WO1988006764A2

WO1988006764A2 - Massively parallel array processing system

Info

Publication number: WO1988006764A2
Application number: PCT/US1988/000456
Authority: WO
Inventors: Robert S. Grondalski
Original assignee: Digital Equipment Corporation
Priority date: 1987-02-24
Filing date: 1988-02-16
Publication date: 1988-09-07
Also published as: CA1312960C; EP0303696A1; WO1988006764A3

Abstract

An array processing system including a plurality of processing elements each including a processor (22A) and an associated memory module (23A), the system further including a router network (30) over which each processing element can transfer messages to other random processing elements, a mechanism by which a processor can transmit data to one of four nearest-neighbor processors. In addition, the processing elements are divided into groups each with four processing elements, in which one of the processing elements can access data in the other processing elements' memory modules. The routing network switches messages in a plurality of switching stages, with each stage connecting to the next stage through communications paths that are divided into groups, each group, in turn being associated with selected address signals. A communications path continuity test circuit associated with each path detects any discontinuity in the communications path and disables the path. Thus, the stage may attempt to transfer a message over another path associated with the same address.

Description

.MASSIVELY PARALLEL ARRAY PROCESSING SYSTEM

BACKGROUND OF THE INVENTION

1. Field of the Invention

The invention relates generally to the field of digital data processing systems, and more particularly to array processing systems which incorporate a large number of processors that are interconnected in a regular connection structure and in which all of the processors receive the same instruction from a common control structure.

2. Description of the Prior Art

A digital data processing system comprises three basic elements, namely, a memory element, an input/output element and a processor element. The memory element stores information in addressable storage locations. This information includes data and instructions for processing 88/00456

-2- the data. The processor element fetches information from the memory element, interprets the information as either an instruction or data, processes the data in accordance with the instructions, and returns the processed data to the memory element. The input/output element, under control of the processor element, also communicates with the memory element to transfer information, including instructions and data to be processed, to the memory, and to obtain processed data from the memory.

Most modern data processing systems are considered "von Neumann" machines, since they are constructed according to a paradigm attributed to John von Neumann. Von Neumann machines are characterized by having a processing element, a global memory which stores all information in the system, and a program counter which identifies the location in the global memory of the instruction being executed. The processing element is executing one instruction at a time, that is, the instruction that is identified by the program counter. When that instruction has been executed, the program counter is advanced to identify the location of the next instruction to be executed. (In most modern systems. the program counter is actually advanced before the processor has finished processing the current instruction) .

Von Neumann systems are conceptually uncomplicated to design and program, since they do only one operation at a time, but they are also relatively slow. A number of advancements have been made to the original von Neumann paradigm to permit the various parts of the system, most particularly the various components of the processor, to operate relatively independently and achieve a significant increase in processing speed. The first such advancement was pipelining of the various steps in executing an instruction, including instruction fetch, operation code decode (a typical instruction includes an operation code which identifies the operation to be performed and in most cases one or more operand specifiers which identify the operands, or data, to be used in executing the instruction) , operand fetch, execution (that is, performing the operation set forth in the operation code on the fetched operands) , and storing of processed data, which are performed relatively independently by separate hardware in the processor. In a pipelined processor, the processor's instruction fetch hardware may be fetching one instruction while other hardware is decoding the operation code of another, fetching the operands of another, executing yet another instruction and storing the processed data of a fifth instruction. Pipelining does not speed up processing of an individual instruction, but since the processor begins processing a second instruction before it has finished processing the first, it does speed up processing a series of instructions.

Pipelining has also been used within several of the circuits comprising the processor, most notably the circuits which perform certain arithmetic operations, to speed processing of a series of calculations. Like pipelining of instruction processing, pipelining arithmetic operations does not speed up an individual calculation, but it does speed up processing of a series of calculations.

A pipelined processor is obviously much more complicated than a simple processor in a von Neumann system, as it requires not only the various circuits to perform each of the operations (in a simple von Neumann processor, many circuits could be used to perform several operations) , but also control circuits to coordinate the activities of the various circuits. However, the speed-up of the system can be dramatic.

More recently, some processors have been provided with execution hardware which include multiple functional- units each being designed to perform a certain type of mathematical operation. For example, some processors have separate functional units for performing integer arithmetic and floating point arithmetic, since floating point arithmetic requires handling two parts of a floating point number, namely the fraction and the exponent, while numbers in integer arithmetic have only one part. Some processors, for example the CDC 6600 manufactured by Control Data Corporation, included a number of separate hardware functional units each of which performs one or only several types of mathematical operations, including addition, multiplication, division, branch, and logical operations, all of which may be executing at once. This can be helpful in speeding up certain calculations, most particularly those in which several functional units may be used at one time for performing part of the calculation.

In a processor which incorporates pipelining or multiple functional units (or both, since both may be incorporated into a processor) , a single instruction stream operates on a single data stream. That is, each instruction operates on data to produce one calculation at a time. Such processors have been termed "SISD", for "single instruction- single data". However, if a program requires a segment of a program to be used to operate on a number of diverse elements of data to produce a number of calculations, the program causes the processor to loop through that segment for each calculation. In some cases, in which the program segment is short or there are only a few data elements, the time required to perform the calculations on the data is not unduly long.

However, for many types of such programs, SISD processors would require a very long time to perform all of the calculations that are required. Accordingly, processors have been developed which incorporate a large number of processing elements, all operating concurrently on the same instruction, with each processing element processing a separate data stream. These processors have been termed "SIMD" processors, for "single instruction-multiple data".

SIMD processors are useful in a number of applications, including image processing, signal processing, artificial « intelligence, database operations and computer simulation of a number of things such as electronic circuits and fluid dynamics. In image processing, each processor performs processing on a pixel ("picture element") to enhance the overall image. In signal processing, the processors concurrently perform a number of the calculations required to produce the Fast Fourier Transform of the signal. In^¬ artificial intelligence, the processors perform searches on extensive databases representing the stored knowledge of the application. In database operations, the processors perform searches, as in the artificial intelligence applications, and they also perform sorting operations. In computer simulation of, for example, electronic circuits, each processor represents one part of the circuit, and the processor's calculations indicates the response of the part to signals from other parts of the circuit. Similarly, in simulating fluid dynamics, which can be useful in a number of applications such as weather prediction and the design of airplanes, each processor is associated with one point in space, and the calculations performed provide information about various factors such as fluid flow, temperature, pressure, and so forth. Typical SIMD processors include two primary components, namely an array of processor elements and a routing network over which the processor elements may communicate the results of a calculation to other processor elements for use in future calculations. In addition, SIMD processors include a control processor for controlling the operations of the processor elements and routing network in response to instructions and data from a host computer system.

Several routing networks have been used in SIMD processors and a number of others have been proposed. In one routing network, which has been used in the Massively Parallel Processor, manufactured by Goodyear Arrowspace Corporation ("Goodyear MPP") , the processor elements are interconnected in a matrix, or mesh,,^{^} arrangement. In such an arrangement, the processor elements are connected in rows and columns and directly communicate only with their four nearest neighbors. This arrangement can be somewhat slow if communications may be to random processor elements, but the number of wires which are required to make the interconnections is lower than in most other arrangements, on the order of 4n, where "n" is the number of processor elements, assuming only unidirectional transfer of messages over each wire. If each wire can transfer bidirectionally, the number of wires is reduced by half, with a possible reduction in the message transfer rate.

The matrix network is also used on the "Connection Machine", manufactured by Thinking Machines Corporation, but that machine also includes a hypercube network allowing " communications between random processor elements (that is, processor elements which are not nearest neighbors) . In a hypercube network, each processor chip connects directly to twelve other processor chip. Each processor chip includes several processor elements and circuits which form part of the routing network. The routing circuits on each chip receive messages from the processor elements on the chip for transmission to processor elements on other processor chips. In addition, the routing circuits receive messages from other processor chips. If a message from another processor chip is to be received by a processor element on the chip, it forwards it to that element; however, if the message is to be received by a processor element on another chip, it transmits the message over a wire to another processor another chip. The procedure is repeated until the message reaches the intended recipient. Thus, the routing circuits on each chip must be able to handle not only messages from the processor elements on the chip, but also from messages from other chips which may or may not be addressed to processor elements on the chip.

A hypercube network handles communications fairly rapidly, but it requires a large number of wires, on the order of nlog₂n, if messages are transferred unidirectionally over each wire. For example, if "n" were 4096 (4K, K=1024) , the hypercube would require on the order of 48K wires. If the wires transfer messages bidirectionally, only 24K wires would be required, but the volume of message traffic that could be carried would also be reduced. Typically, the larger the number of wires in a routing network, the more expensive is the network, and the greater is the likelihood of failure resulting from broken wires or failed switching elements which interconnect the wires.

Another routing arrangement which has been proposed is a crossbar switch, through which each processor element can communicate with any of the other processor elements directly. The crossbar switch provides the most efficient communications of any of the routing networks proposed.. However, a crossbar switch also has the most wires and switching elements, both on the order of n2, and thus is most expensive and most susceptible to failure due to broke wires and switching elements. Using the above example, in which "n" is 4K, the number of wires and switching elements required for the crossbar switch is 16M (M=l,048,576) .

Yet another routing arrangement is an omega network, i which switching is performed through a number of serially- connected stages. Each stage has two inputs, each connecting to the outputs of two prior stages or processor chips and has two outputs. The "Butterfly" manufactured b Bolt Beranek and Newman, uses an omega network.

The cost of a routing network is directly related to the number of wires, as is the likelihood of failure due to discontinuity in a communications path. To reduce the number of wires and achieve a significant fraction of the efficiency of the crossbar switch, a routing network has been proposed in which a multiple-stage omega network performs some portion of the switching. The output from th omega network is connected to a crossbar switch, which woul require many fewer switching connections than would be necessary in the absence of the omega network. Depending o the number of stages in the omega network, the number of wires may be less than in a hypercube, while the transfer efficiency would be greater than a hypercube. For example, if "n" is 4K, and a seven-stage omega network is provided to a crossbar switch, 36K wires (again assuming unidirectional communications over each wire) would be required to form the routing network.

