Automatic Extraction of Compute Patterns from Computer Architecture Descriptions
Bachelor’s Thesis / Master’s Thesis / Student Research Project
Abstract
Finding Neural Network operators in a Neural Network and generating code running on a hardware platform can be done with different deep learning tools like TVM, PyTorch or Tensorflow. These compute patterns are mostly defined manually for each computer architecture.
The goal of this student project is to automatically find and extract those compute patterns (e.g. (parts of) matrix-multiplication/convolution) from an abstract hardware description on the scalar operator level (addition, multiplication, etc.).
Requirements
- Python
- Successfully atteded the lecture “Grundlagen der Rechnerarchitektur” and/or “Parallele Rechnerarchitekturen” (optional)
- Linux (optional)