Machine Independent Optimization¶
TODO
Machine Dependent Optimization¶
python-moa has the ability to have multiple backends. Generally these optimizations fall under loop optimizations. Wikipedia has a good overview approaches are polyhedral and unimodular transformation. These issues of optimization fall outside of the scope of work of MOA and we need to learn how to use other’s approaches. Some of the optimizations that we could expect the framework to perform.
loop fusion
loop ordering
loop reduction (psi reduction)
reorder expressions (distributive property +-*/)
dimension lifting for parallelizing algorithms
cache intermediate results (immediately obvious in multiple matrix multiply and fast fourier transform)
Some frameworks claim to handle pieces of this problem so they should be looked into.