Machine Learning for Vectorization Decision in OpenCL/SYCL Kernel

Abstract

Vectorization of OpenCL/SYCL kernel on CPU device could improve performance significantly. It utilizes single instruction multiple data (SIMD) instruction to process multiple work-items concurrently. However, some applications don't benefit from vectorization. Whether to do vectorization is a challenging problem, since it could vary from case to case. For OpenCL kernels, Intel SYCL CPU device currently uses heuristic to decide whether to discard vectorized kernel. This paper presents a machine learning approach to tackle this problem. Experimental result on Intel Xeon Cascade Lake CPU demonstrates the new approach is better than the heuristic approach.