Kernel fusion is one of the common techniques used to optimize the performance of CUDA programs. In this technique, two kernels are combined into a single one. There are three main fusion approaches and depending on the nature of the problem and the algorithm employed in each program, one approach may yield better performance than the others. Programmers typically determine the appropriate fusion method through trial and error. Developing a performance model capable of predicting the execution time of fused kernels can eliminate this costly and time-consuming process. In this paper, we propose a performance prediction model that estimates the execution time of fused kernels created using each of the three fusion techniques, based on features extracted from the original kernels. This enables programmers to make more informed decisions when selecting the optimal fusion method. The proposed model is developed using machine learning techniques. To construct the dataset, we used eight programs from the NVIDIA CUDA Sample suite and the Rodinia benchmark. Kernels were fused pairwise using the three fusion techniques, and features were extracted from each pair to form the dataset. Nine machine learning algorithms were implemented and evaluated using k-fold cross-validation. Among them, the Random Forest algorithm achieved the best performance and was used to build the final prediction model. Experimental results demonstrate that the proposed model predicts the execution time of fused kernels with an average error of less than 5%. By leveraging machine learning, the proposed approach significantly simplifies the kernel fusion method selection process for programmers.

کلمات کلیدی