TY - JOUR
T1 - Fast electromagnetic integral-equation solvers on graphics processing units
AU - Li, Shaojing
AU - Chang, Ruinan
AU - Boag, Amir
AU - Lomakin, Vitaliy
N1 - Funding Information:
This work was supported in part by NSF CIAN ERC and
Funding Information:
by United States – Israel Binational Science Foundation Grant No. 2008077.
PY - 2012
Y1 - 2012
N2 - A survey of electromagnetic integral-equation solvers, implemented on graphics processing units (GPUs), is presented. Several key points for efficient GPU implementations of integral-equation solvers are outlined. Three spatial-interpolation-based algorithms, including the Nonuniform-Grid Interpolation Method (NGIM), the box Adaptive-Integral Method (B-AIM), and the fast periodic interpolation method (FPIM), are described to show the basic principles for optimizing GPU-accelerated fast integral-equation algorithms. It is shown that proper implementations of these algorithms lead to very high computational performance, with GPU-CPU speed-ups in the range of 100-300. Critical points for these accomplishments are (i) a proper arrangement of the data structure, (ii) an on-the-fly approach, trading excessive memory usage with increased arithmetic operations and data uniformity, and (iii) efficient utilization of the types of GPU memory. The presented methods and their GPU implementations are geared towards creating efficient electromagnetic integral-equation solvers. They can also find a wide range of applications in a number of other areas of computational physics.
AB - A survey of electromagnetic integral-equation solvers, implemented on graphics processing units (GPUs), is presented. Several key points for efficient GPU implementations of integral-equation solvers are outlined. Three spatial-interpolation-based algorithms, including the Nonuniform-Grid Interpolation Method (NGIM), the box Adaptive-Integral Method (B-AIM), and the fast periodic interpolation method (FPIM), are described to show the basic principles for optimizing GPU-accelerated fast integral-equation algorithms. It is shown that proper implementations of these algorithms lead to very high computational performance, with GPU-CPU speed-ups in the range of 100-300. Critical points for these accomplishments are (i) a proper arrangement of the data structure, (ii) an on-the-fly approach, trading excessive memory usage with increased arithmetic operations and data uniformity, and (iii) efficient utilization of the types of GPU memory. The presented methods and their GPU implementations are geared towards creating efficient electromagnetic integral-equation solvers. They can also find a wide range of applications in a number of other areas of computational physics.
KW - Computational electromagnetics
KW - electromagnetic analysis
KW - graphics processing units
KW - high performance computing
KW - integral equations
UR - http://www.scopus.com/inward/record.url?scp=84862788046&partnerID=8YFLogxK
U2 - 10.1109/MAP.2012.6348120
DO - 10.1109/MAP.2012.6348120
M3 - ???researchoutput.researchoutputtypes.contributiontojournal.article???
AN - SCOPUS:84862788046
SN - 1045-9243
VL - 54
SP - 71
EP - 87
JO - IEEE Antennas and Propagation Magazine
JF - IEEE Antennas and Propagation Magazine
IS - 5
M1 - 6348120
ER -