Wed Jul 2 17:14:37 CEST 2008
datatypes and iterators
Starting from the ideas in FP, how can we build a minimalistic algebra
of programs specific for image processing? This means:
* keep the data type simple (tiled images)
* functionals are special cases of map/fold/shift for image ops
starting from this, building a framework for effective loop fusion
should be doable. the problem is composition of shift operators:
combining two convolution maps gives a bigger convolution map. (Is it
possible to work only in terms of the 4 direction unit shifts?)
This can be tested in 1D first, i.e. for block based audio processing.
Algebra of programs. Ingredients:
* binary functions +,-,*
* scalars + vectors
Is it possible to make something smaller than FP?