Using a routing network to transfer data does have a number of limitations. The mesh network is useful generally when transferring data only to the adjacent processors, as each transfer requires commands from the controlling program. A hypercube, crossbar switch, omega, or like network is most useful if message transfers are expected to b^"e to random processors. Some array processors, the connection machine, for example, have two mechanisms for transferring data, one for random transfers and the other for matrix transfers. Under some circumstances, however, it may be faster to provide a processor with direct access to memories associated with other processing elements. This may be useful, for example, when, after performing operations in parallel, a serial operation is to.be performed on the just processed data. If one processing element has access to the data in at least some other processing elements' memories, the processor may perform serial operations using that data. Also, the processing element may use those memories if a problem requires more storage capacity than a single processing element would have.

Typically, array processors are used in performing arithmetic operations on numerical values, which are expressed in "floating point" form. In that form, a floating point number has a fraction portion and an exponent portion, with the value of the number being the value contained in the fraction portion multiplied by the value two raised to the value contained in the exponent portion. When performing arithmetic operations such as addition and subtraction on such numbers, the numbers must be "aligned", that is, they must have the same value of the exponent. To achieve this, the value of the fraction portion of the floating point number must be reduced, which raises the effective value of the number's exponent portion, until the exponent values are equal. After the arithmetic operation, the fraction of the result must be normalized, that is, leading zeroes must be removed by decreasing the value of the fraction portion, while at the same time increasing the value of the result's exponent. In both the alignment and normalization operations, the fractions are reduced and or increased by shifting their values in the locations in which they are stored.

However, in the alignment and normalization operations, since the values of the numbers processed by the various processing elements are all different, the number of shifts required to effect the alignment or normalization will also be different.

SUMMARY OF THE INVENTION

The invention provides a new and improved array processing system having a plurality of processing elements each with a processor and an associated memory.

In one aspect, the invention provides an array processing system including a plurality of processing elements each including a processor and an associated memory, in which the processing elements are divided -into groups each having a selected number of processors and associated memories. In response to control signals, one processor in each group is enabled to obtain data from any of the memories associated with any of the processors in the group. While operating in this mode, the other processors are effectively disabled. This aspect allows computations to be performed in parallel on data in the various memories by all of the processors in the group, and then facilitates the serial processing of that data by one of the processors in the group, without requiring the other processors to transfer the data to the one processor over the interconnection mechanism. In addition, it permits the array to provide larger amounts of memory, using smaller numbers of processors, if necessary to perform a computation.

In another aspect, the invention provides an array processor whose processing elements can process data a bit at a time, in bit mode, or a nibble (four bits) at a time, in nibble mode.

In another aspect, the invention provides a processor having a shift register that can shift both toward a shift out terminal and toward a shift in terminal, and that can shift toward the shift out terminal a bit at a time or a nibble (four bits) at a time. The ability to shift a^nibble at a time facilitates the operation in nibble mode. The ability to shift toward the shift in terminal is useful in floating point arithmetic, particularly shifting toward the shift in terminal in normalizing the fraction portion of a number after an arithmetic operation.

In yet another aspect, the invention provides a mechanism for quickly and easily generating a status signal for indicating when all of the processors in the array have completed an operation such as normalization following arithmetic operations or transmission of messages.

In still another aspect, the invention provides an array processing system in which the processing elements can receive data from, and transmit data to, other processing elements in a matrix arrangement, and in which they can provide that the data transmitted is the logical OR of the received data and the processing element's own data, to effect a logical OR of the data being transmitted by all of the processing elements along an entire row or column.

In a further aspect, the invention provides an array procssing system having a message transfer mechanism that permits one processing element, which might otherwise have to serially transmit messages to a large number of other -17- processing elements, to use other processing elements as assistants in transmitting messages to the other processin elements, thereby reducing the amount of time required to transmit all of the necessary messages.

In a further aspect, the invention provides an array processing system including a routing network for transferring messages, with each message containing an address identifying the intended recipient. The routing network includes a plurality of switching stages each of which uses a selected number of bits of the address to identify a switching path through each stage. The lines connecting between stages are divided into groups, with ea group including a plurality of lines each associated with one encoding of the address bits used by the stage in selecting the switching path. Each switching stage can couple a message having the required address encoding onto any one of the lines in the appropriate group.

In another aspect, the invention provides an array processing system including a routing network in which messages are transferred over communications path, with th routing network including a system for detecting when a communications path has become discontinuous, and for thereafter not using that path.

In yet a further aspect, the invention provides an array processing system including routing network for transferring messages over a communications path. Each message contains an address identifying the intended recipient. The routing network includes a plurality of switching stages each of which uses a selected number of bits of the address to identify a communications path through each stage. The communications paths connecting between stages are divided into groups, with each group including a plurality of paths each associated with one encoding of the address bits used by the stage in selecting the switching path. Each switching stage can couple a message having the required address encoding onto any one of the paths in the appropriate group for transmission to the next stage. If one communications path in a group is not used because it is discontinuous, the stage may transmit a message over any of the other communications paths in the group to the next stage.

In yet another aspect, the invention provides an array processing system having a multiple-stage routing network, in which the first stage is enabled to begin transferring a message, and in which each stage enables the next stage when it is time for that stage to begin transferring the message.

BRIEF DESCRIPTION OF THE DRAWINGS

This invention is pointed out with particularity in the appended claims. The above and further advantages of this invention may be better understood by referring to the following description taken in conjunction with the accompanying drawings, in which:

Fig. 1 is a general organizational diagram depicting the general components of an array processing system, including a processor array constructed in accordance with the invention;

Fig. 2 is a more detailed block diagram of the processor array depicted in Fig. 1, depicting one mechanism for transferring information through the array;

Fig. 3 is another more detailed block diagram of the processor array depicted in Fig. 1, depicting another mechanism for transferring information through the array; Fig. 4 is a detailed block diagram of a processor element forming part of the processor array depicted in Fig. 1;

Fig. 5A is a block diagram depicting the interconnections between several processor elements and memory elements in the processor array depicted in Fig. 1, and Figs. 5B and 5C depict schematic circuit diagrams detailing the interconnections;

Fig. 6A is a block diagram depicting generation of various status signals on each processor chip used by the array control unit depicted in Fig. 1. ;

Fig. 6B is a detailed circuit diagram of a local address register which forms part of processor element chips in the processor array depicted in Fig. 2;

Fig. 7 is a detailed circuit diagram of a shift register which forms part of the processor element depicted in Fig. 4;

Fig. 8A is a detailed block diagram depicting the routing network used in the information transfer mechanism depicted in Fig. 2 and Fig. 8B depicts the detailed structure of a message transferred through the routing network depicted in Fig. 8A;

Fig. 9 is a detailed block diagram depicting the organization of a switching chip used in the routing networ depicted in Fig. 8A; and

Fig. 10, comprising Figs. 10A through IOC, detailed.. circuit diagrams of various circuits comprising the routing chip depicted in Fig. 9.

DESCRIPTION OF AN ILLUSTRATIVE EMBODIMENT

I. General Description Of Array Processor

With reference to Fig. 1, an array processor constructed in accordance with the invention includes three major sections, including an array control unit 10, a processor array 11 and a system interface 12. The control unit 10 receives data and instructions from a host data processing system (not shown) through the interface 12, transmits the data to the array 11 over a bus 13 and issues to the array 11 microwords over a bus 14 and other timing and control signals s described below. The microwords enable the array 11 to process the data in accordance with the host's instructions. The microwords include fields defining various PROC OP processor operation signals to identify selected operations, MEM ADRS memory address to identify a memory address, and other operational control signals.

During execution of the microinstructions, the processor array 11 may transmit various status signals back to the array control unit over a status bus 15. These status signals are used by the array control unit 10, as described below, in the selection of the microinstructions to be issued to the processor array. After the processor array performs all of the operations responsive to the microinstructions from the array control unit in processing the instruction from the host data processing system, the control unit 10 then enables the processing array 11 to transmit the processed data to the array control unit 10 over bus 13, and the array control unit may then transmit it to the host through system interface 12.

II. General Organization of Processor Array 11

Processor array 11 includes two general portions, namely a large number of processor elements (PE's) 22 on a plurality of processor element (PE) chips 20, and a mechanism for transferring information to and from the array control unit 10 and among the processor elements. In one specific embodiment, each PE chip includes thirty-two processing elements. As described in more detail below in connection with Fig. 4, each processor element includes a processor, including circuitry for performing selected arithmetic and logical operations, and an associated memory. In the aforementioned specific embodiment, each memory has a capacity of 1024 bits (1 Kb) organized into 128 bytes.

The PE's have two mechanisms for transferring information to other processors. In one mechanism, a "random transfer" mechanism, which is generally depicted in Fig. 2, each PE can send a message to any other PE. In the second mechanism, a "nearest neighbor" mechanism, which is illustrated in Fig. 3, the PE's are interconnected in a "nearest neighbor" mesh manner with four other PE's, and each PE may send information only to one of those four PE's to which it is directly connected. In the random transfer mechanism, communications are maintained over a routing network 30 (see Fig. 2) , which is described below in connection with Fig. 2 and in more detail in connection with Figs. 8A through IOC. Circuits for implementing the second mechanis (that is, the nearest neighbor mechanism) are depicted in Fig.- 4.

ith reference to Fig. 2, a plurality of PE chips 20A through 20N (generally identified by reference numeral 20) are depicted. Each PE chip 20 has a plurality of PE's. Since the PE chips are similar, only chip 20A is shown in detail. PE chip 20A includes a plurality of PE's 21A through 21M in an ordered heirarchy. Each PE includes a processor 22A through 22M and a memory 23A through 23M. When a PE is to transmit a message, it's TR transmit request flag 24A through 24M is set. When its RTR OUT EN IN router out enable in signal is then asserted, the PE is enabled to transmit the message, which normally includes the identification of the intended recipient PE followed by data, through a router output control circuit. After the message transfer operation completed, the transmitting PE's TR transmit request flag is reset.

The RTR OUT EN IN router out enable in signal is a daisy-chained enable signal. When it becomes the PE's turn to send a message, which occurs when its RTR OUT EN IN router out enable in signal is asserted, it transmits its message signals'through a router output control circuit 25, which transmits the message from the PE chip 20A as a CHP 0 MSG OUT chip message out signal onto line 29(A).

After a PE's message operation has been completed, its TR transmit request flag 24 is reset and the PE 21 asserts a RTR OUT EN OUT router out enable out signal, which is daisy- chained to the next PE as its the RTR OUT EN IN router out enable in signal. If that PE's TR transmit request flag 24 is set, that PE is enabled to transmit a message. If the TR transmit request flag 24 is cleared the PE does not transmit a message, but instead asserts its RTR OUT EN OUT router out enable signal, which is daisy-chained as the RTR OUT EN IN router out enable in signal to the next PE 21. After messages have been sent and acknowledged for all of the PE's whose TR transmit request flags 24A through 24M were initially set, the control circuit 25 generates a LOCAL DONE signal, which is ORed with LOCAL DONE signals from other PE chips to form a DONE signal. The DONE signal is transmitted over status bus 15 to the array control unit 10, to indicate that all PE's have transmitted messages.

