Avoiding Auto-Tuning with the Polyhedral Model

Speaker: Martin Kong (The University of Oklahoma)
Host:       Prof. Marco Alvarez (U. of Rhode Island)
Date:       Friday, April 23, 2021
Time:       3:00 PM to 4:00 PM EST
Location: This talk will be hosted remotely via Zoom.

Abstract:
One of the main tuning practices in high-performance computing (HPC)
and, more generally, in aggressive program optimization, is to conduct
an empirical search in a potentially large candidate space of
transformations to find semantically equivalent program variants that
exhibit superior performance properties. Such auto-tuning schemes are
particularly useful for small and common computational building
blocks, which are typically deployed as high-performance libraries. In
contrast, more general applications, those without a handful of
subroutines that concentrate the bulk of computational work are not
amenable to such exploratory techniques, as this would require
considerable time and compute resources invested in the empirical
exploration of each of its computational kernels. In this talk, I will
briefly recap the core background of the polyhedral model, a formal
model of compilation, and present recent results in compilation and
tuning transformations that substantially reduce the need for
empirical exploration of statically analyzable programs. In
particular, I will discuss how a closed form representation of program
transformations can be navigated with a lexicon of performance
objectives to: a) enable the synthesis of novel loop transformation
sequences for CPUs, and b) facilitate the selection of tile size
parameters for GPUs, while achieving strong performance and tuning
effort reduction.

Short Bio:
Dr. Kong is an Assistant Professor in the School of Computer Science
at The University of Oklahoma, joining the department in August of
2019. He graduated from The Ohio State University in 2016, where he
was a member of the high-performance computing research laboratory. He
was advised by Prof. Louis-Noël Pouchet and Prof. P. Sadayappan. Prior
to joining OU, Dr. Kong spent two years as an Assistant Computational
Scientist at the Computational Science Initiative in Brookhaven
National Laboratory. Before that, he held a post-doctoral research
position in Prof. Vivek Sarkar’s Habanero research group in the
Computer Science Department at Rice University. Dr. Kong’s work seeks
to exploit synergies among programming languages, compilers, computer
hardware and domain specific knowledge to attain high-performance in
today’s evolving complex computing systems. He has served as a member
of the Program Committee (PC) of IPDPS’21, LCPC (2019 and 2020), ICPP
(2020 and 2021), and as a journal reviewer for ACM TOPLAS, ACM TACO,
ACM TOPC and IEEE TPDPS.

Zoom link:
https://uri-edu.zoom.us/j/91929798900?pwd=N1lCNWhyVHU0cW9mTUhMREprdHhaUT09