Implementation and trade-offs of a DCT architecture using high-level synthesis

This paper presents architectural trade-offs of a time-shared implementation of a modified fast discrete cosine transform algorithm using a high-level synthesis tool. The architecture presented here allows time-sharing of operators in different stages. The overhead in control and multiplexing is minimal. A full implementation of an 8×8 2-D DCT outperforms the original pipelined architecture and a hand-crafted time-shared architecture by reducing the required area by up to 50%. It also improves the latency by up to 70%. It achieves these improvements maintaining the throughput for a 5% decrease in the required critical path timing. The complexity of the 2-D DCT used is higher than the traditional benchmarks for high-level synthesis. This paper shows the effectiveness of the synthesis tool used for large, practical algorithms