In one specific embodiment, to reduce the number of input terminals to the routing- network 30, the PE chips 20A through 20N are paired in their transmission of messages through the routing network. With reference to Fig. 2, the last PE 21(M) on PE chip 20A provides the RTR OUT EN IN router out enable in signal to the first PE 21 on PE chip 20B. Thus, after all messages have been transmitted from PE chip 20A and acknowledged, PE chip 2OB begins transmitting messages.

The message outputs of selected ones of the PE chips 20A through 2ON are connected to multiplexers 32 which selectively couple the outputs of the PE chips or the data signals from the data bus 13 to the routing network. The routing network 30 thus facilitates the transfer of data from the array control unit 10 into the processor array 11. More specifically, with reference to PE chip 20A, the output circuit 25 of PE chip 20A is connected to one input terminal of a multiplexer 32. The other input terminal of multiplexer 32 is connected to one line of data bus 13. The multiplexer is controlled by a SEL DB select data bus signal which, when asserted, enables the multiplexer 32 to couple the signals on the line of data bus 13 to the routing network 30. When the SEL DB select data bus signal is not asserted, the multiplexer 32 couples the signals from the PE chips to the routing network 30. The output signals from the routing network 30 (the routing network will be described below in connection with Figs. 8A through IOC) are coupled as CHP X MSG IN chip message input signals ("X" refers to the particular PE chi pair) on lines 31(A) through 31(N) to message input terminals on the PE chips 20A through 20N, and specificall to a router input control circuit 33 shown in PE chip 20A-. The router input control circuit first determines whether not the message is for a PE on the PE chip on which it resides. This determination is made in response to a portion of the address field of the received message, in particular first signal received by the router input contr circuit, and by the condition of a HI 32 signal. As shown in Fig. 2, the HI 32 signal is transmitted in uncomplement form to one of each pair of PE chips, such as PE chips 20A and 20(N-1), and in complemented form to the other PE chip in the pair, such as PE chips 20B and 20N. If the router input control circuit 33 receives the uncomplemented HI 32 signal it accepts the messages for which the first signal asserted, but if the router input control circuit 33 receives the complemented HI 32 signal, it accepts the messages for which the first signal is negated. The remainder of the address field, as explained below in connection with Fig. 8B, identify the PE 21 which is the intended recipient of the message, and the router input άdantώfiedrcuTnediπBαtscohermessageuib feafesPEhelRSS receive flag 26 of the receiving PE and directs the data portion of the message to that PE's memory 23.

In the transfer of a message from a particular transmitting PE 21 through the routing network 30 to a recipient PE 21, the address essentially establishes a transfer path. If a RVM reverse mode signal from the control icroword from array control unit 10 is asserted after the transfer path has been established, the transfer direction through routing network 30 is reversed. Accordingly, the PE 21 whose RCV flag 26 is set is enabled to transmit data onto line 31 and the routing network 30 couples the data to the appropriate line 29. The output control 25 then couples the received data to the memory 23 of the PE 21 which transmitted the original message, that is, the PE 21 associated with the first set TR flag 24 in the heirarchy._ If the RVM reverse mode signal is later negated, data may again be transferred over the established transfer path in the forward direction. Accordingly, by suitable conditioning of the RVM reverse mode signal, transfers may take place in either direction over the transfer path.

As described above, the control microword includes a memory address field which provides MEM ADRS memory address signals to the PE's 21. In addition, a local address register 28 is provided, which can be used by the PE's when a LAM local address mode signal from the control microword is asserted. The local address register can be loaded by a message from routing network 30, and so a message itself can contain the particular destination address in the identified PE's memory into which it is to be stored. After the local address register is loaded, it normally automatically increments to identify the next location in memory, which simplifies use during a message transfer. A PE 21 may also load the local address register 28 and use it, thereby allowing the PE to be indirectly addressed. The local address register is described below in more detail in Fig. 6B.

The output signals from selected ones of the lines from routing network 30 are also coupled to data bus 13 to carry data output signals to the array control unit 10. The "nearest neighbor" message transfer mechanism is depicted in Fig. 3. With reference to Fig. 3, a plurality of PE's 22A through 22N are depicted organized in an array pattern of a plurality of "k" rows and "1" columns, where N=kl. Each PE 22A through 22N can transmit information, in the form of signals, to its four nearest neighbors in the array. Specifically, for example, the PE 22(k+2) can transmit signals "westwardly" to PE 20B and "northwardly" to PE 22k+l (shown on Fig. 3), and "southwardly" to PE 22(k-K3) and "eastwardly" to 22(2k+2) in a row and column not shown in Fig. 3. The PE's in the leftmost column, as shown in Fig. 3, that is PE's 22A through 22K, can transmit signals westwardly to the rightmost column as shown in Fig. 3, that is, to PE's 22(lk+l) through 22N, and the PE's in the rightmost column can transmit signals eastwardly to the PE's in the leftmost column. Similarly, the PE's in the top row, that is, PE's 22A, 22(k+1) ...22(lk+1) , can transmit signals northwardly to the PE's in the bottom row, that is PE's 22K, 22(2k) ...22N, and the PE's in the bottom row can transmit signals southwardly to the PE's in the top row. The mechanism used by the PE's to transmit in the "nearest neighbor" mode will be described ^"below in connection with Fig. 4. -31- The nearest neighbor message transfer mechanism may also be used for input of signals, representing data, to array 11 or output of data from array 11. In particular, with reference to Fig. 4, a multiplexer network (not show may be connected at the left side of the array, for exampl to permit data to be loaded into the leftmost column of PE data from either the rightmost column or from an external data source. The data is initially latched by the leftmos column and is transmitted to the next column, that is PE's 22(K+1) through 22(2K) as more data is latched by the leftmost column. That process continues until data has been passed throughout the entire array.

The output of data is effectuated by means of the nearest neighbor transfer mechanism by providing a set of output drivers (not shown) which are connected to a set of appropriated receiving and recording circuits. The output drivers may be connected to, for example, the eastward transmission terminals of the rightmost column of the arra to receive data signals from the lines which also connect the PE's along that column to the PE's on the leftmost column. To allow data to be retrieved from the array 11, the PE's are enabled to transmit the data in an easterly, direction while recording the outputs of the output drivers. After the data from the leftmost column has returned to it, data from all of the PE's in the array will have passed through the rightmost column of PE's and have been transmitted by the drivers.

III. Processing Element

A. General Description (Discussion of Fig. 4)

With reference to Fig. 4 , a PE constructed in accordance with the invention includes a full adder 51 which receives input signals from a one of a pair of shift registers, namely A shift register 52 or B shift register 53, from a Boolean logic circuit 54, and from a carry latch 55. A multiplexer 56 selects the output signals from the S/0 shift out output terminals of one of the A or B shift registers 52 or 53, in response to an A/B select signal representing a bit in the control microword received from the array control unit 10 over bus 14 (Fig. 1) . As is typical in an array processing system, during any given operating cycle, as governed by one or more clock signals (not shown) , a single control microword is transmitted by the array control unit 10 to ?11 of the PE's in the array 11 -33- to control the operations of the PE's in the array 11 dur that cycle. Depending on the condition of one or more flags, as described below, a PE may or may not perform th operations enabled by the control microword.

If a CLO clear shift register output signal from the control microword is negated, the multiplexer is enabled couple the output signal from the selected shift register the connected input of the adder 51. By asserting the CL signal, the array control unit enables the multiplexer to couple an output signal having an effective value of zero the input of adder 51. This provides a mechanism for loading a value into a shift register prior to performing arithmetic operation.

The A and B shift registers 52 and 53 operate in response to a number of control signals from the control microword. Both the A and B shift registers are of variab effective length, that is, the output bit position from which output signals are transmitted to multiplexer 56 can be varied in response to A SHFT SZ (A shift register size) and B SHFT SZ (B shift register size) signals from the control microword. In one embodiment, the A shift registe 52 has effective lengths of four, eight, twelve, sixteen. twenty, twenty-four, twenty-eight and thirty-two bits, and may be used for performing integer arithmetic and arithmetic on the fraction portion of floating point numbers. In the same embodiment, the B shift register 53 has effective lengths of four, eight and twelve bits, and may be used for performing arithmetic on the exponent portion of floating point numbers.

The contents of the A and B shift registers 52 and 53 are shifted in response to a SHFT EN shift enable signal generated by an OR gate 60. The OR gate 60 is energized by an SH shift signal from the control microword, and by an asserted signal from an AND gate 61. The AND gate 61 is enabled by an M signal from an M flag 62 and is energized by an MSH (M shift) signal from the control microword. Thus, the array control unit may enable an unconditional shift by asserting the SH shift signal, or it may alternatively enable a shift conditioned on the state of the M flag 62 by asserting the MSH signal in the control microword.

The PE's operate in either a bit serial mode or a nibble mode, in which a nibble (four bits) is processed during a control cycle. The PE operates in nibble mode when an NM nibble mode signal from the control microword is -35- asserted. When the NM nibble mode signal is asserted, the and B shift registers 52 and 53 shift four bits in a cycle, otherwise they shift one bit in a cycle. Also, when in nibble mode, the A and B shift registers receive and store four bits in a cycle from their shift in input terminals (designated S/I on the Figure) .

Finally, as described in detail below in connection with Fig. 7, the A shift register 52 also can shift in both directions, that is, toward the S/O output terminal or toward the S/I shift in input terminal. The shift directio is selected in response to a SHFT DIR shift direction signa from the control microword. The ability to shift the contents of the shift register toward the S/I input termina is useful in floating point arithmetic, for example, particularly in normalization, that is, in removing leading zeroes, which are stored in the end of the shift register toward the S/I shift in input terminal.

The Boolean logic circuit 54 couples data signals to the full adder 51 from a four-line data bus 58 or from one of the nearest neighbor PE's. Depending on the state of LG FTN logic function signals from the control microword, and if an LCM load constant mode signal from the control microword is not asserted, the Boolean logic circuit 54 may perform one of sixteen logical operations between an X flag 63 and signals from the data bus 58 or from the nearest neighbor PE's. If a logical operation is performed between the X flag 63 and the signals on the data bus 58, the logical operation is performed with respect to the signal on each line of the data bus 58 individually. For example, if an AND operation is to be performed, four AND operations are actually performed, one between the X flag 63 and the signal on each of the lines of the data bus 58.

