Uniform but non-local DAGS: a trade-off between pure systolic and SIMD solutions
Tanguy Risset, Yves Robert · 2002
The authors derive processor arrays which are synthesized from uniform but non-local DAGs. They introduce a scope-b broadcast transformation that amounts working with dependence vectors of 'length' b. The parameter b can be adjusted to cope with current integration constraints. They explain the transformation with the Gaussian elimination algorithm. For instance with b=3, they derive an array which has the same number of cells as the Ahmed-Delosme-Morf array but whose execution time is /sup 7n///sub 3/+o(n), as opposed to 3n+o(n). They also apply the scope-b broadcast technique to synthesize faster processor arrays for the algebraic path problem.>