Enhancing Yolo-v4 Performance using Scalar Matrix Multiplication in oneAPI

Introduction

This project explores the enhancement of Yolo-v4 performance through Scalar Matrix Multiplication (MM) in oneAPI. We focus on optimizing convolution layers in the Yolo-v4 model, integrating oneAPI with Python.

Background

We utilize oneAPI to optimize deep learning computations in the Yolo-v4 model, aiming for improved efficiency and accuracy in object detection.

Methodology

Our approach covers:

1. Scalar Matrix Multiplication in oneAPI
2. Python-C++ Integration
3. Convolution Layer Wrapper
4. Yolo-v4 Modification and Implementation
5. Performance Analysis

Commands to Run

• Scalar MM: icpx -fsycl smm.cpp -o smm
• Python-C++ Integration: icpx -fsycl -fPIC -shared -o libsmm.so shared.cpp
• Convolution Wrapper: python3 wrapper.py
• Yolo-v4: python3 yoto.py, python3 yolo.py
• Performance Analysis: python3 compare.py

Results

Demonstrates minimal speedup in convolution layers of the Yolo-v4 model on CPU devices.

Challenges

Discusses challenges in FLOPs calculation for complex models with custom implementations.

Conclusion

Highlights the potential of Scalar MM in oneAPI for deep learning optimization.

Authors

• Vikash Singh (vxs465)
• Thomas Bornhorst (thb34)

Institution

Case Western Reserve University