About / FAQ

About / FAQ#

ASTRA KernelKit is a side-project of ASTRA Toolbox, and aims to provide more flexible access to the tomographic projectors and kernels used in ASTRA. It is a minimalistic standalone library.

How does KernelKit work?#

Tomographic projectors are often written in a low-level language, like C/C++, and use NVIDIA CUDA for acceleration on graphical processing units (GPUs).

ASTRA KernelKit makes it easy to use Python to access and modify these projectors. It does so by using CuPy, which in turn uses NVRTC, the NVIDIA runtime compiler, to compile CUDA kernels during the Python script.

A comparison of KernelKit with ASTRA Toolbox and Tomosipo.

The comparison of KernelKit with ASTRA Toolbox and Tomosipo is explained in the paper publication associated with the software package [1].#

ASTRA KernelKit or ASTRA Toolbox?#

KernelKit:

  • directly compile and call ASTRA Toolbox kernels in your Python code;

  • have more flexibility toward data axis order, data types, and memory layout, helping to get better performance;

  • customize ASTRA Toolbox kernels, for, e.g., machine learning applications.

  • integrate kernels with advanced CUDA, such as unified memory and streams.

Toolbox:

  • pre-implemented algorithms (e.g., SIRT, CGLS, FBP, etc.);

  • larger variety of kernels and functionalities;

  • multi-GPU support out of the box;

  • Windows and MatLab support.

Why CuPy?#

A frequently asked question is why KernelKit is based on CuPy, and not any other GPU array library such as PyTorch or PyCUDA.

A main feature of CuPy is that it transparently exposes advanced CUDA functionality to Python, such as the NVIDIA run-time compiler, but also streams, graphs, CUDA arrays, and texture memory. This allows KernelKit to be written with a minimalistic code base and without shipping binary code.

Is KernelKit compatible with …?#

ASTRA Toolbox: KernelKit does not aim to be a drop-in replacement for ASTRA Toolbox. There are ForwardProjector and BackProjector ASTRA Toolbox compatible classes available, making it easy to cross-test implementations with the ASTRA Toolbox.

Tomosipo: Work in progress in a Github branch. Here a KernelKit operator can be created with ts.operator(..., backend='kernelkit').

PyTorch: Torch tensors can be converted to CuPy arrays and vice versa efficiently, without incurring copies. See kernelkit.torch_support.AutogradOperator.

I found an issue / How can I contribute?#

Contributions welcome! Please report any issues on the Github issue tracker.

Contents