BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Computer Laboratory Computer Architecture Group Me
eting
SUMMARY:Optimization Methods for Accelerator Mapping and H
ardware Design Space Exploration - Grace Dinh\, Un
versity of Berkley
DTSTART;TZID=Europe/London:20240308T130000
DTEND;TZID=Europe/London:20240308T140000
UID:TALK213157AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/213157
DESCRIPTION:Obtaining good performance for tensor problems on
accelerators (both mapping workloads to existing a
ccelerators\, and performing design-space explorat
ion to select hardware parameters) requires optimi
zing an objective function over a large\, nonconve
x space. This objective function represents a perf
ormance metric which may be modeled\, simulated\,
or (if possible) measured. Each such objective fun
ction incurs different tradeoffs in terms of speed
\, accuracy\, and the strength of results that can
be formally proven\, and as a result requires its
own optimization methods.\n\nIn this talk\, I wil
l describe optimization approaches for several suc
h models.\nIn a simple communication model\, we de
rive an unconditional communication lower bound fo
r convolutions and "projective" tensor operations\
, and show that this can always be attained (up to
a constant factor) by solving a mathematical opti
mization problem.\nWe then add additional hardware
constraints to this optimization program\, result
ing in a fast one-shot mapper that encompasses loo
p tiling\, permutation\, and spatio-temporal order
ing.\nFurthermore\, we describe ways to incorporat
e measured or simulated performance results into t
he objective for this mapping\, and discuss recent
progress on solving this problem when these resul
ts may be expensive to collect.\n\nI’ll also descr
ibe ongoing work on extending this approach to mod
el performance\, compute lower bounds\, and determ
ine optimizations for sparse tensor operations.
LOCATION:SS03\, Computer Laboratory\, William Gates Buildin
g
CONTACT:Tobias Grosser
END:VEVENT
END:VCALENDAR