If the LCM load constant mode signal is asserted, the Boolean logic circuit 54 couples the LGC FTN logic function signals from the control microword to the address 51 as LOG OUT (3:0) logic out signals. In this way, the control microword may provide a constant, corresponding to the LGC FTN logic function signals, to all of the PE's in parallel.

In either bit-serial or nibble mode, the carry latch 55 supplies a one bit CARRY signal to the full adder 51. The carry latch 55 latches the CARRY signal whenever the SHFT EN shift enable signal from OR gate 60 is asserted, A SET C set carry signal from the control microword, forces the CARRY IN output signal of the carry latch 55 to an asserted level, or a logical one value, and a CL C clear carry sign from the control microword, when asserted, forces the CARR IN output signal to a negated level, or a logical zero value. The SET C set carry signal and CL C clear carry signal are useful for the same purpose as the CLO clear ^■ shift register output signal, namely, initial conditioning of the carry latch 55 prior to performing an arithmetic - operation.

Furthermore, in either bit serial or nibble mode, the full adder 51 generates four ADD OUT (3:0) adder output signals and a C OUT carry out signal. The ADD OUT (3:0) adder output signals are coupled to the S/I shift in terminals of the A and B shift registers 52 and 53 and to input terminal of an adder latch 57, The C OUT carry out signal represents the carry signal from the high-order add stage of the full adder 51, that is, .the stage that generates the high-order ADD OUT (3) adder output signal.

The C OUT signal and the C OUT (0) carry signal from the low order stage of the adder, which represents the car from that stage, are coupled to a multiplexer 64 which, in response to the NM nibble mode signal from the control microword, couples one of the signals to an input of a second multiplexer 68. Multiplexer 68, in turn, in response to a LD C load carry signal from the control microword, selects either the CARRY OUT signal from multiplexer 64 or the LOG OUT (0) logic output (0) signal from Boolean logic circuit 54 for storage in the carry latch 55. If the LD C load carry signal is asserted, the multiplexer 68 is conditioned to couple the LOG OUT (0) logic output (0) signal to the carry latch 55 and the carry latch is enabled to latch the signal. This facilitates initializing the carry latch 55 to a known state through Boolean logic 54. If, on the other hand, the LD C load carry signal is negated, the multiplexer 68 is conditioned to couple the - CARRY OUT signal from multiplexer 64 to the input of the carry latch. The CARRY OUT signal is latched if the SHFT EN shift enable signal from the control microword is asserted.

Thus, in nibble mode, the carry latch 55 receives the carry from the high-order stage of the full adder 51, representing the carry from the high order stage of the nibble. In bit-serial mode, the carry latch 55 receives the carry from the low-order stage of the adder. In bit-serial mode, the arithmetic operations in full adder 51 are essentially carried out in the low-order stage, as the signals from the higher-order stages are effectively ignored.

The M flag 62 is also conditioned by signals from the full adder 51 through a multiplexer 65. The multiplexer couples either the ADD OUT (0) signal from the full adder 51, or, if the NM nibble mode signal is asserted, the ADD- OUT (3) signal. The signal from the multiplexer is inve^~rted by an inverter 66 and the result coupled to an input terminal of a second multiplexer 67. The multiplexer 67 also includes a second input terminal which receives the low-order signal from the Boolean logic circuit 54. If the SHFT DIR shift direction signal signal from the control microword is asserted, indicating that data in the shift register 52 is shifted toward the shift out terminal S/O, the multiplexer 67 couples the complement of the high-order signal from the A shifter 52 (left shift) to the M flag 62 for storage in response to an LD M load M signal from the control microword. Otherwise, the M flag 62 is loaded from the signals from the Boolean logic circuit 54. An AND gate 70 couples the condition of the M flag 62 to data bus 58, particularly the low-order DATA (0) line thereof, in response to a RD M,read M signal from the control microword. If the M flag 62 is loaded from the A shifter 52, it may be used, for example, in normalization during floating point arithmetic. In this operation, the fraction portion of the floating point number, which is stored in the A shift register 52, is shifted toward the S/I shift in terminal (that is, the high-order end) until the high-order stage contains the value "one". In response to each shift of the contents of the A shift register 52, the exponent, which is in the B shift register 53, is incremented. Since different PE's in array 11 may require different numbers of shifts to normalize the contents of their respective A shift registers 52, the shift operations then depend on the state of the M SH (M conditioned shift) signal from the control microword. Until a "one" reaches the high-order stage of the A shift register 52, the inverter 66 conditions the M flag 62 to be set (that is, store the complement of the zero value from the shift register 52) , and thus the M signal is asserted. Thus, the M SH signal enables the A and B shift registers 52 and 53 to operate. However, when a "one" reaches the high-order stage of the A shift register 52, the inverter 66 conditions the M flag to be clear, negating the output M signal. When the M SH (M conditioned shift) signal from the control microword is asserted, the negated M signal disables the shift registers 52 and 53. The M flag 62 can also be used for initial alignment of floating point numbe prior to an addition or subtraction in a similar manner.

Two additional flags are loaded from the low-order output signal from Boolean logic circuit 54, namely a WRT memory write enable flag 71 and the TR transmit request fl 24. The WRT memory write enable flag 71 is loaded in response to an LD WRT load memory write enable signal from the control microword and the TR transmit request flag 24 loaded in response to an LD TR load transmit request signal from the control microword. The WRT memory write enable flag 71 generates a WRT memory write signal which is coupl to the PE chip's router control circuit 33 which, in turn, generates a WRT EN write enable signal associated with eac of the memories 23 on the chip to control writing of data thereto.

The router control circuit 33 enables writing data to the memory under four circumstances. Under normal operation, that is, when the routing network 30 (Fig. 2) is not in use, the router control circuit 33 generates a PE SE processor element select signal when the WR write signal from the control microword is asserted if the associated PE's WRT flag is set. Since the RVM reverse mode signal from the control microword is negated, a multiplexer 500 is enabled to assert the WRT EN write enable signal, which is coupled to the memory 23. The WRT write enable signal enables the associated memory 23 to store the data bits from the data bus 58 through transceiver 72 in the location specified by the address signals in the control microword or from the local address register 28 (see Fig. 6B, below) .

Alternatively, when the routing network is in use and actually transferring data, storage of data in the memory depends on the condition of the RVM reverse mode signal. The router control circuit 33 receives the data as an MSG IN message in signal. Contemporaneously, the ENR enable router signal is asserted. As will be explained below in connection with Fig. 10A, the first six bits received by the PE chip 20 identify the chip of the chip pair and PE on the chip which is to receive the message data. The router control circuit 33 receives the PE identification portion of the MSG IN message in signals and asserts a SET R set receive flag signal which is associated with the identifed PE 21. This enables the RCV receive flag 26 to be set, which results in the assertion of an RCV receive signal. In addition, the router control circuit 33 asserts the PE SEL processor element select signal associated with the identified PE. A multiplexer 501 is enabled to couple the data portion of the MSG IN message in signals on line 31 as a DATA (0) signal onto the data bus 58. In addition, if th RVM reverse mode signal is not asserted, the multiplexer ,50 asserts the WRT EN write enable signal associated with the intended recipient PE to enable the DATA (0) signal to be written into memory 23. This enables the data from the dat field of a message during a forward transfer through routin network 30 to be written into the memory 23 of the receivin PE 21.

If, on the other hand, the RVM reverse mode signal fro the control microword is asserted, the multiplexer 501 is enabled to couple the signal on line 29, which normally carries the MSG OUT message out signal, to the data bus 58 as the DATA (0) signal. In addition, the multiplexer 500 couples a TR EN transmit enable signal to the memory 23 as the WRT EN write enable signal. As explained below, the TR EN transmit enable signal is asserted when the PE 21 is enabled to transmit during a forward transfer through routing network 30, and so identifies the recipient of data during a reverse transfer.

Finally, the router control circuit 33 also enables the memory 23 to store data when the routing network is in use but between messages. This may occur when, for example, several messages are to be directed to a single PE by several PE's during one major message transfer cycle, and between messages the recipient PE is to perform a calculation on the just received data. A major message transfer cycle comprises a plurality of minor message transfer cycles each initiated by the ENR enable router signal and terminated by a BND branch if not done signal. If the DONE signal is not asserted at the end of a minor message transfer cycle, some PE's have messages yet to transfer, and the BND signal from the control microword indicates that the array control unit will return to the beginning of the message transfer sequence. If a calculation is to be performed before the beginning of the next minor message transfer cycle, the ENR enable router signal is negated, but the BND branch if not done signal is delayed until the calculation is performed and the results written into memory 23. In this case, the memory 23 is enabled by the router control circuit 33 if the WRT flag set and the WR write signal from the control microword is asserted. However, the WRT EN write enable signal will b asserted only for those PE's which just received a messag during the current minor message transfer cycle because t condition of the RCV flag 26 is normally coupled onto the data bus and is stored in WRT flag 71.

The data output from memory 23 is coupled to a memor data output control circuit 73 described below in connect with Figs. 5A and 5B. In brief, the memory data output control circuit 73 enables four memory modules 23, each normally associated with a separate PE, to be associated with one of the PE's in response to the conditions of the ADRS (11:10) high order address signals from the control microword. Thus, in one embodiment, in which the memory modules 23 each have on the order of 1 Kb (1024 bits) , wh either of the ADRS (11) or ADRS (10) high order address b are asserted, one PE has an effective memory of 4 Kb (4096 bits), and the other PE's, which are normally associated with the PE's whose memory modules 23 are available to th one PE, are effectively disabled. This may be useful in several situations, including cases requiring more storag than is normally available to one PE. In addition, this may be useful in cases in which operations are performed in each PE on data in parallel,- and then operations are performed serially by one PE using the data stored in all of the memory modules. Making the data in the memory modules directly available to the PE performing the operations serially avoids requiring the PE's to use message transfers to transfer the data to that PE.

The data from the memory data output control circuit 73 is stored in a read data latch 74 and coupled by the latch to transceiver 72. The transceiver 72 in turn couples the data onto data bus 58. The data bus 58, in turn, couples the data to one input terminal of a multiplexer 75. If an NR neighbor signal from the control microword is not asserted, the multiplexer 75 couples the read data from the data bus 58 onto a bus 77 to the input terminal of Boolean logic circuit 54.

