The Research Of Acceleration Methods In Communication Simulation Based On GPU

In recent years, graphic processing unit (GPU) is increasingly applied to general-purpose computing, because of its high parallel computing performance. GPU is also used in communications field. Open computing language (OpenCL) is an open industry standard for general purpose programming across heterogeneous platforms consisting of central processing units (CPUs), GPUs, digital signal processors (DSPs) and other processors. Channel model simulation is one of the most time-consuming modules in communication system. In this paper, we mainly research the acceleration of channel model simulation based on GPU with OpenCL.Channel model mainly includes three parts, which are generating channel coefficients, generating additive white Gaussian noise (AWGN) and filtering input signal. This paper mainly research parallel simulation of Jakes model and spatial channel model (SCM) which are commonly used. OpenCL doesn’t offer libraries to generate random numbers. Therefore, Gaussian random number generation and channel coefficients generation are the most two critical problem to solve in this paper.First, this paper analyzed the OpenCL architecture and optimizations techniques. We tested the performance of transmission bandwidth and computing power of GTX660, which is the GPU in our platform. These experiment results supply basis for the future research on optimization. Secondly, we research Gaussian random number generator. We derived and implemented parallel congruential method and Mersenne Twister (MT) algorithm. And this part also showed the statistical properties and the time performance of random numbers generated by GPU. The results show that the implementation achieves30to150speedup compared with central processing unit (CPU). Third, we studied the principles and analyzed the parallelism of Jakes and SCM model. Then sub-modules including channel coefficients generation and input signal filtering were implemented by OpenCL. A lot of effective GPU accelerating approaches were employed to make all those GPU functions highly optimized. These approaches include static threads, global memory coalesced access, sharing local memory, out-of-order command queue and vectorization. Finally, we verified our approaches on Nvidia’s mid-range GPU GTX660and showed the experiment results.The results showed that GPU based channel model simulation with LTE system configurations can achieve real-time simulation. Even for the more complex SCM model, when sampling rate is30.72MHz and MIMO channel is2x2, the total kernel execution time is about0.93584ms, which is less than1ms during one sub-frame based on mid-range GPU GTX660. With the introduction of3D/Massive MIMO technology, the channel model simulation is more complex and timing-consumed. The results of this paper should have significant application value in practice.