The Synchronization Treatment in Implementing Data-Parallel Programming Languages on CPUs
Feng Yue, Jianmin Pang, Rongcai Zhao, Chao Dai · 2013
When implementing data-parallel programming languages such as CUDA, OpenCL on CPUs, synchronization must be simulated correctly. The basic method is thread-based, which means all thread must execute one instruction in turn before execute the next one. In this paper, we propose function splitting to treat synchronization in a co routine style but not just thread-based. It splits the data-parallel function presented by low-level intermediate representation into several parts by simulating synchronization. We evaluate our method in translating PTX kernels to multi-core CPUs, the result of which shows this method could promotes performance by 15% compared to thread-based method. Our main contribution is a generous synchronization treatment that performs on low-level intermediate code given by a control flow graph in SSA form.