A multiplexer 76 selects a data signal from one of the neighboring PE's, in response to the condition of D(1:0) direction signals from the control microword. If the NR neighbor signal from the control microword is asserted, multiplexer 75 couples the signal selected by multiplexer 76 onto the low-order line of bus 77, and transmits negated signals onto the remaining lines of the bus, thereby facilitating the data reception portion of the nearest- neighbor message transfer mechanism described above in connection with Fig. 3.

Data is coupled to the memory modules 23 for storage as follows. Data from the data bus 58 is coupled through transceiver 72 to the data input terminals DI of the associated memory module 23. Depending on the condition of the ADRS (11:10) signals, the memory data output control circuit 73 may instead couple the data from the transceiver 72 to a memory module 23 normally associated with another PE for storage therein.

The data transmission portion of the nearest-neighbor message transfer mechanism, which was described above in connection with Fig. 3, is performed as follows. The DATA (0) signal from data bus 58 is connected to one terminal of a multiplexer 80. If neither the ROW signal nor the COL column signal from the control microword is asserted, the multiplexer 80 couples the DATA (0) signal to the input terminal of a demultiplexer 81. If the NR neighbor mode signal from the control microword is asserted, the demultiplexer 81 couples the DATA (0) signal to the PE determined by the D(1:0) direction signals.

If, on the other hand, either the ROW or the COL column signal from the control microword is asserted, the multiplexer 80 couples the output signal from an OR gate 82 to the demultiplexer 81. The OR gate 82 is enabled by the DATA (0) signal from data bus 58 or the output of multiplexer 76. The signal transferred to the neighboring PE thus represents the logical OR of the signal from a neighboring PE received during the nearest neighbor transfer and the data signal on the low-order line of data bus 58. By this mechanism, the row and column data comparisons may be easily facilitated, which is useful in connection with finding minimum and maximum values along a row or column of PE's. If both the ROW and COL signals are asserted at once, the comparisons may be facilitated with respect to all of the PE's in the array 11 at one time.

Also depicted on Fig. 4 is circuitry associated with the PE chip's router control circuit 33 for enabling the PE to transfer messages over the routing network. As was note above in connection with Fig. 2, the PE's on a chip are iteratively enabled to transfer messages over the routing network. Each PE receives an RTR OUT EN IN router out enable in signal which, when asserted, enables the PE to transmit message data signals over a common line 29. After the message is acknowledged, the PE generates an RTR OUT EN OUT router out enable out signal which is coupled to the next PE in line as that PE's RTR OUT EN IN router out enable in signal.

Specifically, when the TR transmit request flag 24 is set, a TR transmit request signal is asserted, which enables one input of an AND gate 94. When the PE's RTR OUT EN IN router out enable in signal is next asserted, an AND gate 94 asserts the TR EN transmit enable signal, which is coupled to one input of a multiplexer 93. If the RVM reverse mode signal is not asserted, indicating the transfer is in the forward direction through the routing network 30, the TR EN transmit enable signal enables a driver 95 to couple RD LTH (0) read latch output signals onto a line 90. If the RVM reverse mode signal is not asserted, an inverter 97 enables a driver 98 to couple the signal on line 90 onto line 29 as the MSG OUT message out signals.

On the other hand, if the RVM reverse mode signal from ' the control microword is asserted, indicating the transfer is in the reverse direction, the multiplexer couples the RCV receive signal from the receive flag 26 to control the driver 95. When the RCV signal is asserted, which occurs when the PE 21 the intended recipient of a message, the driver 95 is enabled to couple the RD LTH (0) read latch signals onto line 90. Since the RVM reverse mode signal is asserted, driver 98 is disabled. However, a multiplexer 96 is enabled to couple the signal on line 90 onto line 31 for transfer in the reverse direction through routing network 30.

As described below in connection with Fig. 8A, a message also includes an acknowledgement field during which the PE chip 20 containing the PE identified by the address transmits an acknowledgement signal. If the RVM reverse mode signal is not asserted, multiplexer 96 couples the ACK acknowledgement signal onto line 31 for transmission to the PE transmitting the message.

After a message operation has been completed, that is, after data has been transmitted in forward and reverse directions, as required by the control microwords from array control logic 10, the TR transmit request flags 24 in the transmitting PE's 21 are reset. When the TR transmit request flag 24 is reset, the TR transmit request signal i negated. Since the RTR OUT EN IN router out enable in signal is asserted, the AND gate 101 is energized to transmit the RTR OUT EN OUT router out enable out signal t the next PE in the hierarchy to enable that PE to engage i a message operation. If the PE's TR transmit request flag is not set when the RTR OUT EN IN router out enable in signal is asserted, its AND gate corresponding to AND gate 101 generates the asserted RTR OUT EN OUT signal for transmission to the next PE.

B. Memory Interconnection

As has been noted, the array 11 can operate in either 1 Kb mode, in which each PE has an associated 1 Kb memory module 23 (Figs. 2 and 4) , or a 4 Kb mode. In the 4 Kb mode, the PE's are divided into groups of four PE's with on PE having access to its associated 1 Kb memory module 23 as well as to the 1 Kb memory modules that are normally (that is, in 1 Kb mode) associated with the three other PE's. This allows operations to proceed in parallel by all of the PE's, and then permits the results, which are contained in all of the memory modules 23 to be accessed by one PE for future processing, without requiring the results of the initial operations to be transferred by way of messages. Figs. 5A and 5B depict circuits for facilitating this operation.

With reference to Fig. 5A, a group of four processors 22(0) through 22(3) are connected to four memory modules 23(0) through 23(3) through memory control circuit 73. A multiplexer/demultiplexer 110 selectively couples data signals between all of memory modules 23(0) through 23(3) and processor 22(0), and a plurality of transceivers 111 through 113 each couple data signals between one memory module 23(1) through 23(3) and one associated processor 22(1) through 22(3). The ADRS (9:0) low order address signals from the control microword are coupled to all of the memory modules 23(0) through 23(3) in parallel to select one location in each of the memory modules, and ADRS (11:10) high order address signals from the control microword control the memory control circuit 73.

If the ADRS (11:10) signals are both negated, the array 11 (Fig. 1) operates in the 1 Kb mode. In that mode multiplexer/demultiplexer 110 transfers data signals between the processor 22(0) and memory module 23(0). The negated ADRS (11) and ADRS (10) signals energize a NOR gate 114, which, in turn, energizes output enable terminals on each o the transceivers 111 through 113. This enables the transceivers to couple data signals between memory modules 23(1) through 23(3) and respective processors 22(1) through 22(3).

If, on the other hand, either of the ADRS (11) or ADRS (10) signals are asserted, the array 11 operates in 4 Kb mode. In that case, NOR gate 114 disables the transceivers, which enables them to transmit data signals to the processors 22(1) through 22(3) having zero values. The ADRS (11:10) address signals enable the multiplexer/demultiplexer to transfer data signals between one of the memory modules 23(0) through 23(3), as determined by the binary encoded value of the encoded ADRS (11:10) address signals, and processor 22(0).

Fig. 5B depicts a schematic diagram illustrating one embodiment of a circuit of the portion of the memory control circuit 73 between memory module 23(3) and processor 22(3), and showing the circuit for coupling data signals to processor 22(0). When a location, as identified by the ADRS (9:0) address signals from the control^microword, is being read, memory module 23(3) transmits four data bit signals. -54- identified as DAT(O) through DA (3) on respective lines 120(0) through 120(3). The DAT(O) data signal corresponds to the contents of the bit location identified by the ADRS (9:0) address signals, and the other DAT(l) through DAT(3) signals correspond to the contents of the bit locations having the respective next sequential bit address locations

If the array 11 is not in the 4 Kb mode, an OR gate 12 will be de-energized so as to negate a 4K MODE signal, and an inverter 122 will generate an asserted NOT 4K MODE signal. The negated 4K MODE signal turns on four pass transistors 123(0) through 123(3), which couples the respective DAT(0) through DAT(3) data signals to lines 130(0) through 130(3) as DAT (0)-(3) TO PE (3) data to processing element signals for storage in read latch 74, which is represented by four latches 74(0) through 74(3). The latches store the respective signals and couple them to the PE 21(3) as LAT DAT (0)-(3) TO PE (3) latched data to processing element signals.

If the array is in the 4 Kb mode, the ADRS(11:10) address signals enable the OR gate 121 to assert the 4K MOD signal and negate the NOT 4K MODE signal. The negated NOT 4K MODE signal de-energizes transistors 123(0) through 123(3), thereby blocking the transmission of the DAT(O) through DAT(3) data signals to lines 130(0) through 130(3), and the asserted 4K MODE signal energizes transistors 131(0) through 131(3) to ground the lines 130(0) through 130(3) and negate all of the DAT (0)-(3) TO PE (3) data to processing element signals. As a result, all of the latches 74(0) through 74(3) are reset, resulting in negated LAT DAT (0)- (3) TO PE (3) latched data to processing element signals coupled to PE 21(3).

In addition, if both ADRS(11:10) address signals are asserted, indicating the data signals from memory module 23(3) are to be coupled to PE 21(0), a demultiplexer 124 energizes a line 125(3) to assert a MEM DAT (3) memory data signal, which energizes pass transistors 126(0) through 126(3). Energized pass transistors 126(0) through 126(3) couple the DAT(0) through DAT(3) data signals on lines 120(0) through 120(3) onto lines 127(0) through 127(3) as DAT (0)-(3) TO PE (0) data to processing element (0) signals, which are coupled to processor 3 (0) .

Data may be written from the PE 21(3) to memory module 23(3) if the NOT 4K MODE signal is asserted when the WRT EN (PE3) write enable signal is asserted by multiplexer 500 (Fig. 4). The WRT EN (PE3) signal enables AND gates 132(0) through 132(3), which couple the DAT (0)-(3) FRM PE(3) data signals from the PE(3) to lines 130(0) through 130(3), respectively. The asserted NOT 4K MODE signal turns on the transistors 123(0) through 123(3), enabling them to couple the DAT (0)-(3) FRM PE(3) data from processing element (3) signals onto lines 120(0) through 120(3), after which they are stored in memory module 23(3).

The memory control circuits 73 associated with processors 22(1) and 22(2) are similar to that depicted in Fig. 5B except that transistors 126(0) through 126(3) are connected to lines 125(1) and 125(2), respectively, and controlled by the MEM DAT (1) and MEM DAT (2) memory data signals from demultiplexer 124.

