A Reconfigurable FPGA Architecture for DSP TransformsOUTLINEMotivationNeed for Reconfigurable ArchitectureAREA & POWERSlide 6Discrete Gabor TransformDiscrete Convolution2-D Fourier TransformConvolution OperationConvolution Operation (Contd.)Imaginary Radix RepresentationConvolution Operation (using Complex Representation)Convolution using Complex Representation - Communication GraphGabor Transform Communication GraphReconfigurationWork so farCONCLUSIONA Reconfigurable FPGA Architecture for DSP TransformsSubramanian Rama Vishnu VijayaraghavanOUTLINEMotivationReconfigurable FPGA’sDSP Transforms, Breakdown & ApplicationsCommunication Graphs& Proposed ArchitectureImaginary Radix Complex MultiplicationAccomplished WorkConclusionMotivation Dedicated VLSI Architectures for Orthogonal Transforms – FFT, DCT, Convolution, CorrelationDedicated VLSI Architectures for Non- Orthogonal Transforms – Gabor, WaveletNot many Architectures for Both – Current Day Applications like Handhelds, Mobile Phones, etc. require such DSP capabilitiesNeed for Reconfigurable ArchitectureMultiple Orthogonal & Non-Orthogonal Transforms can be broken down to a basic set of Building blocks (DCT,DST, multipliers and Adders)Handheld devices don’t require much Multiprocessing – No need to waste hardwareIncreased Fault-Tolerance By Reconfiguration and RedundancyAREA & POWERINCREASING PROMINENCE OF PORTABLE SYSTEMSCell PhonesPersonal Digital AssistantsTablet PC’sNeed for Low Power & AreaBattery Technology not kept pace with Semiconductor TechnologyBATTERY(40+ lbs)DISCRETE FOURIER TRANSFORMAPPLICATIONS:Image Processing Orthogonal Frequency Division MultiplexingTraditional DFTBreakdown of 2D DFTBreakdown of 1D DFTDiscrete Gabor TransformGabor Transform and CoefficientsBreakdownApplicationsSpeech Processing / Voice RecognitionImage CompressionDiscrete ConvolutionApplicationsImage ManipulationSound Processing2-D Fourier TransformConvolution OperationConvolution Operation (Contd.)Computational complexity:2 DCT, 2 DST,4 real multiplications and 2 real additionsImaginary Radix RepresentationA imaginary number system, Donald Knuth, Communications of the ACM Concept:a + ib = A – Interleave both real and Imaginary parts# of multiplications get reduced to onePreserve Interleaving even during multiplicationRequires slight modifications in multiplier design (one reason for migrating to FPGA)Convolution Operation (using Complex Representation)Computational complexity:2 DCT, 2 DST,1 complex multiplication (same as real multiplication methodology)Convolution using Complex Representation - Communication GraphGabor Transform Communication GraphReconfigurationWork so far Design & Synthesis of Basic Building BlocksDCTDSTParallel Array MultiplierReconfiguration UnitPartial IntegrationWork to be done:Complete IntegrationFunctional Correctness CheckCONCLUSIONNeed for multiple transforms on same chipMobile devices, HandheldsNot much multiprocessing requiredUse of Reconfigurable FPGA’s ReducesAREAIncreasesFunctionalityFault
View Full Document