CA2037732C

CA2037732C - Matrix sorting network

Info

Publication number: CA2037732C
Application number: CA002037732A
Authority: CA
Inventors: Ernst August Munter
Original assignee: Northern Telecom Ltd
Current assignee: Nortel Networks Ltd
Priority date: 1990-07-12
Filing date: 1991-03-07
Publication date: 1998-01-06
Anticipated expiration: 2011-03-07
Also published as: JP3194597B2; US5216420A; JPH0630455A; CA2037732A1

Abstract

In a switching system, or the like, an apparatus for sorting N signals comprises log2 N stages of sorting matrices wherein each stage comprises M (N/2M)-by-(N/2M) matrices of sorting cells and the final stage (M=1) comprises an N/2-by-N/2 matrix of sorting cells. In practice, the matrices are almost square (N/2M)-by-([N/2M]+1).

Description

MATRIX SORTING NETWORK

CROSS-REFERENCE TO R~LATED APPLICATION

This application is related to commonly assigned United States Patent Application 07/39~J,925 filed on August 29, lg89 by E.A. Munter and I. Perryman, which is incorporated herein by reference.

BACKGROUND OF THE INVENTION

1. Field of the Invention This invention relates to data processing systems in general, and to packet switching systems in particular. More particularly, this invention relates to sorting networks for use in, for example, asynchronus transfer mode (ATM) switch controllers or computer co-processors.

2. Prior Art of the Invention In ATM packet switching networks internal buffering is often used to resolve output contention, such as in a buffered Banyan network. An alternative to buffering is the use of sorting networks, the most well-known of which is the Batcher bitonic sorter as implemented, for example, in Batcher-Banyan networks.

In ~eneral, in communication systems and computer systems it may often be either necessary or desirable to sort signals received by the system. The sorting results in the system distinguishing, for example, signals of higher priority from signals of lower 203773~

priority, The priority of a signal may be established on the basis, for example, of the magnitude of a particular constituent of it.

Prior art sorters have comprised sets of sub-stages of sorter cells forming sorting modules which increase in size, by a factor of two per stage. Such prior art sorters perform a binary merge operation on pairs of sorted lists. The connection patterns between stages and sub-stages in such prior art sorters are highly 10 regular but are best implemented in a 3-dimensional way, since projections onto a plane requires a large amount of space for such sorters. Such sorters require three or four interconnection layers to accommodate the crossing signals and clock/power. Because of such 15 interconnection topology it is not possible to remove a fractional part of the Batcher sorter.

SUMMARY OF THE INVENTION
,;
It has been found that better performance and greater flexibility may be obtained using sorting 20 networks comprising identical simple sorting cells in a compound matrix arrangement. Specifically, for N inputs already sorted by row and column a sorting matrix of N12 rows of cells and N/2 columns of cells may be used to generate N sorted outputs on its diagonals by merging 25 the two input sets. The presorted input sets can be the outputs of two smaller matrices of the same kind, down to the level of single inputs.

Accordingly, the present invention provides an apparatus for sorting N signals. comprising (log2 N) stages of sorting matrices wherein each stage comprises M (Nl2M)-by-(N/2M) matrices of sorting, M being equal to Nl2, N/4, N/8, ...,N/N.

In a preferred implementation of such apparatus, the sorting matrices are almost square comprising (Nl2M)-by-(lNl2M]+l) sorting matrices.

An advantage of the above described apparatus is 10 that fractional slices of network along either edge between a set of inputs and the corresponding outputs may be removed without affecting the operation of the remaining network. This may facilitate testing and fault diagnosis in that the size of the fully configured 15 network can be dynamically reduced.

A disadvantage of the matrix sorting apparatus of the present invention is that its size grows faster than that of the Batcher bitonic sorter with increasing N;
(size growing as the square of N as opposed to the 20 square of (log2 N) for the Batcher sorter). However, over a certain size range, the matrix sorter of the present invention is much faster. For small N, say around 16 or 32, the hardware requirements of both types of sorters are comparable, but the matrix sorter is 25 significantly faster.

BRIEF DESCRIPTION OF THE DRAWINGS

The preferred embodiment of the present invention 7~32 will now be described in detail in con~unction with the annexed drawings, in which:

Figure l illustrates schematically a prior art 32-port batcher sorting network;

Figure 2 illustrates schematically a 32-port matrix sorting network according to the present invention;

Figure 3 illustrates schematically in more detail the 32-port matrix sorting network shown in Figure 2;

Figure 4 illustrates the 8-by-9 matrix shown in Figure 3 in more detail; and Figure 5 is a block-schematic of the cell identified in Figure 4.

DETAILED DES~RIPTION Of' 'I'flE PREFERRED EMf30DIMENT

Figure l of the drawings shows a prior art 32-port batcher sorting network. whose cells operate essenti.ally as double-pole-double-throw switches with two input and two output ports. The 240 cells, interconnected as shown in the figure. may be used to sort 32 input signals into 32 output signals. Whi]e it has fewer and simpler cells than does the matrix sorter of the present invention~ it exhibi.ts comp]ex interconnect.ion topology, pa~ticularly toward the Olltpl.lt, anll is less f lexil)le.

