a gpu implementation of extragalactic radio source detection a cuda approach

A GPU Implementation of Extragalactic Radio Source Detection

A CUDA Approach

Gary Resnick Scott Badenhorst

Department of Computer ScienceUniversity of Cape Town

17 March, 2010

Introduction

Approach

Plan

Outcomes

Overview

Radio Astronomy

A radio image of the sky

Radio Interferometry

What is a Source?

PipelineTelescope Array

Correlation

Converting to frequency space

Cleaning the dirty map

Source Extraction

Source Extraction

Source Extraction PipelineReading in the data

Compute Threshold

Apply threshold

Blob detection

Generate Catalogue

FITS Data Cube

Blob Detection (XY)

Characterisation (Z)

HI (Neutral Hydrogen)

AIPS◦ IMFIT◦ IMSAD

Source Extractor◦ Deblend

Duchamp◦ Own variation

Related Work

PlanBack down to Earth

Can the GPU be utilized for efficient image source extraction?

How well does a GPU solution scale

to handle large image data volumes?

Can we improve the processing time of current source extraction software?

Research Questions

The accuracy of the GPU implementations are greater than or equate to current extraction algorithms ◦ Sufficient reliability of faint source detection

Spurious detections (Type I error) are constrained by a p-value of 0.05

Attaining a speedup on the GPU implementations in the order of 10x or greater than traditional CPU implementations

Key Success Factors

Work Allocation

Input Handling• L

ogical segmentation of data cube

Thresholding• D

etermine a threshold cut-off value

Source Extraction• C

PU (Serial) & GPU (Parallel) Blob detection

Validation & Filtration• S

plit/merge sources and check errors

Catalogue Generation• C

alculate source parameters (velocity, position, etc) and create listing

TeamIndividually

Each member will independently implement a CPU implementation of a source extraction algorithm and then port it to the GPU

Source Extraction

Gary

•Duchamp• GPU

Scott

•SeXtractor• GPU

Interfaces with Framework

Source Extraction

Segments (grids) of pixels

List of sources

Validation & Filtration

Validated source list

Factor Probability Impact RiskFailure to acquire data low low lowStakeholder departure low high lowData/code loss low high mediu

mUnsuccessful module implementation medium mediu

mmedium

Hardware failure low medium

high

Cascading delay medium high medium

Risks

Risk Factors

Radio Survey Data◦ Synthesized◦ Real Sky Map

Implemented in C/C++ & CUDA on GNU/Linux◦ CFITSIO & WCSLIB Libraries

Developed on NVIDIA (CUDA Compliant) Fermi architecture◦ High-end machine with current gen NVIDIA card(s)

People

Resources Required

Outcomes

Software developed for source extraction and catalogue generation.

Two independent reports A website and poster Presentation to demonstrate the

project.

Deliverables & Milestones

Open Source◦ Distributed under GNU GPL

Akin to Duchamp & SeXtractor GNU/Linux Compatibility

◦ Comparable performance/accuracy across distros Institutions maintain data ownership

◦ Retains rights to catalogues produced Data Integrity

◦ Standardized input and output formats

Ethical, Professional & Legal Issues

We present an approach to solving a problem within Astronomy that needs to be addressed

Our approach is well thought out and relevant

A balanced and fair work load has been allocated to each member

The project has minimal unmitigated risks

Conclusion

Questions?

Project Timeline