Parallel turbo decoder using a low-latency Max-Log-MAP kernel for a VLIW DSP

Turbo-Codes have attracted great interest in digital mobile radio due to their remarkable error correcting capabilities. In this paper, we present an optimized Turbo decoder for wireless systems following recommended CDMA2000 standard (192 time frames with 8 states per frame). This decoder is implemented using TI's latest C64x digital signal processor. To achieve the maximum parallelism in TI's VLIW architecture, we have specially redesigned the trellis computation algorithm to improve the decoder throughput and reduce the number of computation operations required. In particular, our proposed algorithm transforms a number of add/subtract operations to multiplication operations. In this way, previously unused functional units become available and therefore, more parallel instructions can run simultaneously, leading to throughput increase and latency reduction. To our findings, current TI's compiler (Code Composer V1.2) fails to generate the optimized assembly code when our algorithms are directly implemented in C. To this end, we have illustrated optimized resource binding and timing schedules by applying code motion and loop transformation techniques. The optimized Turbo decoder can finish one decoding stage in 18.1 microseconds for a C64x DSP running at 400 MHz.