The Compute Unified Device Architecture (CUDA) programming environment from NVIDIA is a milestone towards making programming many-core GPUs more flexible to programmers. However, there are still many challenges for programmers when using CUDA. One is how to deal with GPU device memory, and data transfer between host memory and GPU device memory explicitly. In this study, source-to-source compiling and runtime library technologies are used to implement an experimental programming system based on CUDA, called memCUDA, which can automatically map GPU device memory to host memory. With some pragma directive language, programmer can directly use host memory in CUDA kernel functions, during which the tedious and error-prone data transfer and device memory management are shielded from programmer. The performance is also improved with some near-optimal technologies. Experiment results show that memCUDA programs can get similar effect with well-optimized CUDA programs with more compact source code.