20377~2 Referring now to ~'igure 2 of the drawings, a so~-t~1lg network 1() for sorting 3) input signaJs (applied at inputs 11a and 11'a to llp and 11'p) according to the present inventlon comprises tive (1~g2 32j stages of sorting matrices. The matrices 12a to 12p, 13a to 13h, 14a to 14d, 15a to 15b, and Ib, wherein the diagonal outputs are row/column inputs to the succeeding matr-ix.
The 32 diagonal outputs O~: the f-inal m.11:rix 1~ (17a and 17'a to 17p and 17'p) represent the outputs of the entire network 10.

In Figure 3. which correspon(1s substa11tial1y ~o Fig1Jrc 2 but shows details ot' interconrlectio1l~ each of the matrices 1'~a - 12p. 13a - 13h. 14a - 14d, 15a - ISh, and 16 are not square matrices, but are, respectively, 1-by-2, 2-by-3, 4-by-5, ~-by-~, and 16-hy-17. The reason for this practical difference from the theoretical square matrices is that it is necessary to split the main diagonal of each matrix into two diagonals in order not to have special cells along the 20 main diagonal of a matrix.

For purposes of explanation, the matrix 15a is shown in more detail in Figure 4. It comprises 8 rows and 9 columns of cells, one of which, designated 18 in Figure 4, is shown in detail in Figure 5. The matrix 25 15a comprises horizontal bus inputs Al to A8 and vertical bus inputs Bl to B8. Diagonal buses C1 to (~16 are the output buses which are applied to the matrix 16 as inputs.

In order to generate 1~ diagonal outputs, the main diagonal C8 of the matrix is split into diagonals C8 and C9 by adding another column of cells, so that the Ib inputs sorte~ by row and co1umn are output, cornpletely sorted by the 8-by-'3 rnatrIx. onto the sixteen diagonal outputs Cl to Clh. Alternatively, "specialized" cells (not shown) along the rnain diagona] C8 of the matrix 15a may be constructed so as to be able to achieve that result. Although. in the N / 2-by-(( N / 2) t 1 i configuration, slightly more cells are used than in the N/2-by-N/2 confi.gurati.on. al1 the cells are exactly alike.

Tlle circui.t and orient.at:ion o~ the sorting cells will now be described with reference to Figures 4 and 5.

The organization of cells (such as 18) in the upper right triangular half of the matri~ 15a is a mirror image, about the line between the diagonal C8 and C9, of the cells in the lower left triangular half of the matrix l5a. All of the cells are identical to the cell 18 shown in Figure 5.

The cell 18 has inputs ~rom input buses A2 and B3.
The cell 18 comprises a comparator 19 which compares the analog or the digitally encoded magnitude of the signals input on A2 and B3. a selector 20 and a switch 21 to connect the selected signal to the diagonal output hus C4.

Many cells are potentiallY connectable to an output d:iagonal bu~ ~I to ~16 and so could be in conflict.

20377~2 Tllerefore~ the cell 18 furtheY comprises two inhibit inputs 2 and 23 and two inhibit outputs 24 and 25.
The two inhibit inputs 22 and 23 and the two inhihit outputs 24 and 25 are connected vertically and hor:izontally between adJacerlt cells such th~lt the flow of control is in the direction of the output diagonals C8 and C~ ~the main diagonals).

The cells around the perimeter of the matrix are not ordinary cells like the cell 18, even though they are identical in structure tor reasorls of economy. They do not receive inhibit inputs trom outside the rnatrix.
Sirnilarly, the cells 26 and 26' alollK the dia~onals C8 arl(l C~, respectively. do not ~cnera~e inhihi~ outputs such that inhibit control does not cross trorr, the diagonal C8 to C~ or vice versa.

The cell 18 independently compares the input received at the buses A2 and B3 and makes a selection as between them by means of the selector 20. The selection is switched to the bus C4 only if NOR gate 27 indicates both inhibit inputs 22 and 23 are OFF, as per the following truth table:

203773~

lnhi~it Inhibit N()~
lnput 2I Input 23 ~ate 27 O O
O l O
5 1 0 ~
O

This. of course. ensures that the diagonal C4 does not receive rnultiple signals at the same time.

At the same time, arId totaIIy independently of the 10 inhibit inputs 22 and 23~ the cell 18 turns one of its inhibit outputs 24 and 25 ON. Specifically~ the horizontal output 24 is turned ON if the signal from bus A2 is selected (in Figure 5~ for example, if A>B) and the vertical output 25 is turned ON if bus B3 is 15 selected (in Figure 5~ for example, if A<B).

Thus the cell 18 selects a bus and inhibits a neighboring cell simultaneously, Such selection and inhibition has the important consequence that no ripple through delay is required. Since selection and inhibit 20 mechanisrns operate in parallel, the network is very fast, requiring on the order of (log2N) cell delays (with N=32 for the network 10), where a cell delay is made up of delaY from the comparator 19 and the selector 20.

g The sense of sorting mav be selected arbitrarilY, to be ascending or descending, but all cells rnust operate in the same sense in the upper triangle, and the opllc)site sense in thc lower tr-iang1e nf the matrix 15a.

