Conclusions

Summary

In this book we have described a pipelined code mapping scheme which can be used to produce machine code to exploit the potential parallelism in application programs intended for static dataflow computers. Our mapping scheme is particularly suitable for compiling the kernel of scientific numerical applications in which strong regularity of array operations is a major characteristic.

The functional language Val is our choice of high level language to express these applications. In particular, we define a subset of Val, called PIPVAL, to represent the programs to be transformed by our basic mapping rules. Two important features of PIPVAL are its array creation constructs: forall and for-construct. With these two constructs, array creation operations on regular structures can be expressed without using Val array append operations. Other array regularities are also utilized by our code mapping scheme, including the form of our array selection indexing scheme.

Pipelined code mapping schemes were described for PIPVAL expressions, with special ttention given to mapping rules for array creation constructs, especially primitive forall and or-construct expressions. Optimizations were presented which take advantage of the regularity in array operations as found in these expressions, removing these operations from the data flow graph or replacing them by ordinary machine operations. In addition, certain related ptimization echniques were discussed which can be used along with the basic mapping scheme to improve ipelining performance.

The code mapping scheme described in this monograph assumes the use of a dataflow omputer having sufficient computing power to exploit the parallelism made available through pipelined execution of machineevel data flow programs. Important factors in architecture design, aricularly instruction design to support pipelining, were also addressed.Finally, we presented xperimental work using pipelined code mapping n the McGill Dataflow Architecture.