A Highly Efficient Implementation of I/O Functions on GPU

The API interfaces provided by CUDA can help programmers develop CUDA applications and get high performance in GPU. However, many of the I/O operations are not supported in device codes. This paper has implemented most of the I/O functions through host's agent by using the characteristics of mapped memory in CUDA, such as read/write file and 'printf'. The methods that used to implement these I/O functions will not affect the performance of original applications, users' I/O requirements can be responded quickly, even more, the performance of 'printf' implemented in this paper is higher than that provided by CUDA. This paper supports easy and effective real-time debug method to GPU users, the research in this paper can improve productivity of converting legacy C/C++ codes to CUDA codes, and it is a valuable investigation for broadening CUDA's functions.