Stencil computation research project presentation #1
- 1. Stencil Computation Research Project Jishnu P | Reshmi Mitra Presentation #1 | Date: 11-Jul-2017
- 2. The Agenda Discuss 2 or 3 Optimization techniques from An Auto-Tuning Framework for Parallel Multicore Stencil Computations
- 3. Optimizations Techniques used in the auto tuning framework. Several common optimizations have been implemented in the framework as AST transformations, including ● Loop Unrolling ● Cache Blocking ● Arithmetic Simplification
- 12. Cache Blocking To expose temporal locality and increase cache reuse
- 13. Cache Blocking ● An important class of algorithmic changes involves blocking data structures to fit in cache. ● By organizing data memory accesses, one can load the cache with a small subset of a much larger data set. ● The idea is then to work on this block of data in cache. ● By using/reusing this data in cache we reduce the need to go to memory (reduce memory bandwidth pressure).
- 14. An example.
- 15. Example contd...
- 25. AST - Abstract Syntax Tree ● Abstract syntax trees are data structures widely used in compilers, due to their property of representing the structure of program code. ● An AST is usually the result of the syntax analysis phase of a compiler. ● It often serves as an intermediate representation of the program through several stages that the compiler requires, and has a strong impact on the final output of the compiler.
- 26. AST example
- 27. These were some of the serial optimizations. ● Although the current set of optimizations may seem identical to existing compiler optimizations, future strategies such as memory structure transformations will be beyond the scope of compilers, since such optimizations are specific to stencil-based computations. ● Additionally, the fact that the framework’s transformations yield code that outperforms compiler-only optimized versions shows compiler algorithms cannot always prove that these (safe) optimizations are allowed. ● Thus, a domain-specific code generator run by the user has the freedom to implement transformations that a compiler may not.
- 29. Parellel Optimization ● The shared-memory parallel code generators leverage the serial code generation routines to produce the version run by each individual thread. ● Since the parallelization strategy influences code structure, the AST — which represents code run on each individual thread — must be modified to reflect the chosen parallelization strategy. ● The parallel code generators make the necessary modifications to the AST before passing it to the serial code generator.
- 30. Stencil auto-tuning framework flow
- 31. References ● http://people.csail.mit.edu/cycha n/papers/ipdps10.pdf ● https://en.wikipedia.org/wiki/Abs tract_syntax_tree ● https://www.youtube.com/watch ?v=SfV8aRX0YY0 ● https://software.intel.com/en-us/ articles/cache-blocking-techniqu es
- 32. Sometimes it is good to revisit our learnings. It helps to be a good competitor and also to be prepared for grabbing opportunities. Thank you