Fig. 5C depicts the circuit of the memory control circuit 73 associated with PE 21(0) . It can be seen that the memory control circuit 73 depicted in Fig. 5C is similar to the circuit depicted in Fig. 5B, except that there are no transistors 123(0) through 123(3) or 131(0) through"131(3) ; instead, the lines 130(0) through 130(3) are connected directly to lines 127(0) through 127(3), respectively, and the DAT (0)-(3) TO PE^" (0) data to processing element (0) -57- signals control the latches 74(0) through 74(3) directly. Furthermore, transistors 126(0) through 126(3) are controlled by the MEM DAT (0) signal from demultiplexer 12 to couple the DAT(O) through DAT(3) data signals from memo module 23(0) directly onto the lines 127(0) through 127(3) when the MEM DAT (0) signal is asserted.. The MEM DAT (0) signal is asserted by demultiplexer 124 when the ADRS (11) and ADRS (10) signals are both negated, that is, when the array 11 is not in the 4 Kb mode.

C. Status Signal Generation

As was noted above in the discussion regarding Figs. 2 and 4, the array control unit 10 (Fig. 1) uses a DONE statu signal transmitted over status bus 15 to determine if various operations have been completed by the PE's 21 in array 11 and to control further processing. The DONE statu signal is asserted, for example, during normalization durin floating point arithmetic when the fraction portion of the floating point numbers in all of the PE's have been normalized.

During normalization, until the DONE signal is asserted, the array control unit 10 repetitively transmits sequences of control microwords which enable the array 11 to engage in a normalization operation. Whether or not a particular PE actually engages in the operation depends on the state of various ones of the PE's flags. Furthermore, whether or not a PE engages in a normalization operation depends on the condition of the PE's M flag 62 (Fig. 4).

Fig. 6A depicts circuitry within a PE chip 20 for generating a DONE OUT signal which is used by array 11 (Fig. 1) in connection with generating the DONE status signal. With reference to Fig. 6A, each PE 22(0) through 22(N) on the PE chip 20 has extending therefrom a wire 150(0) through 150(N) which carries signals M(0) through M(N) . The asserted or negated condition of each signal M(0) through M(N) corresponds to the set or cleared condition of the PE's M flag 62 (Fig. 4). Thus, if a PE's M flag 62 is set, the PE's M (x) signal ("x" a value from zero to N) is asserted, otherwise the M(x) signal is negated.

All of wires 150(0) through 150(N) are connected to a common wire 151 in a wired-OR configuration. Wire 151 carries an M COMP composite "M" signal whose asserted or negated condition depends on the conditions of the M(0) through M(N) signals from all of the PE's on the chip. Thus, if the M flag 62 (Fig. 4) of any of processors 22(0) through 22(N) is set, the PE's corresponding M(x) signal i asserted, and the M COMP signal is asserted. If the M fla 62 of all of processors 22(0) through 22(N) is cleared, all of the M(0) through M(N) signals will be negated, and the COMP signal will be negated. As described above in connection with Fig. 4, when a PE has completed the normalization operation, its M flag 62 will be cleared; thus, when the M COMP composite M signal is negated, all of the PE's on the chip will have completed the normalization operation.

With reference to Fig. 6A, line 151 is coupled to one input terminal of an OR gate 155. If the signal on line 15 is asserted, the DONE X signal is also asserted, otherwise, the DONE X signal is negated. If the DONE X signal is asserted, the OR gate 155 also asserts the DONE OUT signal.

The OR gate 155 has a second input terminal which receives a DONE PREV done previous signal from an OR gate 156. OR gate 156, in turn receives as input signals DONE I A through DONE IN D signals, each of which corresponds to a DONE OUT signal from another PE chip. The DONE OUT signal from the PE chip 20 depicted in Fig. 6A may also be directe as an input signal to an OR gate corresponding to OR gate 156 on another PE chip. The resulting connection, an "OR" tree, results in one DONE OUT signal from one PE which is coupled to array control unit as the DONE signal. Thus, when the DONE signal is in a low voltage condition, depending on the condition of the DN done select signal, all of the M COMP composite M signals in all of the PE chips are negated, indicating that all of the M flags 62 are cleared. Thus, the operation (such as floating point number normalization) which was governed by the M flags 62 in the PE's has been completed.-

D. Local Address Register

As mentioned above in connection with Fig. 2, each PE chip 20 includes a local address register 28. Normally, the control microword from control unit 10 contains a memory address which is transmitted to all PE's 22 in parallel and identifies the locations in the memories 23 from which data is obtained or to which data is written. Accordingly, at one time, all PE's are using the same memory location, either for a read operation or for a write operation. In response to an LAM local address mode signal from the control microword, however, the PE's 22 instead use the contents of the local address register 28 on each PE chip 20. This permits the PE's 22 on different PE chips 20 to use different memory locations.

This can be particularly useful- as, for example, messages transmitted through router network 30 can include destination addresses in the memories of the receiving PE's and not just the identification of the receiving PE 21. If for example, a PE is to transfer data from one location in its own memory 23 to another location, that could be done easily through router network 30 by loading the address of the other location in the local address register and using that address to control writing while the data was being transmitted through the router network 30.

In addition, a PE can load the local address register 28 directly through its M flag 62. This permits the PE to indirectly address itself by reference to the local address register 28 during an operation.

With reference to Fig. 6B, in one specific embodiment the local address register provides the upper seven address bits ADRS (3:9), and the lower three bits are provided by the control microword, indicated on the Fig. by the MW MEM ADRS (0) through (2) microword memory address signals. Thus, in that embodiment, the transfers using the local address register 28 are "byte aligned". The local address register 28 includes seven stages 401(0) through 401(6) (generally referred to by reference numeral 401) , one stage 401 for each of the seven address bits. Since the stages 401 are similar, only stage 401(0) is depicted in detail.

Stage 401(0) receives an input signal from a multiplexer comprising pass transistors 402 and 403 controlled by the ENR enable router signal from the array control unit 10. When the ENR enable router signal is asserted, indicating the router network 30 is in operation, the transistor 402 passes the MSG IN message in signal from the router network 30 to an input transistor 404. If, on the other hand, the ENR signal is not asserted, an inverter 405 enables pass transistor 403 to couple the M COMP "M" composite signal from line 151 (Fig. 6) to the pass transistor 404.

Pass transistor 404 controls the loading of the stage 401(0) with a local address in response to a LAR LD local address load signal from the array control unit 10. The LAR LD local address load signal is pulsed in synchronism with a PH2 phase two timing signal during loading of the local address register 28. When the LAR LD local address load signal is asserted, transistor 404 is enabled to transfer the input signal selected by transistors 402 and 403 to the input of an inverter 405.

Pass transistor 404, inverter 405, a pass transistor 406 and a second inverter 407 form one stage of a seven- stage shift register, with the other stages of the shift register being in the other stages 401(1) through 401^'(6) of the local address register. After the LAR LD local address load signal has enabled inverter 405 to be energized by the input signal, the PH2 phase two timing signal negates the LAR LD local address load signal to turn off pass transisto 404. Thereafter, a PHI phase one timing signal is asserted, coupling the output signal from the inverter 405 to inverte 407, whose output signal is coupled to stage 401(1) of the local address register. If the LAR LD local address register load signal is asserted in synchronism with the next PH2 phase two timing signal, a transistor correspondin to transistor 404 in stage 401(1) is turned on to couple th signal to an inverter corresponding to inverter 405. The process continues as long as the local address register is to be loaded.

The selection of the address signals to be used as the ADRS (9:3) address signals, whether the contents of the local address register or the MW MEM ADRS (9:3) microword memory address signals, is made in a multiplexer comprising transistors 410 and 411, which are controlled in response to the LAM local address mode signal from the control microword. When the LAM signal is negated, an inverter 412 turns on transistor 411 which couples the MW MEM ADRS (3) microword memory address signal onto line 412(3) as the ADRS(3) address signal.

On the other hand, when the LAM local address mode signal is asserted, which can occur after the local address has been loaded into the stages 401, transistor 410 is turned on to couple the output of an inverter 413, which is controlled, in turn, by the output of inverter 405, onto line 412(3).- The output signal from inverter 413 has the same condition as the input signal to inverter 405, and so the signal coupled onto line 412(3) has the same condition as the signal supplied by pass transistor 404. Each stage also includes a network 414 for incrementi the local address register as data is being written into t successive locations of the memory 23. In particular, whe all of the MW MEM ADRS (2:0) microword memory address signals (2:0) are asserted, indicating a byte boundary has been reached, an AND gate 420 is energized. If the NM nibble mode signal from the control microword is asserted, and the MW MEM ADRS (2) signal is asserted, also indicatin a byte boundary has been reached during nibble mode, the A gate 420 is also asserted. When the PHI phase one timing signal is next asserted, a pass transistor 421 is enabled couple an asserted LAR CI local address register carry in signal to a NAND gate 422. If the LAM local address mode signal is asserted, the NAND gate 422 couples a negated signal to an input of a NAND gate 423 and an inverter 424 i stage 401(0) of the local address register. Until a byte boundary is passed, however, the LAR CI local address register carry in signal is negated, and so the output signal from NAND gate 422 will be asserted.

Network 414 includes two general portions, one generating an LU latch up signal which is coupled to a pass transistor 425 in the same stage 401(0) of the local addres register 28, and the other generating a CO carry out signal which is coupled to the next stage 401(1) which is used in the same way as the output signal from NAND gate 422 in stage 401(0). Transistor 425 is controlled by an LAR INCR local address register increment signal, which is also pulsed in synchronism with the PH2 phase two timing signal, after the local address register 28 has been loaded. Transistor 425, when turned on, couples the LU latch up signal from network 414 to the input of inverter 405 to control the signal therefrom.

Until a byte boundary is passed, the LAR CI local address register carry in signal is negated, and so the output signal from NAND gate 422 will be asserted. Accordingly, one input of NAND gate 423 in network 414 will be enabled, and the output signal from inverter 424 will be negated. Since the input to NAND gate 423 is enabled, its output will be the complement of the signal from inverter 407, which, in turn, is the same as the input to inverter 405. The signal from NAND gate 423 is coupled to one input of NAND gate 428.

The signal from inverter 407 is..also complemented by inverter 426 and coupled to one input of a NAND gate 427. Since the signal from inverter 424 is negated, the output signal from NAND gate 427 is always asserted. Thus, one input signal to NAND gate 428 corresponds to the complement of the signal from inverter 407 and the other input is asserted, and so the LU latch up signal from NAND gate 428 follows the signal from 407 when the LAR CI local address register signal is negated.

