RaptorQ code, the next generation of Raptor code for forward error correction (FEC), is proposed to significantly reduce the redundant information. However, the improved coding performance comes at the expense of increased encoding and decoding complexity. On the other hand, graphics processing units (GPUs) are finding their way beyond graphics processing into general purpose computing in the consumer market. This paper investigates the suitability of GPU for RaptorQ code to process large block and symbol sizes in FEC transmission. The paper explores serial and parallel implementations of Raptor code on CPU and GPU, respectively. Our work shows that efficient parallelization on the GPU can improve the performance of the decoder significantly. Furthermore, simulations are performed for the practical real time requirement in multimedia broadcast/multicast service (MBMS) and digital video broadcasting (DVB) in highspeed downlink packet access (HSDPA) network. Conclusions are drawn with respect to the applicability of this new code for realtime multimedia broadcasting and content delivery on GPU.

In this paper, we present a first analysis of the application of Raptor codes in the domain of P2P streaming. With the help of fountain codes, such as Raptor codes, it is possible to completely omit content reconciliation in P2P networks. Hereby, the scheduling complexity of the data dissemination is greatly reduced. The contributions of the paper are the following: First, we present our implementation of the Raptor code used in the performed experiments and elaborate the application of the Raptor code in the scenario of P2P streaming. Second, we investigate the choice of the prevalent parameters, necessary to achieve the best trade-off between performance, computational complexity and resilience of the Raptor code. We use the obtained results to evaluate the general feasibility of using Raptor codes to improve the performance of P2P streaming networks. In addition, we report some insights arising from the practical experience with Raptor codes.

Raptor codes have been proven very suitable for mobile broadcast and multicast multimedia content delivery, and yet their computational complexity has not been investigated in the context of embedded systems. At the heart of Raptor codes are the matrix inversion and vector decoder operations. This paper analyzes the performance, energy profile, and resource implication of two matrix inversion and decoding algorithms; Gaussian elimination (GE) and third Generation Partnership Group (3GPP) standard (SA), for the Raptor decoder on a system on a chip (SoC) platform with a soft-core embedded processor. We investigate the effect of the cache size, memory type, and mapping on the performance of the two algorithms under consideration. We show that with an appropriate data to memory mapping, a speedup factor of 5.77 can be obtained for GE with respect to SA. This paper also proposes a dedicated peripheral hardware block that achieves 5.90 times better performance compared with the software, requiring an energy consumption that is lower by a factor of 5.5, when the symbol size and the data path word length are small (32 bits). We show that with parallel processing in hardware, using the wider word lengths, and employing bigger symbol sizes T, we can improve the performance, while reducing the energy consumption. Extending the hardware word length and symbol size T to 128 bits will result in a performance improvement factor of 6.73 in favor of the hardware; while energy consumption reduces by a factor of 3.8.

In this paper we propose a novel parallel hardware architecture for two binary matrix inversion and vector decoding algorithms, for hard Raptor decoder. We compare the achieved performance to a software based implementation in an embedded processor. We demonstrate the superiority of our proposed architecture in terms of performance (by a factor 12), power and energy dissipation (by a factor of 15). We also include the hardware resource requirements in the comparison. Furthermore, the proposed hardware architecture is parameterized and easily scalable. The data processing word size has been successfully extended up to 1024 bits and fitted within the chosen FPGA hardware platform.