Fraunhofer-Gesellschaft

Publica

Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Portability of the fragmented l1-norm transform for massively parallel processing

 
: Scholz, T.; Rosenberger, M.; Notni, G.

:

Rosenberger, M. ; Society of Photo-Optical Instrumentation Engineers -SPIE-, Bellingham/Wash.:
Photonics and Education in Measurement Science 2019 : 17-19 September 2019, Jena, Germany
Bellingham, WA: SPIE, 2019 (Proceedings of SPIE 11144)
ISBN: 978-1-5106-2981-3
ISBN: 978-1-5106-2982-0
Paper 111441A, 5 pp.
Conference "Photonics and Education in Measurement Science" <2019, Jena>
English
Conference Paper
Fraunhofer IOF ()

Abstract
White light interferometry is a major optical non-contact and therefore nondestructive testing method for nanostructures and surface reconstruction. By the reason of scanning, the data throughput is high and the resulting data stack can exceed gigabytes of raw data. Effective data compression was realized in an FPGA early in the signaling cascade. On the one hand this can significantly boost the achievable data throughput from the sensor, on the other hand the compression results in fragmented raw data with non-equidistant sampling steps and is therefore incompatible with FFT based reconstruction algorithms. In order to face this issue the fragmented l1-norm transform (flot) was developed. The flot reconstruction algorithm is a symbiosis of the l1-norm known from compressive sensing and additionally the wavelet-transform. In contrast to the traditional wavelet-transform the flot algorithm has no dependence on FFT and can quickly handle non-equidistant sampled data. Raw data is heavily independent between pixels in white light interferometry by design. Therefore, implementing the re-construction algorithm on massively parallel hardware is promising. In the last decade this usually meant data processing on GPUs. Nowadays alternatives in the form of affordable CPU clusters or easy to program FPGAs gain importance. OpenCL is a framework to accelerate highly parallel problems on all of the three platforms. In this paper the implementation of the flot algorithm in OpenCL will be explained, compared by speed and power consumption and categorized for suitable use-cases.

: http://publica.fraunhofer.de/documents/N-574975.html