A Parallel Viterbi Decoder Implementation for High ThroughputOutlineMotivationActual Algorithmic FlowOptimizationOptimization Contd.Slide 7Slide 8ApproachResultsResults Contd..ConclusionMuhammad Shoaib Bin AltafOutlineMotivationActual FlowOptimizationsApproachResultsConclusionMotivationConvolutional coding with Viterbi decoding a powerful method for FEC in Communication SystemsViterbi Algorithm is based on Maximum Likelihood Estimation which is sequential. Thus slow.Modern Communications Standards like Wimax support very high throughputData speed is increasing so is the need for high speed Viterbi decodingWe are looking for such a scheme which gives vectorized output bitsActual Algorithmic FlowWe have done this stuff in our Homework as wellOn building trellis, at each stage path metric will be computed Best path metric computation at each stageTraceback decoding done bit by bitEach clock cycle, one bit will be decodedOptimizationVA is sequential but the “Good” thing is, it’s Recursive Various optimization possibilities can be employed for speed-up.Since the purpose was to have vectorized output, the only viable option is ‘Look Ahead Transformation’Discussed Look Ahead transformation for Hoffman decoding in the classBlock processing of the dataOptimization Contd.Decoding using 2 Look Ahead step.Optimization Contd.Increasing the number of Look Ahead stepsOptimization Contd.Instead of 2 paths, we have to select the minimum among the 4 possible pathsLookup table needs to be changedApproachMatlab SimulationN=10^5 bits of dataTwo implementations of VA Constraint Length K=3One based on simple decodingOther based on Look Ahead TransformationPerformance comparison to justify the correctness of the suggested approachResultsData processing speed nearly doubles on taking a single Look Ahead step.Sequential VA Optimized VAExecution time in Seconds 38.3294 20.0305Results Contd..Performance Comaprsion0 1 2 3 4 5 6 7 8 9 1010-510-410-310-210-1Eb/No, dBBit Error RateBER comparision for different Viterbi decoding imlementations for BPSK in AWGN theory - uncodedsimulation - Viterbi sequentialsimulation - Viterbi parallelConclusionLook Ahead Transformation is very attractive for increasing the throughput for Recursive AlgorithmsNo loss in decoding abilitiesDepending on the Application Look Ahead step can be increased to any valueThe extra hardware cost is nominal as compared to the achieved performanceIn this Project the main focus was on speeding up the decoding rate irrespective of the extra hardware cost
View Full Document