A Functional Methodology for Parallel Image Processing Development

ion abstraction parallelisation translation translation translation system transformation user transformation third party transformation A L application in language L HL harness in language L Figure 1. intermediate ML development ML software to C++ (this stage has not yet been automated, though previous work shows it is possible [5]), and this is done at the end of the development cycle. Translating at the earlier stage is pure sequential ML prototyping, though it is then also possible to parallelise within C++ using the proposed method by interfacing to the harness C++ API which is the same as the ML API. For C++ coders, the method introduces a new module, but one providing flexibility and composability, deriving from the functional way it was conceived. 3 SIMPLE EXAMPLE Before introducing the application under discussion, it is useful to illustrate the functional style by the tractable and complete expression of a convolution algorithm. fun convolve2D (src, filter) = let (* get sizes of src image and filter *) val (sh,sw) = dimensions src; val (fh,fw) = dimensions filter; (* produces a single element of convolution matrix *) fun single_convolve (r0,c0) = let (* multiply corresponding filter and src elements *) (* offset into src is (r0,c0) *) val product = tabulate (fh,fw,fn(i,j) => sub(src,i+r0,j+c0) * sub(filter,i,j) ); in (* add together elements of product matrix *) (* reduction/fold operation is add *) fold ( fn (a,b) => a + b ) 0.0 product; end; in (* result is size of src minus size of filter *) tabulate ( (sh-fh)+1, (sw-fw)+1, single_convolve ); end;