On the other hand, when the LAR CI local address register carry in signal is asserted, the output signal fro NAND gate 422 is negated. Accordingly, the output signal from NAND gate 423 is asserted. This output signal is coupled to one input of NAND gate 428. The output signal from inverter 407 is also complemented by inverter 426 and the result coupled to one input of NAND gate 427.

Since the output signal from NAND gate 422 is negated, the signal from inverter 424 to the second input of NAND gate 427 is asserted. Accordingly, the output of NAND gate 427, which is coupled to the second input of NAND gate 428, follows the complement of the signal from inverter 426, and thus follows the signal from inverter 407. Thus, one input signal to NAND gate 428 follows the signal from inverter 407 and the other input is asserted. Accordingly, the LU latch up signal generated by NAND gate 428 is the complement of the signal from inverter 407 when the LAR CI local address register signal is asserted.

With the LAR CI local address register carry in signal asserted, the inverter 424 enables one input of a NAND gate 430, which thereby couples the complement of the signal -from the inverter 407 as the CO carry out signal to stage 401(1) of the local address register. Stage 401(1) uses the CO carry out signal in the same way that stage 401(0) uses the output signal from NAND gate 422. Thus, when the output signal from inverter 407 in stage 401(0) is negated, the LU latch up signal in stage 401(1) follows the signal from its inverter 407, and when the output signal from inverter 407 in stage 401(0) is asserted, the LU latch up signal in stage 401(1) follows the complement of the signal from its inverter 407.

On the other hand, with the LAR CI local address register carry in signal negated, the inverter 424 enables the NAND gate 430 be asserted and the LU latch up signal in stage 401(1) follows the signal from its inverter 407 because there is no carr in. -69- It will be appreciated that, while one local address register 28 has been described as being included in the PE chip 20, and being used one at a time by all of the PE's o that chip, a local address register may be provided for ea of the PE's 22 on the PE chip 20.

E. Shift Register

As described above in connection with Fig. 4, the A shift register 52 (Fig. 4) can shift not only toward its S/ output terminal, but also toward its S/I input terminal. This capability permits the contents of the shift register to be easily shifted after such operations as floating poin addition and subtraction which require normalization.

Fig. 7 depicts a schematic circuit of a portion of the A shift register 52. The A shift register 52 includes, in one embodiment, eight cells 160(0) through 160(7), each of which includes four stages. Each stage stores one data bit Cell 160(0) contains the least significant data bits (indicated by the legend "LSB" in the Figure) and the succeeding cells 160(1), 160(2), and so forth, contains successively more signficant data bits, so that, depending on the A SHFT SZ shift register size signals from the control microword (see Fig. 4) cell 160(7) may contain the most significant data bits (as indicated by the legend "MSB" in the Figure) . Data is shifted into A shift register 52 through cell 160(0). The cell from which data bits are shifted out depends on the condition of the A SHFT SZ A shift register shift size signals from the control microword (see Fig. 4) .

Since the circuits of the cells are similar, the circuit of only cell 160(4) will be described in detail. The cell 160(4) includes four stages 161(0) through 161(3) each of which includes a pair of inverters, including input inverters 162(0) through 162(3) and output inverters 163(0) through 163(3). Stage 161(0) stores the low-order data bit of the data in the cell, and stage 161(3) stores the high- order data bit of the data in the cell.

The pair of inverters 162 (n) and 163(n) ("n" is an integer from 0 to 3) in each stage are connected through control transistors 180(n) and 181(n) to form a flip-flop which actually stores the data bit. In each stage, the output of inverter 162(n) is coupled through a control transistor 180(n), which is normally held on by an asserted PH 1 control signal, to the input of inverter 163 (n . The

output of inverter 163(n) is coupled through control transistor 181(n) to a node 182(n), to which the input of inverter 162(n) is also connected. The transistor 181(n) , which is normally held on by an asserted HOLD control signal, controls the feedback from the output of inverter 163 (n) to the input of inverter 162(n) which enables the inverters to latch the input at node 182(n) in a known manner.

During shift operations, the HOLD signal is negated, turning off transistor 181(n) and breaking the feedback path. In addition, the PH 1 timing is also negated, turni off transistor 180(n). Transistor 180(n), when off, isolates the output inverter 163(n) from the input inverte 162(n) . This allows the output inverter 163(n) of the sta to transmit a signal representing the stored data bit at t same time that the input inverter 162(n) of the stage receives a signal representing a signal representing a dat bit from another stage as described below.

Cell 160(4) receives four bits of data, comprising signals DATA IN (0) through DATA IN (3), from cell 160(3) respective lines 164(0) through 164(3). Cell 160(4) transmits four bits of data, comprising signals DATA OUT (

through DATA OUT (3), on respective lines 165(0) through 165(3), to stage 160(5). In addition, if the stage 160(4) is one from which data signals are transmitted to multiplexer 56 (Fig. 4) , each stage includes a pass transistor 170(0) through 170(3), connected to lines 165(0) through 165(3) and controlled by a CELL 4 OUT signal-, in ^"the A SHFT SZ A shift register 52 size signals. When the CELL 4 OUT signal is asserted, the transistors 170(0) through 170(3) are enabled to couple the DATA OUT (0) through DATA OUT (3) signals to the A shift register input terminal of multiplexer 56.

The direction and amount of shift in A shift register 52 are controlled by the SHFT DIR shift direction signal and the NM nibble mode signal, both of which are provided by the control microword from the array control unit 10 and by an EN SHFT enable shift timing signal, in addition to the timing control signals HOLD and PH 1. A shift register 52 generates three additional control signals in response to the SHFT DIR and NM signals. Specifically, A shift register 52 includes an AND gate 171 which receives the SHFT DIR and NM signals and generates an NS nibble shift signal, which, when asserted, enables one input of an AND gate 176. If the

NM signal is asserted, when the EN SHFT enable shift timin signal is then asserted, AND gate 176 asserts an NSU nibbl shift up signal. As will be described below, when the NSU nibble shift up signal is asserted, the A shift register 52 is conditioned to shift the data bits up a nibble (4 bits) at a time, that is, to the cell 160 which stores the next most significant data bits. Thus, when NSU nibble shift up signal is asserted, the data bits are conditioned to be shifted from cell to cell, and are stored in the same stage in each cell. The data bits shifted into cell 160(0) are provided by full adder 51 (Fig., 4) .

A second AND gate 172 generates a BS bit shift signal in response to the complement of the NM nibble mode signal, as inverted in inverter 173, and the SHFT DIR shift direction signal. Thus, when the NM nibble mode signal is negated and the SHFT DIR shift direction signal is asserted the BS bit shift signal from AND gate 172 will be asserted. When the EN SHFT enable shift timing signal is asserted, AN gate 177 asserts the BSU bit shift up signal. Since, however, the NM signal is negated, the NS nibble shift signal from AND gate 171 will be negated and the NSU nibble shift up signal from AND gate 176 will remain negated when

the EN SHFT enable shift signal is asserted. However, as will be described below, when the BSU bit shift up signal i asserted, the A shift register 52 is conditioned to shift the data bits up a -stage at a time. The data bit in the most significant stage in each cell is shifted into the least significant stage of the next more significant cell. The data bit shifted into the least significant stage [corresponding to stage 161(0)] is provided by the ADD OUT (0) signal from full adder 51 (see Fig. 4) .

Another AND gate 174 generates a DOWN signal in response to the complement of the NM nibble mode signal and the complement of the SHFT DIR shift direction signal, as inverted in inverter 175. When the EN SHFT enable shift timing signal is next asserted, an AND gate 178 enabled by the DOWN and EN SHFT enable shift signal asserts the SD shift down signal. Thus, when both the SHFT DIR shift direction signal and the NM nibble mode signal are negated, the DOWN signal is asserted. In addition, with the SHFT DI and NM signals both negated, both the NSU and BSU nibble an bit shift up signals from AND gates 176 and 177 are negated As will be described below, when the SD shift down signal i asserted, the A shift register 52 is conditioned to shift

the data bits down a stage at a time. The data bit in the least significant stage in each cell, except for the most significant stage in cell 160(7), is shifted into the the most significant stage of the next lesser significant cell. A data bit having the value zero is stored in the most significant stage [corresponding to stage 161(3)] of the most significant cell 160(7).

Each stage 161(0) through 161(3) further includes thre transistors, namely transistors 183 (n), 184(n) and 185(n) each enabled by one of the BSU bit shift up, NSU nibble shift up, and SD shift down control signals, respectively. When a control signal is asserted, the enabled transistor couples a signal representing the state of a data bit from another stage to the input node 182(n) of its respective stage.

In response to the assertion of the BSU bit shift up signal, with reference first to stage 161(0), the transisto 183(0) is turned on, coupling the DATA IN (3) data signal o line 164(3) to node 182(0). The DATA IN (3) data signal corresponds to the data bit from the most significant stage of cell 160(3). Further, in response to the BSU bit shift up signal, transistors 183(1) through 183(3) are turned on.

coupling the DATA OUT (0) through DATA OUT (2) signals to nodes 182(1) through 182(3), respectively. As a result, the data bits in each of the stages, namely the most significant stage of cell 160(3) and the stages 161(0) through 161(2) are shifted up one stage and are stored in stages 161(0) through 161(3) of cell 160(4).

In response to the assertion of the NSU nibble shift up signal, transistors 184(0) through 184(3) are turned on, . coupling the DATA IN (0) through DATA IN (3) data signals on lines 164(0) through 164(3), respectively to nodes 182(0) through 182(3). The DATA IN (0) through DATA IN (3) data signals correspond to the data bits from the respective stages of cell 160(3). As a result, the data bits in each of the stages of cell 160(3) are shifted up one cell (four stages) and are stored in stages 161(0) through 161(3) of cell 160(4) .

In response to the assertion of the SD shift down signal, transistors 185(0) through 185(2) are turned on, coupling the DATA OUT (1) through DATA OUT (3) data signals on lines 165(1) through 165(3), respectively to nodes 182(0) through 182(2) for storage in stages 161(0) through 161(2). In addition, transistor 185(3) is turned on to couple the

DATA OUT (0) [CELL 160(5)] data signal from the least significant stage in cell 160(5) [that is, the stage corresponding to stage 161(0)] to node 182(3) for storage. stage 161(3). A transistor in the most significant stage. cell 160(3) corresponding to transistor 185(3) receives th DATA OUT (0) data out signal from stage 161(0) for storage therein. As a result, the data bits in each of the stages are shifted down one stage.

