ROS Theses Repository

View Item 
  •   ROS Home
  • Mathematical & Computer Sciences
  • Doctoral Theses (Mathematical & Computer Sciences)
  • View Item
  •   ROS Home
  • Mathematical & Computer Sciences
  • Doctoral Theses (Mathematical & Computer Sciences)
  • View Item
  •   ROS Home
  • Mathematical & Computer Sciences
  • Doctoral Theses (Mathematical & Computer Sciences)
  • View Item
  • Admin
JavaScript is disabled for your browser. Some features of this site may not work without it.

Data layout types : a type-based approach to automatic data layout transformations for improved SIMD vectorisation

View/Open
SinkarovsA_0815_macs.pdf (2.114Mb)
Date
2015-08
Author
Šinkarovs, Artjoms
Metadata
Show full item record
Abstract
The increasing complexity of modern hardware requires sophisticated programming techniques for programs to run efficiently. At the same time, increased power of modern hardware enables more advanced analyses to be included in compilers. This thesis focuses on one particular optimisation technique that improves utilisation of vector units. The foundation of this technique is the ability to chose memory mappings for data structures of a given program. Usually programming languages use a fixed layout for logical data structures in physical memory. Such a static mapping often has a negative effect on usability of vector units. In this thesis we consider a compiler for a programming language that allows every data structure in a program to have its own data layout. We make sure that data layouts across the program are sound, and most importantly we solve a problem of automatic data layout reconstruction. To consistently do this, we formulate this as a type inference problem, where type encodes a data layout for a given structure as well as implied program transformations. We prove that type-implied transformations preserve semantics of the original programs and we demonstrate significant performance improvements when targeting SIMD-capable architectures.
URI
http://hdl.handle.net/10399/2880
Collections
  • Doctoral Theses (Mathematical & Computer Sciences)

Browse

All of ROSCommunities & CollectionsBy Issue DateAuthorsTitlesThis CollectionBy Issue DateAuthorsTitles

ROS Administrator

LoginRegister
©Heriot-Watt University, Edinburgh, Scotland, UK EH14 4AS.

Maintained by the Library
Tel: +44 (0)131 451 3577
Library Email: libhelp@hw.ac.uk
ROS Email: open.access@hw.ac.uk

Scottish registered charity number: SC000278

  • About
  • Copyright
  • Accessibility
  • Policies
  • Privacy & Cookies
  • Feedback
AboutCopyright
AccessibilityPolicies
Privacy & Cookies
Feedback
 
©Heriot-Watt University, Edinburgh, Scotland, UK EH14 4AS.

Maintained by the Library
Tel: +44 (0)131 451 3577
Library Email: libhelp@hw.ac.uk
ROS Email: open.access@hw.ac.uk

Scottish registered charity number: SC000278

  • About
  • Copyright
  • Accessibility
  • Policies
  • Privacy & Cookies
  • Feedback
AboutCopyright
AccessibilityPolicies
Privacy & Cookies
Feedback