In the case of equa]ity of the signals input at the respective inputs A2 and B3 for the cell 18, either signal (it does not matter which) may be selected and the corresponding inhibit output is generated. But generally speaking~ it should be noted that the lO selection criteria may be different from simply comparing magnitude, For example, for digitally packetized signals. certairl priority encoding port~ons may be compared and a decision made accordin~ly to select either signal A or signal B.

Preferrably, the network 10 further comprises drivers or buffers (not shown) between the stages, at least for longer signal paths in the last stage of the sorting network. In the input stages, the busses are very short, such as fanouts of 2, 3, 5 or 9, and the 20 busses may be joined by simple switches, The network lO of the present invention requires all data to be available for the comparisons before the correct inhibit decisions can be made, This is naturally the case where the data is propagated in 25 analog or in parallel digital form. Serial digital transmission on the busses A, B, C and/or between the stages is possible if conversion to parallel takes place in each cell or between stages.

~037732 Serial transmission on the busses A, B, C, combined with serial (l-bit) comparison circuits in the cells considerably reduces the amount of hardware required, potentially allowing large sorting matrices to be built on a single device, but at the expense of added delay.
The extra delay is dependent on the word length to be compared.

The pattern of the sorting network lO is fundamentally two-dimensional, conceptually consisting as it does of only horizontal busses, vertical busses, and one direction of diagonal busses. Clock and power buses blend in right along with these busses. The linear layout also means that growth, while ultimately limited by cost, is not difficult topologically. A
plane may simply be covered with cells and the three busses A, B and C interrupted and spliced together.

In the matrix sorter of thP present invention slices of network lO along either edge between one input or set of inputs and the corresponding outputs can be removed without affecting the operation of the rest of the network lO. This feature may be useful for testing and fault diagnostic purposes in that the size of a fully configured network lO can be dynamically reduced.

Claims

1. A matrix sorting network for sorting N inputs onto N outputs wherein N is an integer of the power of 2, said matrix comprising:

log2N stages;

each said stage comprising M matrices, where M is an integer such that M=(N/2L) for the Lth stage of said matrix sorting network;

each said matrix of each said stage including an (N/2M)-by-((N/2M)+1) array of sorting cells including N/2M rows of said cells and N/2M + 1 columns of said cells wherein said cells of each said matrix define one or more diagonals;

N/2M first input buses extending along said rows of each said matrix and N/2M second input buses extending along said columns of each said matrix;

each said sorting cell having a first cell input means for carrying an input signal having a value and connected to one of said first input buses, a second cell input means for carrying an input signal having a value and connected to one of said second input buses, a comparator for selecting either the said value carried on said first cell input means or the value carried on said second input means by comparing values to one another, and one cell output means carrying a cell output signal of value equal to the selected one of said values;

output buses such that each output bus is connected to the output means of the cells disposed at one diagonal and such that each output bus may receive the cell output signal of connected sorting cells;

at least one of said sorting cells further comprising an inhibit output means for providing a disabling signal having a value dependent upon which one of said values carried on the cell input means of such cell is selected by the comparator of such said cell; and at least one of said sorting cells further comprising a disabling means connected to the inhibit output means of at least one other cell and for preventing the cell output signal of such cell from being conveyed to any of said output buses dependent upon the value of the disabling signal.

2. A matrix sorting network as claimed in claim 1 wherein each said disabling means comprises a first inhibit input means for accepting a disabling signal from the inhibit output means of one other cell, and a second inhibit input means for accepting a disabling signal from the inhibit output means of a second other cell.

3. A matrix sorting network as claimed in claim 2 wherein each said inhibit output means included two inhibit output connections, and means for providing different disabling signals of said two inhibit output connections.

4. A matrix sorting network as claimed in claim 1 wherein said array of cells of at least one said matrix define a main diagonal bisecting the array, said disabling means and said inhibit output means in each such matrix being connected to one another so that said disabling means of said cells in such matrix closer to the main diagonal are connected to the inhibit output means of cells farther from said main diagonal.

5. A matrix sorting network as claimed in claim 4 wherein said comparators of said array of cells of each said matrix are arranged such that said comparators on one said of said main diagonal select the greater of the values carried on said cell input means, and said comparators on the other side of said main diagonal select the lesser of the values carried on said cell input means.

6. A matrix sorting network as claimed in claim 5 wherein each said cell having inhibit output means further comprises one row inhibit output means and one column inhibit output means, such that said row inhibit output means are connected to said disabling means of the adjacent cell close to said main diagonal and in the same row as said cell having said inhibit output means, and said column inhibit output means are connected to said disabling means of the adjacent cell close to said main diagonal and in the same column as said cell having said inhibit output means.

7. A matrix sorting network as claimed in claim 1 wherein each said output bus within one said stage of said matrix sorting network is connected to one said input bus of one said matrix within the next said stage, for all said stages of said matrix sorting network except for the last said stage.