GPU Based Polyphase Filter Banks for VLBI - MIT Haystack ...

Introduction Implemenation Results Future Work ConclusionsGPU Based Polyphase Filter Banksfor VLBIMark McCurryChristopher Beaudoin & Geoff CrewMIT Haystack REU Program 2011GPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsOutlineIntroductionWhat is being done?PFBHardwareImplemenationCUDAResultsQualityQuantityFuture WorkConclusionsGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsWhat is being done?DSP◮ Input of Analog values from data source near 4 Gb/sGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsWhat is being done?DSP◮ Input of Analog values from data source near 4 Gb/s◮ Perform Polyphase Filter Bank (PFB) on a Nvidia GPUGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsWhat is being done?DSP◮ Input of Analog values from data source near 4 Gb/s◮ Perform Polyphase Filter Bank (PFB) on a Nvidia GPU◮ Output of channelized frequency spectrumGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsHardwareRoach Digital Backend (RDBE)GPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsHardwareNvidia GPUsGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsCUDAWhat is CUDA?GPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsCUDACUDAg l o b a l voidc u u n q u a n t i z e ( f l o a t ∗ dest , const i n t 8 t ∗ s r c ,s i z e t N){const s i z e t i = blockDim . x∗ b l o c k I d x . x +t h r e a d I d x . x ;i f ( i

Introduction Implemenation Results Future Work ConclusionsCUDACUDA 2// Setup v i a cudaMemcpy// i n p u t and output a r e on the GPUc u u n q u a n t i z e (output , input , 1 0 0 ) ;// Return v i a cudaMemcpyGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsQualityQuality of output: Actual1Magnitude response of 256 tap FIR, 32 tap FFT based PFB0.90.80.7Normalized Gain0.60.50.40.30.20.100 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1Frequency (half−cycles/second)GPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsQualityQuality of output: Matlab1Matlab Prototype PFB0.90.80.7Normalized Gain0.60.50.40.30.20.100 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1Frequency (half−cycles/second)GPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsQuantitySpeedTable: Performance of GPU Code 1Performance MetricsData Input RateReference Implementaiton744 MB/sNo extra channels540 MB/sHardcoding FIR size756 MB/sHardcoded FIR, Hand Tuned Block Size 890 MB/sUsing 1/10 cost 470GTX637 MB/s1 As of August 1stGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsIncreasing SpeedsHow to proceed:◮ Profiling, and extensive testingGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsIncreasing SpeedsHow to proceed:◮ Profiling, and extensive testing◮ Effective use of CUDA resourcesGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsIncreasing SpeedsHow to proceed:◮ Profiling, and extensive testing◮ Effective use of CUDA resources◮ Waiting for new cardsGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsConclusions◮ CUDA makes heavyweight software processing possibleGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsConclusions◮ CUDA makes heavyweight software processing possible◮ Real time processing is possible at lower ratesGPU Based Polyphase Filter Banks for VLBI

Introduction Implemenation Results Future Work ConclusionsConclusions◮ CUDA makes heavyweight software processing possible◮ Real time processing is possible at lower rates◮ GPU processing is applicable to Astronomy and scientificcomputationGPU Based Polyphase Filter Banks for VLBI