Exploration and Implementation of Winograd Convolution using the Streaming Architecture on FPGA

Type of work:

Master Thesis

Assignment:

During this thesis, the student will focus on acceleration of the Winograd convolution algorithm using the streaming architecture approach on FPGA. The responsibilities of the student will be: state-of-the-art exploration with respect to streaming architectures on FPGA and Winograd accelerators. Implementation of understanding of HLS4ML and FINN CONV1D and CONV2D layers for comparison with Winograd implementation. Implementation of a parametrizable/flexible Winograd convolution architecture for different kernel sizes and precisions using the streaming architecture approach and HLS library. Comparison of Winograd implementation with state-of-the-art by parameters: resource utilization, throughput, latency, power consumption, energy, energy efficiency, and area efficiency.

Skills:

• Neural Network Framework (Pytorch, Brevitas)
• Hardware/Software Development (Python, C/C++, HLS, Vitis)
• NN FPGA Frameworks (HLS4ML, FINN-HLS)

Supervisor:

M. M. Ghaffar, M.Moursi, L. Krupp

Student:

Muhammad Adnan

Year:

2024