Summary

• Scalar optimization:
  • improving ILP by code transformation and grouping independent instructions
  • improving memory access by restructuring loop nests to take better advantage of memory hierarchy
• compilers are good at instruction level optimizations and loop transformations. It depends on the language, however:
  • F77 is the easiest for compiler to work with
  • C is more difficult
  • F90/C++ are most complex for compiler optimizations
• the user is responsible to present the code in a way that allows for compiler optimizations:
  • don’t violate the language standard
  • write clean and clear code
  • consider the data structures for (false) sharing and alignment
  • consider the data structures for data dependencies
  • most natural presentation of algorithms using multi-dimensional arrays

Previous slide

Next slide

Back to first slide

View graphic version