[Colloquium] Reminder: Wingerter/MS Presentation/Jun 7, 2017

[Margaret Jaffey via Colloquium]

This is a reminder about Joe Wingerter's MS Presentation tomorrow.

------------------------------------------------------------------------------
Date:  Wednesday, June 7, 2017

Time:  1:00 PM

Place:  Ryerson 277

M.S. Candidate:  Joseph Wingerter

M.S. Paper Title: Lambda cu — An Intermediate Representation for
Compiling Nested Data Parallelism

Abstract:
GPUs provide more raw processing power than has been accessible in the
past, but making effective use of this processing power is a
forbidding programming task. With high memory bandwidth and vast
capacity for arithmetic operations, GPUs are well-equipped for
executing data-parallel programs, but the programming patterns
required to hide memory latency impose strict requirements on the
design of GPU programs, and using the low-level hardware programming
interfaces remains very difficult. The Nessie compiler facilitates
pro- gramming for data-parallel GPGPU applications by implementing a
translation from NESL–a high-level, strict, functional programming
language–to efficient low-level C++ code invoking nVidia CUDA kernels
on the GPU. The compiler achieves efficiency in generated code through
optimizations on a specialized intermediate representation, λcu,
which constitutes the primary contribution of this work. λcu
represents arrays as the application of per-element generator
functions to index spaces, separates serial and parallel portions of
input programs, and makes explicit a set of building blocks for
parallel array operations. The most significant optimiza- tion pass in
the Nessie compiler fuses together parallel array operations.
Alternative choices of which operations to fuse are evaluated by
constructing integer linear programming instances using a heuristic
model of performance benefit associated with each optimization
opportunity, constrained by the compatibility relation between
different array operations. This approach al- lows Nessie to avoid
exhaustively searching the space of potential optimized programs,
which is combinatorially large due to incompatibilities between fusion
choices.

Joseph's advisor is Prof. John Reppy

Login to the Computer Science Department website for details:
 https://www.cs.uchicago.edu/phd/ms_announcements#wings

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey            margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 156)               (773) 702-6011
The University of Chicago      http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=