This document summarizes a poster presentation about using GPUs to parallelize synthetic aperture radar (SAR) image generation. It describes a proposed parallel algorithm that divides SAR data into blocks and processes the range and azimuth dimensions concurrently on GPU cores. Experimental results show that an Nvidia Fermi S2050 GPU achieved the highest performance and fastest speeds, processing a 1024x1024 SAR image over 74 times faster than a CPU.

1. Abstract
This poster presents a design for parallel processing of synthetic
aperture radar (SAR) data using multi-core Graphics Processing
Units (GP-GPUs). In this design a two-dimensional image is
reconstructed from received echo pulses and their corresponding
response values. Then the comparative performance of proposed
parallel algorithm implementation is tabulated using different set of
nvidia GPUs i.e. GT, Tesla and Fermi. Analyzing experimentation
results on image of various size 1,024 x 1,024 pixels to 4,096 x
4,096 pixels led us to the conclusion that nvidia Fermi S2050 have
maximum performance throughput .
Introduction
Generating images from Synthetic aperture radar (SAR) video’s
raw data require complex computations that is tabulated in data
sample for RADARSAT-1[1].
In SAR system, fundamental principle is based on matching of
received(Rx) signal phase with the transmitted(Tx) signal phase at
a stable frequency of transmission. Fast Fourier Transform (FFT)
calculation is used to calculate phase of received and transmitted
signal.
The modified algorithm that is presented in the paper [2] is
implemented by dividing the impulse response of Rx/Tx signal into
several blocks of equal length and then computing FFT on each
block. It help to process the dimension of range and azimuth in
parallel.
Since GP-GPU has huge computational capability which process
data in parallel, modified SAR image generation on GPU is
implemented. New proposed parallel algorithm helped to achieve
faster execution speed for range and azimuth dimension.
Algorithm And Method
Data Sample
Implementation
Typical Parallel Algorithm
Proposed Parallel Algorithm
Process distribution to generate SAR
Range Image Generation Timing (1024x1024)
Azimuth Image Generation Timing (1024x1024)
Final SAR Image Generation (1024x1024)
Result - SAR Image Generation
References
[1] Ian G Cumming and Frank H Wong, Digital Processing of Synthetic Aperture Data,
Artech House, Boston, London, Jan 2005.
[2] Bhattacharya C, “Parallel processing of satellite-borne SAR data for accurate and
efficient image formation,” EUSAR 2010, pp. 1046-49, 7-10 June 2010, Aachen,
Germany.
[3] NVIDIA Corporation, Compute Unified Device Architecture (CUDA),
http://developer.nvidia.com/object/cuda.html.
[4] AK Agarwal and et al, “Accelerated SAR image generation on GPGPU platform,”
APSAR, 2011
Multi-core GPU – Fast parallel SAR image generation
Mahesh Khadtare1, Pragati Dharmale2, Prakalp Somwanshi3 & C Bhattacharya4
1 – I2IT, Pune, IN; 2 – SNHU, NH, US; 3 – CRL, Pune, IN; 4 – DIAT, Pune, IN
Parameter Typical Value
Number of range samples 6840
Number of Azimuth samples 4096
Block size (image size) 8192 X 8192
Data size (actual size) 11.27 M samples /sec
Total Computational requirement (range) 1.83 GFLOPS
Total Computational requirement (azimuth) 3.28 GFLOPS
Total load computing requirement 5.11 GFLOPS
NOTE: All calculations in O (GFLOPS)
6840 pixels, 50 km
Range (R)
Azimuth(s)
4096 lines,
3.6 km
RADAR-SAT-1
Data from C-band Sensor
• 50 km Range
• 3.6 km Azimuth
• 7.2 m Ground Range
Resolution
• 5.26 m Azimuth
Resolution
Image (1) Image (2) Image (N)
GPU-PE(1)
Process Image (1) Process Image (2) Process Image (N)
GPU-PE(2) GPU-PE(N)
Image
1
G-PE
1
G-PE
2
G-PE
N
P-Image 1
Image N
G-PE
1
G-PE
2
G-PE
N
P-Image N
Exploit Linearity of Correlation process
1)1(),()(
1
0
 


KiniKnunu
P
i
i
1)1(),()(
1
0
 


KjnjKnyny
Q
j
j
1),()()(
1
0
22
0
1
0
  






KlKnulnylr i
P
j
lK
n
j
Q
j
Perform Block Correlation in Transform domain
     120],[)( *
 KkkYkUIDFTlr jiij
Arrange Correlation vector as output of Matrix
transform
X = A*T
Mathematical Model
Phase-I
Range
Processing
Phase-II
Azimuth
Processing
Input Data Stream Output data stream
Time
Signal
Data
Sensor
RangeFFT
Multiplicationof
Reference
Function
RangeIFFT
Corner-turn
AzimuthFFT
RangeMigration
Multiplicationof
Reference
Function
AzimuthIFFT
Recons-
tructed
Picture
Range Azimuth Range Azimuth
Range Compress Processing Azimuth Compress Processing
The rate of
Execution Time* 34 % 6 % 60 %
*1 Thread,
Conventional Corner Turn
CPU / GPU Core
Details
Time(sec) Speedu
p
AMD Athlon X2 Dual
Core
1 45.337 1x
Fx1800 64 19.50 2.32x
GeForce GT 525M 96 4.150 10.92x
Tesla GeForce GTX 275 192 1.421 31.90x
Fermi S2050 (Multi
GPU)
4x448 0.321 141.21x
0
10
20
30
40
50
AMD
Athlon X2
Dual Core
Fx1800 GeForce
GT 525M
Fermi
S2050
Range Image Processing
0
10
20
30
40
50
60
AMD
Athlon X2
Dual Core
Fx1800 GeForce
GT 525M
Fermi
S2050
Azimuth Image ProcessingCPU / GPU Core
Details
Time(sec) Speedup
AMD Athlon X2 Dual Core 1 50.065 1x
Fx1800 64 36.60 1.368x
GeForce GT 525M 96 8.501 5.884x
Tesla GeForce GTX 275 192 3.725 13.44x
Fermi S2050 (Multi
GPU)
4x448 0.981 51.01x
CPU / GPU Core
Details
Time(sec) Speedup
AMD Athlon X2 Dual Core 1 95.00 1x
Fx1800 64 36.60 2.6x
GeForce GT 525M 96 9.23 10.29x
Tesla GeForce GTX 275 192 4.4563 21.31x
Fermi S2050 (Multi
GPU)
4x448 1.280 74.21x
0
20
40
60
80
100
AMD
Athlon X2
Dual Core
Fx1800 GeForce
GT 525M
Fermi
S2050
SAR Image Processing
Transmit Signal
( M Blocks)
Formation of M*N
Blocks
Received Signal
(N Blocks)
Formation of M*N
Blocks
FFT of M*N Block
vector of Transmit
Signal
FFT of M*N Block vector
of Received Signal
IFFT with Normalization
of correlated M*N blocks of
data
Rearrange the
DATA
Processed O/P Image
contact Name
Mahesh Khadtare: maheshkha@gmail.com
Poster
P5142
Category: Video & Image Processing - Vi02