Report Number: CS-TR-71-239
Institution: Stanford University, Department of Computer Science
Title: Large [g,d] sorting networks
Author: Van Voorhis, David C.
Date: August 1971
Abstract: With only a few exceptions the minimum-comparator N-sorter
networks employ the generalized "divide-sort-merge" strategy.
That is, the N inputs are divided among g $\geq$ 2 smaller
sorting networks -- of size $N_1,N_2,...,N_g$, where $N =
\sum_{k=1}^{g} N_k$ -- that comprise the initial portion of
the N-sorter network. The remainder of the N-sorter is a
comparator network that merges the outputs of the $N_1-,
N_2-, ...,$ and $N_g$-sorter networks into a single sorted
sequence. The most economical merge networks yet designed,
known as the "[g,d]" merge networks, consist of d smaller
merge networks -- where d is a common divisor of
$N_1,N_2,...,N_g$ -- followed by a special comparator network
labeled a "[g,d] f-network." In this paper we describe
special constructions for $[2^r,2^r]$ f-networks, r > 1,
which enable us to reduce the number of comparators required
by a large N-sorter network from $.25N {log_2 N)}^2 -
.25N(log_2 N) + O(N) to .25N{(log_2 N)}^2 - .37N(log_2 N) +
O(N)$.
http://i.stanford.edu/pub/cstr/reports/cs/tr/71/239/CS-TR-71-239.pdf