BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Asia/Tokyo
BEGIN:STANDARD
DTSTART:19510909T000000
TZOFFSETFROM:+1000
TZOFFSETTO:+0900
TZNAME:JST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260408T065611Z
UID:21285AC8-9E89-4C4E-B989-8C09FF763BBB
DTSTART;TZID=Asia/Tokyo:20260521T100000
DTEND;TZID=Asia/Tokyo:20260521T103000
DESCRIPTION:&quot;The increasing diversity of heterogeneous architectures presen
 ts a major compiler challenge\, with current toolchains often being rigid\
 , vendor-specific\, and difficult to extend. To address this\, we are deve
 loping the OFA-Compiler (One For All) ecosystem\, a suite of modular MLIR-
 based tools for heterogeneous targets. This presentation introduces OFA-Fl
 ex\, the component of this ecosystem designed as a design-space exploratio
 n (DSE) framework.\n\nThe core of OFA-Flex is a powerful pattern that deco
 mposes the compilation process into three distinct and modular stages: sch
 eduling\, resource allocation\, and placement. These stages are implemente
 d ashigh-level algorithms operating on MLIR Linalg-on-tensors dialect. To 
 ensure modularity\, a custom dialect is used to communicate decisions betw
 een stages\, allowing them to operate independently yet cohesively. The OF
 A-Flex key innovation is its &quot;pluggable&quot; nature: users can select from a l
 ibrary of provided algorithms for each stage or seamlessly integrate new\,
  custom ones.\n\nThe execution plan generated by OFA-Flex is then handed o
 ff to OFA-Backend. This second component of our ecosystem generates high-p
 erformance code by lowering the abstract operations and linking them to a 
 library of optimized hardware kernels.&quot;\n\nRoom: 711\, Bldg: West Building
  #2\, Kyushu University\, ITO Campus\, 744 Motooka\, Nishi-ku\, Fukuoka\, 
 Fukuoka\, Japan
LOCATION:Room: 711\, Bldg: West Building #2\, Kyushu University\, ITO Campu
 s\, 744 Motooka\, Nishi-ku\, Fukuoka\, Fukuoka\, Japan
ORGANIZER:inoue@ait.kyushu-u.ac.jp
SEQUENCE:16
SUMMARY:Invited talk: Dr. Omar Ragheb and Mr. Stephen Wicklund
URL;VALUE=URI:https://events.vtools.ieee.org/m/553594
X-ALT-DESC:Description: &lt;br /&gt;&lt;div class=&quot;elementToProof&quot;&gt;&quot;The increasing d
 iversity of heterogeneous architectures presents a major compiler challeng
 e\, with current toolchains often being rigid\, vendor-specific\, and diff
 icult to extend. To address this\, we are developing the OFA-Compiler (One
  For All) ecosystem\,&amp;nbsp\;a suite of modular MLIR-based tools for hetero
 geneous targets. This presentation introduces OFA-Flex\, the component of 
 this ecosystem designed as a design-space exploration (DSE) framework.&lt;/di
 v&gt;\n&lt;div class=&quot;elementToProof&quot;&gt;&amp;nbsp\;&lt;/div&gt;\n&lt;div class=&quot;elementToProof&quot;
 &gt;The core of OFA-Flex is a powerful pattern that decomposes the compilatio
 n process into three distinct and modular stages: scheduling\, resource al
 location\, and placement. These stages are implemented ashigh-level algori
 thms operating on MLIR Linalg-on-tensors&amp;nbsp\;dialect. To ensure modulari
 ty\, a custom dialect is used to communicate decisions between stages\, al
 lowing them to operate independently yet cohesively. The OFA-Flex key inno
 vation is its &quot;pluggable&quot; nature: users can select from a library of provi
 ded algorithms&amp;nbsp\;for each stage or seamlessly integrate new\, custom o
 nes.&lt;/div&gt;\n&lt;div class=&quot;elementToProof&quot;&gt;&amp;nbsp\;&lt;/div&gt;\n&lt;div class=&quot;element
 ToProof&quot;&gt;The execution plan generated by OFA-Flex is then handed off to OF
 A-Backend. This second component of our ecosystem generates high-performan
 ce code by lowering the abstract operations and linking them to a library 
 of optimized hardware kernels.&quot;&lt;/div&gt;
END:VEVENT
END:VCALENDAR

