Automatic Parallelism in Differentiation of Fourier Transforms.

Abstract

For functions given in the form of a computer program, automatic differentiation is an efficient technique to accurately evaluate the derivatives of that function. Starting from a given computer program, automatic differentiation generates another program for the evaluation of the original function and its derivatives in a fully mechanical way. While the efficiency of this black box approach is already high as compared to numerical differentiation based on divided differences, automatic differentiation can be applied even more efficiently by taking into account high-level knowledge about the given computer program. We show that, in the case where the function involves a Fourier transform, the degree of parallelism in the program generated by automatic differentiation can be increased leading to a rich set of automatic parallelization strategies that are not available when employing a black box automatic parallelization approach. Experiments of the new automatic parallelization approach are reported on a SunFire 6800 server using up to 20 processors.