Development of a stencil compiler for one-dimensional convolution operators on the CM-5

Abstract

Finite difference methods are ideally suited to modern computer architectures and in particular parallel ones. Convolution operators or “stencils” are commonly used to solve a wide variety of problems. These operators can be elegantly written in data-parallel language but the performances obtained can be lower than expected.

With its cache-less vector node architecture, the CM-5 is particularly well suited to this type of operation on large volumes of data. We have developed a code generator that can improve performance over Connection-Machine Fortran (CMF) by a factor of sometimes 3 and reach computation rates in excess of 100 Mflops/node.

We will first describe the performance characteristics of stencil operators expressed in Connection Machine Fortran. Then, based on a performance model of the CM-5 processors, we will describe the three vectorization strategies that were implemented in the code generator.