Fast Gridding on Commodity Graphics Hardware

Thomas Sangild Sørensen; Tobias Schaeffter; Karsten Østergaard Noe; Michael Schacht Hansen

Fast Gridding on Commodity Graphics Hardware

Thomas Sangild Sørensen, Tobias Schaeffter, Karsten Østergaard Noe, Michael Schacht Hansen

Research output: Contribution to book/anthology/report/proceeding › Conference abstract in proceedings › Research

Abstract

The most commonly used algorithm for non-cartesian MRI reconstruction is the gridding algorithm [1]. It consists of three steps: 1) convolution with a gridding kernel and resampling on a cartesian grid, 2) inverse FFT, and 3) deapodization. On the CPU the convolution step is the far most time consuming of the three steps (Table 1). Modern graphics cards (GPUs) can be utilised as a fast parallel processor provided that algorithms are reformulated in a parallel solution. The purpose of this work is to test the hypothesis, that a non-cartesian reconstruction can be efficiently implemented on graphics hardware giving a significant speedup compared to CPU based alternatives. We present a novel GPU implementation of the convolution step that overcomes the problems of memory bandwidth that has limited the speed of previous GPU gridding algorithms [2].

Original language	English
Title of host publication	Proceedings of ISMRM Workshop on Non-Cartesian MRI
Publication date	2007
Publication status	Published - 2007
Event	Workshop on Non-Cartesian MRI - International Society for Magnetic Resonance in Medicine - Sedona, Arizona, United States Duration: 25 Feb 2007 → 28 Feb 2007

Conference

Conference	Workshop on Non-Cartesian MRI - International Society for Magnetic Resonance in Medicine
Country/Territory	United States
City	Sedona, Arizona
Period	25/02/2007 → 28/02/2007

Cite this

@inbook{c4eae580296011dcbee902004c4f4f50,

title = "Fast Gridding on Commodity Graphics Hardware",

abstract = "The most commonly used algorithm for non-cartesian MRI reconstruction is the gridding algorithm [1]. It consists of three steps: 1) convolution with a gridding kernel and resampling on a cartesian grid, 2) inverse FFT, and 3) deapodization. On the CPU the convolution step is the far most time consuming of the three steps (Table 1). Modern graphics cards (GPUs) can be utilised as a fast parallel processor provided that algorithms are reformulated in a parallel solution. The purpose of this work is to test the hypothesis, that a non-cartesian reconstruction can be efficiently implemented on graphics hardware giving a significant speedup compared to CPU based alternatives. We present a novel GPU implementation of the convolution step that overcomes the problems of memory bandwidth that has limited the speed of previous GPU gridding algorithms [2].",

author = "S{\o}rensen, {Thomas Sangild} and Tobias Schaeffter and Noe, {Karsten {\O}stergaard} and {Schacht Hansen}, Michael",

year = "2007",

language = "English",

booktitle = "Proceedings of ISMRM Workshop on Non-Cartesian MRI",

note = "Workshop on Non-Cartesian MRI - International Society for Magnetic Resonance in Medicine ; Conference date: 25-02-2007 Through 28-02-2007",

}

TY - ABST

T1 - Fast Gridding on Commodity Graphics Hardware

AU - Sørensen, Thomas Sangild

AU - Schaeffter, Tobias

AU - Noe, Karsten Østergaard

AU - Schacht Hansen, Michael

PY - 2007

Y1 - 2007

N2 - The most commonly used algorithm for non-cartesian MRI reconstruction is the gridding algorithm [1]. It consists of three steps: 1) convolution with a gridding kernel and resampling on a cartesian grid, 2) inverse FFT, and 3) deapodization. On the CPU the convolution step is the far most time consuming of the three steps (Table 1). Modern graphics cards (GPUs) can be utilised as a fast parallel processor provided that algorithms are reformulated in a parallel solution. The purpose of this work is to test the hypothesis, that a non-cartesian reconstruction can be efficiently implemented on graphics hardware giving a significant speedup compared to CPU based alternatives. We present a novel GPU implementation of the convolution step that overcomes the problems of memory bandwidth that has limited the speed of previous GPU gridding algorithms [2].

AB - The most commonly used algorithm for non-cartesian MRI reconstruction is the gridding algorithm [1]. It consists of three steps: 1) convolution with a gridding kernel and resampling on a cartesian grid, 2) inverse FFT, and 3) deapodization. On the CPU the convolution step is the far most time consuming of the three steps (Table 1). Modern graphics cards (GPUs) can be utilised as a fast parallel processor provided that algorithms are reformulated in a parallel solution. The purpose of this work is to test the hypothesis, that a non-cartesian reconstruction can be efficiently implemented on graphics hardware giving a significant speedup compared to CPU based alternatives. We present a novel GPU implementation of the convolution step that overcomes the problems of memory bandwidth that has limited the speed of previous GPU gridding algorithms [2].

M3 - Conference abstract in proceedings

BT - Proceedings of ISMRM Workshop on Non-Cartesian MRI

T2 - Workshop on Non-Cartesian MRI - International Society for Magnetic Resonance in Medicine

Y2 - 25 February 2007 through 28 February 2007

ER -

Fast Gridding on Commodity Graphics Hardware

Abstract

Conference

Fingerprint

Cite this