Novel automatic mapping technology on CPU-GPU heterogeneous systems
Zhu Zhengdon · Computer Engineering and Applications Journal · 2015
Aiming at the developing and porting difficulties of GPU-based applications, a mapping approach is proposed,which converts serial computing source code into equivalent parallel computing source code. This approach acquires hierarchies of parallelizable loops from serial sources, establishes the correspondence between loop structures and GPU threads, and generates the core function code for GPU. Meanwhile, CPU control code is generated according to read/write attributes of variable references. A compiler prototype is implemented based on this approach, which translates C code into CUDA code automatically. Functionality and performance evaluations of the prototype show that the CUDA code generated is functionally equivalent to the original C code, with significant improvement in performance, thus overcomes the difficulty in porting compute-intensive applications to CPU-GPU heterogeneous systems.