Prediction of SGEMM GPU Kernel Performance using Supervised and Unsupervised Machine Learning Techniques

This paper proposes a novel approach for the prediction of computation time of kernel's performance for a specific system which consists of a CPU along with a GPU (Graphical processing unit). The prediction is done on the basis of 14 different configurations of processor such as local work group size, local memory shape, kernel loop unrolling factor, vector widths for loading and storing etc. A proposed system accomplished with multiple advanced techniques like PCA (Principal Component Analysis), backward elimination for feature reduction, and KNN Regression, and Random Forest Regression for making the predictions. Finally, the performance of the proposed system for predicting the running time of processors has the accuracy of 98.46% after using feature reduction techniques.