto. This technology allows us to process image faster, GPU accelerated. Classical algorithms and GPU-accelerated image processing may differ in the very details but we users should not recognize that. A specific image processing operation should deliver similar results independent from how it is computed.
Installation of requirements#
User of Windows and Mac should not need to install OpenCL. Everything you need should be pre-installed. Linux users need to install an OpenCL-ICD-Loader.
Hence, linux users may have to run commands like this, depending on the linux distribution:
sudo apt update
sudo apt install ocl-icd-opencl-dev
Afterwards, installation can proceed using conda and pip:
mamba install -c conda-forge l pyclesperanto-prototype
Afterwards, you can test it for example by executing these commands in a python script or jupyter notebook:
import pyclesperanto_prototype as cle
print("Used GPU:", cle.get_device())
Also feel free to install the napari-pyclesperanto-assistant plugin in napari.
Installation of optional requirements#
In this chapter, we will also take a look at cupy, an NVidia CUDA based GPU-accelerated processing library and napari-cupy-image-processing, a scriptable napari plugin. These two can be installed using the following commands. This will however only work on computers that have a CUDA-compatible NVidia graphics card.
mamba install -c conda-forge cupy cudatoolkit=10.2
mamba install -c conda-forge napari
pip install napari-cupy-image-processing
Note: Depending on your CUDA installation, you may want to replace the “10.2” in the command line above with the CUDA version you installed on your computer.
See also