ASC Group Meeting Telecon

1
ASC Group Meeting / Telecon

• Adam Jacobs
• February 6, 2007

2
Outline

• Hyperspectral Imaging (HSI)
• HSI Profiling
• Hardware Performance Estimates
• Auto-Correlation
• Target Detection
• Synthetic Aperture Radar (SAR)
• SAR Profiling
• Hardware Performance Estimates
• Range Compression
• Range Migration
• Azimuth Compression

3
Hyperspectral Imaging

• Three Main Tasks
• Auto-Correlation Sample Matrix (ACSM) calculation
• Extract covariance information from the acquired
  image
• Matrix Inverse / Weight Computation
• Combine covariance information and knowledge
  about possible targets to generate weights for
  the target detection algorithm
• Target Detection
• Linear combination of data samples using the
  weights previously calculated

4
HSI Profiling

• ACSM
• Largest percentage of execution time (gt80)
• Target Detection
• Another significant computational section (20)
• Matrix Inverse / Weight Computation
• Only significant for small image sizes
• Takes 40 of execution time for 64 x 64 images
• Takes 0.5 of execution time for 512 x 512 images

5
ACSM Calculation

• Autocorrelation Sample Matrix
• Each pixel in a given pixel gives an independent
  contribution to the final ACSM
• For each inner-loop, an L-element vector will
  create an L2-element matrix
• The summation of the matrices gives the ACSM

6
ACSM Bandwidth Analysis
7
Performance Estimation

• From a bandwidth perspective, ACSM is a good
  kernel for a hardware implementation
• The amount of data that must be transferred to
  and from the FPGA is reasonable
• A large amount of data must be transferred to the
  FPGA but only a fraction must actually be
  returned to the host processor
• How realistic is 500x speedup?
• Not very, but thats OK
• Now that we know the system architecture isnt a
  bottleneck we can examine the actual FPGA
  architecture

8
Hyperspectal Imaging - ACSM

• Possible FPGA Implementation
• Store an intermediate matrix containing the
  running summation of previous matrices on the
  FPGA or nearby memory
• The main function in this kernel is an
  outer-product which exhibits a large amount of
  inherent parallelism
• The ACSM matrix for each pixel takes L2/P cycles
  to compute
• L Spectral Bands
• P processing elements
• 3.5x speedup running at 100 MHz
• L 512, P 4

9
HSI - Target Detection

• Main Computational Kernel
• Dot Product
• Data Reduction
• Each pixel originally contains information about
  L spectral bands
• At the end of computation, each pixel will have a
  single value

10
Synthetic Aperture Radar (SAR)

• 4 Major Computational Tasks
• Range Filtering
• Azimuth Transform
• Range Migration
• Azimuth Compression
• Corner turns are necessary between certain tasks
  so that needed data is in contiguous memory
• This fact also isolates tasks so that it is
  difficult to combine them in hardware

11
SAR Profiling

• Each of the four tasks takes a substantial amount
  of time
• Range Filtering, Azimuth Transform, and Azimuth
  Compression are similar
• FFT followed by multiplication
• Or multiplication followed by IFFT
• Azimuth Compression is slightly more
  computationally intensive
• A new filter must be calculated for each range
• Range Migration is an interpolation kernel

12
Range Compression
13
Azimuth Transform
14
Range Migration
15
Azimuth Compression