BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20210916T132455Z
LOCATION:Jean Calvin
DTSTART;TZID=Europe/Stockholm:20210709T153000
DTEND;TZID=Europe/Stockholm:20210709T160000
UID:submissions.pasc-conference.org_PASC21_sess116_msa131@linklings.com
SUMMARY:MLIR as a Framework for DSLs for Weather and Climate
DESCRIPTION:Minisymposium\n\nMLIR as a Framework for DSLs for Weather and 
 Climate\n\nGysi\n\nMachine learning today drives the development of domain
 -specific hardware and software solutions. Developers write their programs
  in a high-level language and rely on a framework for the target-specific 
 code generation. This separation of concerns is essential for productivity
  and enables code generation for an ever-growing range of targets. Traditi
 onally weather codes are written in Fortran, mixing domain-specific algori
 thms and target-specific optimizations. This monolithic approach hinders p
 ortability and algorithmic innovations. Efforts to improve the situation f
 ailed to find wide adoption since developing a universal framework for wea
 ther and climate is costly. The extensible compiler framework MLIR (Multi-
 Level Intermediate Representation) promises to reduce this development cos
 t. It is part of LLVM and shares the same open development model. At its c
 ore, the framework enables the definition of custom intermediate represent
 ations called dialects that, compared to traditional compilers, raise the 
 level of abstraction from basic blocks, phi-nodes, and low-level instructi
 ons to higher-level concepts. Existing dialects, for example, implement li
 near algebra primitives, parallel loops, or vector operations. In theory, 
 MLIR thus reduces the entry barrier for developing a domain-specific compi
 ler and is an opportunity for the weather and climate community to share i
 nfrastructure with the well-founded machine learning community. We designe
 d a stencil dialect and, using MLIR, implemented a vertical prototype that
  lowers our stencil dialect to efficient GPU code. In this talk, we discus
 s the experiences gained during the project and overview recent developmen
 ts relevant to weather and climate.\n\nDomain: CS and Math, Climate and We
 ather
END:VEVENT
END:VCALENDAR