Share Email Print

Spie Press Book • new

GPU Acceleration for Optical Measurement
Author(s): Tianyi Wang; Qian Kemao
Format Member Price Non-Member Price

Book Description

Optical measurement techniques have been successfully applied to measure various physical quantities in experimental mechanics, material science, biomechanics, etc. To faciliate larger amounts of data and increased accuracy, graphics processing units (GPUs) have entered the mainstream due to their cost-effectiveness, short development cycle, and transparent scalability. This Spotlight not only demonstrates the power of GPUs to accelerate optical measurement algorithms but also provides a hands-on approach for the use of Compute Unified Device Architecture (CUDA) to accelerate existing sequential algorithms on CUDA-capable GPUs. Readers familiar with basic C/C++ programming can integrate CUDA with their existing optical algorithms for higher computing performance.

Book Details

Date Published: 20 December 2017
Pages: 50
ISBN: 9781510617353
Volume: SL34

Table of Contents
SHOW Table of Contents | HIDE Table of Contents

1 Introduction
1.1 Acceleration by parallel computing
1.2 Graphics processing unit
1.3 Heterogeneous computing with CUDA
1.4 Common parallel patterns in optical measurement
1.5 Use of the source code
1.6 Organization of the Spotlight

2 CUDA Basics for GPU Programming
2.1 NVIDIA GPU architecture and CUDA programming model
2.2 CUDA memory hierarchy
2.3 First glimpse: parallel vector addition
2.4 Kernel running time and synchronization
2.5 Timing the vector addition
2.6 Data manipulation
2.7 Pointwise image gradient calculation
2.8 Tiled image gradient calculation

3 Parallel WFF2 Algorithm for Fringe Pattern Analysis
3.1 Basics of WFF2
3.2 Problem decomposition
3.3 Accelerating FFT analytically
3.4 Accelerating FFT with the CUFFT
3.5 Flexible pointwise kernels with grid-stride loops
3.6 Performance verification

4 Parallel IC-GN Algorithm for DIC
4.1 Basics of IC-GN
4.2 Problem decomposition
4.3 Tiled precomputation of the bicubic interpolation coefficients look-up table
4.4 IC-GN iterations
4.5 Parallel reduction for inner-subset sums
4.6 Performance verification

5 Conclusion and Further Reading

Acknowledgments

References

Preface

Due to their advantages of non-destructiveness, high accuracy, and high sensitivity, optical measurement techniques have been successfully applied to measure various important physical quantities in experimental mechanics, materials science, biomechanics, etc. In order to deal with increasingly larger amounts of data and increase accuracy, the computational burden of optical measurement techniques has become heavier. In the past decade, parallel computing devices have been applied to accelerate these techniques, among which graphics processing units (GPUs) have become mainstream due to their high parallelism, cost effectiveness, short development cycle, and transparent scalability. Additionally, compute unified device architecture (CUDA), invented by NVIDIA, provides an easy-to-use C/C++ programming interface, which has opened the possibility to program GPUs without having to learn complex shading languages and the graphics pipeline. This Spotlight not only demonstrates the power of GPUs in accelerating optical measurement algorithms but also provides a hands-on approach for the acceleration of existing sequential algorithms on CUDA-capable GPUs. Readers who understand basic C/C++ programming can then attempt to integrate CUDA with their existing optical algorithms for higher computing performance.

Tianyi Wang
Qian Kemao
December 2017


© SPIE. Terms of Use
Back to Top