國立高雄大學圖資館 |

登入

回首頁

Conditional techniques for stream processing kernels

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Conditional techniques for stream processing kernels
作者:	Kapasi, Ujval J.
面頁冊數:	155 p.
附註:	Adviser: William J. Dally.
附註:	Source: Dissertation Abstracts International, Volume: 65-04, Section: B, page: 1954.
Contained By:	Dissertation Abstracts International65-04B.
標題:	Computer Science.
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3128659
ISBN:	0496759213

Conditional techniques for stream processing kernels
Kapasi, Ujval J.

Conditional techniques for stream processing kernels [electronic resource] - 155 p.

Adviser: William J. Dally.

Thesis (Ph.D.)--Stanford University, 2004.

The stream programming model casts applications as a set of sequential data streams that are operated on by data-parallel computation kernels. Previous work has shown that this model is a powerful representation for media processing applications, such as image- and signal-processing, because it captures the locality and concurrency inherent to an application. This careful handling of important application properties results in kernels that are compute-intensive---i.e., kernels that perform a large number of arithmetic operations per unit of inter-kernel communication bandwidth. Furthermore, the stream model can be implemented with efficient VLSI designs, such as the Imagine Programmable Stream Processor. The Imagine chip supports 48 ALUs on a single die, operating at over 200MHz. This large number of ALUs provides a high peak performance, but makes efficiently executing kernels with conditional code a challenge.

ISBN: 0496759213Subjects--Topical Terms:

212513
Computer Science.

Conditional techniques for stream processing kernels
LDR:02948nmm _2200289 _450 001 162642
005 20051017073508.5
008 230606s2004 eng d
020 $a 0496759213
035 $a 00149143
035 $a 162642
040 $a UnM $c UnM
100 0 $a Kapasi, Ujval J. $3 227786
245 1 0 $a Conditional techniques for stream processing kernels $h [electronic resource]
300 $a 155 p.
500 $a Adviser: William J. Dally.
500 $a Source: Dissertation Abstracts International, Volume: 65-04, Section: B, page: 1954.
502 $a Thesis (Ph.D.)--Stanford University, 2004.
520 # $a The stream programming model casts applications as a set of sequential data streams that are operated on by data-parallel computation kernels. Previous work has shown that this model is a powerful representation for media processing applications, such as image- and signal-processing, because it captures the locality and concurrency inherent to an application. This careful handling of important application properties results in kernels that are compute-intensive---i.e., kernels that perform a large number of arithmetic operations per unit of inter-kernel communication bandwidth. Furthermore, the stream model can be implemented with efficient VLSI designs, such as the Imagine Programmable Stream Processor. The Imagine chip supports 48 ALUs on a single die, operating at over 200MHz. This large number of ALUs provides a high peak performance, but makes efficiently executing kernels with conditional code a challenge.
520 # $a We will introduce two techniques for efficiently supporting kernel conditionals, such as if-statements, case-statements, and while-loops: conditional routing and conditional streams. Conditional routing is a code transformation that exploits the trade-off of increasing inter-kernel communication in order to increase the performance of kernel inner-loops containing conditional code. The second technique we will discuss is the use of a conditional stream, which is a mechanism to reduce the amount of load-imbalance that arises between parallel processing clusters on a single stream processor chip. Load-imbalance results when different conditional paths are taken on different processing clusters, and causes one or more of the clusters to wait idle for the others to complete a kernel. We will also present a case study of the impact of these techniques on a programmable polygon rendering pipeline that contains many unpredictable conditionals. We show that our techniques improve the performance of this application by 1.9x.
590 $a School code: 0212.
650 # 0 $a Computer Science. $3 212513
650 # 0 $a Engineering, Electronics and Electrical. $3 226981
690 $a 0544
690 $a 0984
710 0 # $a Stanford University. $3 212607
773 0 # $g 65-04B. $t Dissertation Abstracts International
790 $a 0212
790 1 0 $a Dally, William J., $e advisor
791 $a Ph.D.
792 $a 2004
856 4 0 $u http://libsw.nuk.edu.tw/login?url=http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3128659 $z http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3128659