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 Systems Research Group Seminar
SUMMARY:The Lift Project: Performance Portable GPU Code ge
 neration via Rewrite Rules - Michel Steuwer (Unive
 rsity of Edinburgh)
DTSTART;TZID=Europe/London:20161209T110000
DTEND;TZID=Europe/London:20161209T123000
UID:TALK69538AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/69538
DESCRIPTION:Computers have become increasingly complex with th
 e emergence of heterogeneous hardware combining mu
 lticore CPUs and GPUs. These parallel systems exhi
 bit tremendous computational power at the cost of 
 increased programming effort resulting in a tensio
 n between performance and code portability. Typica
 lly\, code is either tuned in a low-level imperati
 ve language using hardware-specific optimizations 
 to achieve maximum performance or is written in a 
 high-level\, possibly functional\, language to ach
 ieve portability at the expense of performance.\n 
 \nIn this talk\, I will present our novel approach
  Lift which aims for combining high-level programm
 ing\, code portability\, and high-performance. Sta
 rting from a high-level functional program we appl
 y semantic preserving rewrite rules to transform i
 t into a low-level functional representation\, clo
 se to the OpenCL programming model\, from which ev
 entually OpenCL code is generated. Our rewrite rul
 es define a space of possible implementations whic
 h we automatically explore for generating hardware
 -specific OpenCL implementations.\n \nI will discu
 ss the design and implementation of the Lift langu
 age and compiler showing the advantages of a funct
 ional internal compiler representation for high-pe
 rformance code generation. I will show experimenta
 l results of the high-performance OpenCL code gene
 rated with Lift. Our experiments show how hardware
  specific implementations are automatically derive
 d from simple functional high-level algorithmic ex
 pressions offering performance on a par with highl
 y tuned code for multicore CPUs and GPUs written b
 y experts.\n\nBio: Michel Steuwer is a Postdoctora
 l Research Associate in the CArD group at the Univ
 ersity of Edinburgh. He received his PhD in 2015 f
 rom the University of Muenster in Germany. His res
 earch interests span all areas of parallel program
 ming from languages and programming models to thei
 r implementation in compilers and libraries as wel
 l as their execution at runtime. Recently\, his re
 search has particularly focused on structured para
 llel programming models\, heterogeneous and GPU co
 mputing\, and novel compilation techniques. He is 
 the principle designer of the skeleton library Ske
 lCL (github.com/skelcl) and the Lift language and 
 compiler (github.com/lift-project).
LOCATION:FW11\, Computer Laboratory\, William Gates Buildin
 g
CONTACT:Liang Wang
END:VEVENT
END:VCALENDAR