[Colloquium] Press/MS Presentation/Dec. 7, 2005

Margaret Jaffey margaret at cs.uchicago.edu
Wed Nov 23 09:48:50 CST 2005


This is an announcement of David Press' Master's Presentation.

------------------------
Date:  Wednesday, December 7, 2005

Time:  9:00 a.m.

Place:  Ryerson 255

M.S. Candidate:  David Press

M.S. Paper Title: High-level Language Support for Stream Programming  
on GPUs

Abstract:
Using the streaming program architecture is a natural way to
increase performance on highly-parallel hardware.  However, there is
little language-level support for the paradigm. It is observed that
the model easily maps to the functional programming language paradigm,
so adapting such a language to the streaming model would be clear and
appropriate.  As such, the author extends Standard ML to support the
concept of a streaming ``kernel'', along with the ability to compose
and concatenate kernels to create new kernels.
   The author implements a backend to execute kernels on the GPU using
the BrookGPU compiler and a fallback CPU backend for when the kernel
cannot be executed on the GPU because of its limitations.  The stream
programming paradigm also provides the opportunity for many
optimizations. Any stream program with a kernel that performs
operations that do not depend on the streaming data can be optimized
by performing binding-time analysis to find these operations and
perform them only once at the time of the kernel call.  The author
describes an analysis on an intermediate  representation of a kernel
to detect such redundant computations.  The results of performing the
optimization on sample stream programs are presented.

Advisor:  Prof. John Reppy

A draft copy of David Press's MS Paper is available in Ry 161A.

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey                             margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 161A)        (773) 702-6011
The University of Chicago                  http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=





More information about the Colloquium mailing list