link to homepage

Navigation and service

Master thesis: Efficient algorithms for multiplication of narrow rectangular matrices

Advertising institute: ER-C-1 - Physics of Nanoscale Systems
Reference number: D037/2018, Physics, informatics/computer science

Fast pixelated direct electron detectors for scanning transmission electron microscopy (STEM) allow capturing the entire transmitted beam as a high-resolution detector frame for each scanned pixel. Subsequent processing involves element-wise multiplication and integration with a set of masks or weighing matrices for each frame in order to extract relevant information for each pixel.

This can be expressed as a rectangular matrix multiplication of the scan result with a stack of masks. In principle, this is a task for one of the many BLAS libraries. However, the common ones do not perform well for very narrow matrices, i.e. with only a few rows or columns, like the mask set in this application.

The subject of your master thesis is to identify the reason for this weak performance and subsequently design, implement and test a superior solution as an extension of a common open source BLAS library. Ideally, you will work with the upstream project to integrate your code so that it finds widespread application.

The Forschungszentrum Jülich is home to some of the most advanced electron microscopes in the world and houses several supercomputers that currently take ranks 22 and 29 on the Top500 list. Experts both in electron microscopy and computer science will guide you with your project.

Contact person:
Dr. Dieter Weber
Forschungszentrum Jülich
Ernst Ruska-Centre for Electron microscopy