Efficient Sum of Absolute Difference Computation on FPGAs - - PowerPoint PPT Presentation

Efficient Sum of Absolute Difference Computation on FPGAs

International Conference on 
 Field Programmable Logic and Application (FPL)
 2016

Martin Kumm, Marco Kleinlein and Peter Zipf

University of Kassel, Germany

Sum of Absolute Difference (SAD)

SAD is an important operation in image and video processing Metric to measure the distance between two blocks of an image Applications are, e.g., motion estimation or stereo matching An R×C SAD operation of two matrices and is defined as:

2

SAD(A, B) =

R

X

i=1 C

X

j=1

|ai,j − bi,j| A B

Previous Work

3

Sequential AD [1]

CPA

MSB

CPA

Parallel AD [2]

CPA

MSB

CPA

FPGA optimized [3]

CPA

• SAD is computed with N absolute difference (AD) units
• N-input adder tree/compressor tree required
• LUTs of best reported circuit grow with 2.5NB (B: word size)

Proposed SAD

4

• SAD is computed with N/2 1x2 SAD units
• N/2-input adder tree/compressor tree required
• LUTs of proposed SAD grow with 2.0NB (B: word size)

Proposed 1×2 SAD

CPA CPA

MSB MSB

Results

5

2 4 6 8 10 12 14 16 18 20 22 24 0.5 1 1.5 2 ·104 √ N #LUTs SAD sequential [1] SAD parallel [2] SAD Perri [3] SAD proposed Estimations

(a) Required and estimated LUTs

2 4 6 8 10 12 14 16 18 20 22 24 10 20 30 √ N LUT reduction (%) actual LUT reduction limit for N → ∞

(b) Relative LUT reduction compared to [3]

Check out uni_ks git branch of

https://scm.gforge.inria.fr/anonscm/git/flopoco/flopoco.git

See you at the poster at 3:30!

6

Literature:
 [1] Kanoh, Absolute Value Calculating Circuit Having a Single Adder, US Patent US 4,953,115, 1990
 [2] Chirila-Rus (Xilinx Inc.), Determining Sum of Absolute Differences in Parallel, US Patent US 8,131,788, 2012
 [3] Perri, Zicari & Corsonello, Efficient Absolute Difference Circuits in Virtex-5 FPGAs, MELECON 2010