Parallel Algorithms And System Architecture For Software Radio On GPUs

With the development of information technology, the Wide-band Wireless Communication(WWC) makes people change living mode. People can enjoy the convenience that the WWC bring to us everywhere. However, the mobile terminals in hand should support various wireless protocols, such as the personal area network, local area network,metro area network and wide area network. Software-defined radio is designed to handle this situation that the wireless communication system can support multiple communication modes and protocols without altering the hardware architecture. In the wireless communication technology, multiple input multiple output(MIMO) systems enable the data throughput increasing linearly with the minimum numbers of antennas at transmitter and receiver, so that it is one of the most attractive techniques in wireless communication. Orthogonal frequency-division multiplexing(OFDM) is also a key technique for WWC because of its high spectral efficiency and capability to combat multipath fading. At present,many advanced WWC protocols integrate both MIMO and OFDM techniques, forming the MIMO-OFDM WWC, which has the characteristic of high-throughput, high-fidelity,low-latency and multi-user. As a result, the WWC system not only has high-throughput and high processing ability, but also excellent bit error ratio(BER) performance. In the PHY layer of MIMO-OFDM WWC, there are three types of algorithms, i.e. the channel encoding/decoding algorithms, the OFDM algorithms and the MIMO detecting algorithms.In this thesis, the fine-grained parallel algorithm for the software radio algorithms are presented and a real-time software radio system with graphics processing unit(GPU)integrated is proposed. And the detailed works are bellows.(1) A software radio system, named as Cu Sora, is presented. In Cu Sora, the Central Processing Unit(CPU) is used as controller and the GPU is used as baseband processor.Sora platform is a soft radio platform developed by the Microsoft Asian Research Center.Cu Sora combines the Sora platform with the GPU processors. In Cu Sora, the radio-front(RF) signal is received from the air, the frequency down-conversion and analog-to-digital conversion(ADC) are performed through the Sora front-end. Cu Sora reuses the software framework and MAC layer of Sora to control the radio platform. At the meantime, Cu Sora exploits GPU as the modem processor to achieve high-speed PHY signal processing,which make the throughput and BER performance meet the requirement of wireless communication protocols. A multi-mode software controller is also designed in Cu Sora to support multiple protocols.(2) A fine-grained parallel channel encoding and decoding algorithm for forward error correction(FEC) codes and architecture for the encoder and decoder are presented.Three popular FEC codes, i.e. convolutional codes, Turbo codes, and LDPC codes, are analyzed to find the computation characteristic. Appropriate revised encoding or decoding algorithms for the GPU platform are chosen and their parallel algorithms are presented. In the proposed parallel encoding or decoding algorithms, excellent parallelism is exploited and efficient guarding schemes are used to reduce the effect to BER performance. The encoder and decoder for the three FEC codes are implemented on the GPUs with the Fermi architecture. The throughput of decoders is about two orders of magnitude faster than that of CPU implementations and superior to current GPU-based FEC decoders.(3) A fine-grained parallel algorithm and GPU accelerator architecture for the OFDM algorithms are presented. Three types of OFDM algorithms, i.e. OFDM modulation/ demodulation, synchronization, and channel estimation, are analyzed to find the computation characteristic. Then the fine-grained parallel algorithms for such modules are presented, where not only large parallelism is exploited, but also the multiple sub-carriers can carry correct samples and communication with other sub-carriers. The GPU accelerators for the three OFDM modules are implemented on the GPUs with the Fermi architecture. Several optimization methods are exploited and the throughput is about two orders of magnitude faster than that of CPU implementations.(4) A fine-grained parallel algorithm and GPU accelerator architecture for the MIMO detectors are presented. In this paper, PIC MIMO detector is chosen for its low complexity and high BER performance. The fine-grained parallel algorithms for such detectors are presented, where not only large parallelism is exploited, but also the original transmitted sample vectors can be recovered correctly. The GPU accelerators for the three MIMO detectors are implemented on the GPUs with the Fermi architecture. Several optimization methods are exploited and the throughput is about two orders of magnitude faster than that of CPU implementations.(5) A GPU-based software radio prototype system is presented and implemented.The transceiver modules in the physical layer of two wireless protocols, i.e. Wi Fi(802.11a)and Wi MAX(802.16d), are analyzed. GPU-based OFDM parameterized software defined radio(SDR) framework is presented. The encoders, decoders, OFDM modules, and MIMO detectors presented in this thesis are combined to form the physical layer of the two protocols. These modules are accelerated on the Cu Sora, and the real wireless communication is realized between two Cu Sora platforms. Compared to the Sora platform, the data rate for the wireless communication on Cu Sora can achieve about 10% to 30% performance improvement. The throughput of each module on Cu Sora is superior to that of Sora platform and CPU, DSP, FPGA in other works.