IV. Routing Network 30

A. General Description

As described above with reference to Fig. 2, the routing network 30 is used to transmit messages between PE' which are typically not likely to be on the same PE chip. general organizational diagram of one embodiment of a routing network constructed in accordance with the inventio is depicted in Fig. 8A. With reference to Fig. 8A, the routing network 30 comprises a plurality of identical switching chips (each switching chip is depicted as a rectangle in the Figure) organized in three switching stage 201, 202 and 203. The internal switching circuits of chips will be described below in connection with Figs. 9 and 10A -78- through IOC.

Before proceeding further, it would be helpful to describe the format of a message transmitted through routin network 30. With reference to Fig. 8B, a message begins with a header 204 of twenty-three bits, namely bits (0) through (22) , which identify the intended recipient PE, followed by the message data bits beginning with bit (23) . The header includes three router control fields, identified by reference numerals 205-207, each of which controls the switching through one of the switching stages 201-203, respectively. Each switching stage 201-203 retires one router control field, that is, it does not pass the bits i that field on to the next switching stage or to the recipient PE chips.

Each router control field begins with a protocol bit which, when asserted, indicates that message bits follow. If the protocol bit P is not received at an input terminal, the switching chips ignore succeeding signals at that inpu terminal during the minor message transfer cycle. The fou bits following each protocol bit are RTR ADRS router addre bits which are used by the router to establish a switching path through the stage as described below in connection wi

Figs. 9 and 10A through IOC. If a switching path is established through all of the switching stages, a protocol bit P (bit 15) is transmitted to the PE chip pair (see Fig. 2) connected to the output line from stage 203. In response, during bit 16, the receiving PE chips generate an ACK acknowledgement signal which is transmitted over the switching path established through the routing network 30. At this time, the switching chips condition the switching path to receive a signal from the recipient PE chip pair an couple it to the PE chip 20 which is originating the message. The PE chip 20 originating the message uses the ACK acknowledgement signal to clear the transmitting PE's T flag 24 (Figs. 2 and 4) .

After the transfer of the ACK acknowledgement signal, six-bit PROC ID processor identification is transmitted ove the switching path. The first bit identifies the PE chip 2 in the chip pair which contains the PE to receive the message, and the last five bits identify the PE on the identified chip. The next.-bits are the message bits which are coupled by the PE chip's router control circuit 33 (Figs. 2 and 4) to the PE. -80- With reference again to Fig. 8A, the first two switching stages 201 and 202 are divided into four switchin groups, one of which is depicted in th Figure. The last switching stage 203 operates as a crossbar switch to switch the outputs from the four switching groups to the PE chips. Since the four switching groups are identical, only the one depicted in the Figure will be described in detail. That switching group includes sixteen switching chips, identifie by reference numerals 201(0) through 201(15) in switching stage 201, and an additional sixteen switching chips, identified by reference numerals 202(0) through 202(15) in switching stage 202.

Each switching chip 201(n) and 20 (n) ("n" is an integer) has sixty-four input terminals, generally - identified by reference numeral 210, and sixty four output terminals, generally identified by reference numeral 211. The sixty-four output terminals are grouped into sixteen output groups each having four output terminals. Each of the sixteen output groups is identified by one encoding of the four RTR ADRS router address bits in the portion of the header 204 (Fig. 8B) that is retired by the switching chip. On receiving at an input terminal RTR ADRS router address -81- bits having a particular encoding, the switching chip attempts to open a switching path from that input terminal to one of the four output terminals in the output group identified by that encoding. If RTR ADRS router address bits received at four or fewer input terminals of the switching chip identify a particular output group, switchin paths can be established to the output group from all of th input terminals. If, on the other hand RTR ADRS router address bits identifying a particular output group are received at more than four input terminals, at most four switching paths can be established to the output group and the other messages, for which switching paths cannot be established, are blocked.

The output terminals of the switching chips 201(0) through 201(15) included in the first switching stage 201 are connected to the input terminals of the switching chips 202(0) through 202(15) of the second switching stage as follows. All of the output terminals of the chips 201(0) through 201(15) in groups associated with the "0000" encoding of the RTR ADRS router address retired by switchin chips 201(0) through 201(15) are connected to the input terminals of switching chip 202(0). This is depicted in th

Figure by the wires identified by the legends "4 x 0000" extending from the right sides of switching chips 201(0) through 201(15) to the left side of switching chip 202(0). (In the legend "4 x 0000" in Fig. 8A, the first number, "4", indicates that the associated line includes 4 connecting wires, and the second number "0000" indicates that the wires are associated with the "0000" encoding of the first four RTR ADDRS router address signals, which are retired by the switching stage 201.)

Similarly, all of the output terminals of the chips 201(0) through 201(15) in groups associated with the "0001" encoding of the RTR ADRS router address retired by switching chips 201(0) through 201(15) are connected to the input terminals of switching chip 202(1), as depicted in Fig. 8A by the wires identified by the legends "4 x 0001" extending from the right sides of switching chips 201(0) through 201(15) to the left side of switching chip 202(1). This interconnection pattern between the switching chips in the first two stages 201 and 202 is repeated for all of the other switching chips 202(2) through 202(15); as depicted in the Figure, the input terminals of switching chip 202(15) are connected to the output terminals of switching chips -83- 201(0) through 201(15) that are in groups associated with the "1111" encoding of the RTR ADRS router address retire by the switching chips 201(0) through 201(15). Thus^", the output terminal groups in each of the switching chips 201 through 201(15) are connected to diverse ones of the switching chips 202(0) through 202(15) in the second switching stage 202, with the connection depending on the RTR ADRS router address bits that are retired by the switching chips in the first stage.

As described above, the sixty-four output terminals each of the switching chips 202(0) through 202(15) in the second switching stage 202 are also divided into sixteen groups each having four terminals. Like switching chips 201(0) through 201(15), each output terminal group of switching chips 202(0) through 202(15) is associated with one encoding of the four RTR ADRS router address signals retired by those chips. The output terminal groups of switching chip 202(0), for example, connect to wires havi the legends "4 x 0000 0001" through "4 x 0000 1111", with "4" indicating the number of wires (with each wire being connected to one output terminal) , the first block of fou digits "0000" indicating the RTR ADRS router address bits

retired by the first switching stage 201, and the second block of four digits "0000" and "1111" indicating the RTR ADRS router address bits that are retired by the switching chips 202(0) through 202(15). The wires attached to the output terminal groups of switching chips 202(1) through 202(15) are identified by similar legends.

As described above, switching stages 201 and 202 are divided into four switching groups (one group is depicted i the Figure) the outputs of which are switched by crossbar stage 203. The other three switching groups of switching stages 201 and 202 include chips corresponding to switchin chip 202(0) having output terminal groups associated with the RTR ADRS router address bits "0000 0000" through "0000 1111", chips corresponding to switching chip 202(1) having output terminal groups associated with the RTR ADRS router address bits "0001 0000" through "0001 1111", and so on.

The crossbar switching stage 203 includes sixteen crossbar switching blocks of sixteen crossbar switching chips 203(0) through 203(255) [only switching chips 203(0) through 203(3) are shown in the Figure], with each crossba switching block switching the output signals from output terminal groups in the corresponding switching chips in ea

of the switching groups. Each of the sixteen crossbar switching chips in each crossbar switching block, in turn, switches the outputs from corresponding terminal groups in corresponding stage 202 switching chips from all of the four groups. Thus, crossbar switching chip 203(0) is connected to, and switches the outputs from, the "4 x 0000 0000" output terminal groups from switching chip 202(0) and corresponding switching chips in each of the other switching groups.

Similarly, crossbar switching chip 203(1) is connected to, and switches the outputs from, the "4 x 0000 0001" terminal groups from switching chip 202(0) and corresponding switching chips in each of the other switching groups. In addition, the last crossbar switching chip in the crossbar switching block associated with switching chip 202(0) (and corresponding switching chips in the other switching - - groups), namely crossbar switching chip 203(15), is connected to, and switches the outputs from, the "4 x 0000 1111" terminal groups from switching chip 202(0) and corresponding switching chips in each of the other switching groups.

The crossbar switching chips 203(16) through 203(255) in each of the other crossbar switching blocks are similarly connected to the other switching chips 202(1) through 202(15), and corresponding switching chips in the other switching groups.

Each crossbar switching chip operates as a crossbar switch. That is, the crossbar switching chip associates each of its sixteen output terminals with one encoding of the final four RTR ADRS router address bits which it receives from the switching stage 202 switching chips. Thus, for example, switching chip 203(0) has sixteen output terminals each connected to a wire having the legend "1 x 0000 0000 0000" through "1 x 0000 0000 1111". In this case, the leading "1" -indicates that there is one wire, the first group of digits, namely the first group of "0000", corresponds to the RTR ADRS router address bits retired by switching stage 201, the second group of digits, namely the second group of "0000", corresponds to the RTR ADRS router - address bits retired by switching stage 202, and the third - group of digits, which range from "0000" through "1111", corresponds to the RTR ADRS router address bits received and retired by the crossbar switching chip 203(0). The output terminals of the other crossbar switching chips 203(1) through 203(255) have similar associations wit the RTR ADRS router address bits. It will be appreciated b those skilled in the art that:

(1) each of the sixteen output terminals of each crossbar switching chip 203(0) through 203(255) is associated with one encoding of the third group of RTR ADRS router address bits (that is the third group of digits associated with each output line from the crossbar switchin chips depicted in Fig. 8A, )

(2) each of the sixteen crossbar switching chips in each crossbar switching block is associated with one encoding of the second group of RTR ADRS router address bit (that is the second group of digits associated with each output line from the crossbar switching chips depicted in Fig. 8A, ) , and

(3) each of the sixteen crossbar switching blocks is associated with one encoding of the first group of RTR ADRS router address bits (that is the first group of digits associated with each output line from the crossbar switchin chips depicted in Fig. 8A, ) .

Accordingly, the routing network depicted in Fig. 8A provides switching paths from the input terminals of the switching chips of stage 201 to the output terminals of crossbar switching chips of crossbar stage 203 for all possible encodings of the RTR ADRS router address bits.

The routing network 30 depicted in Fig. 8A h^'as a number of advantages, in particularly relating to the association of multiple output terminals from switching stages 201 and 202 into groups associated with the same encoding of the RTR ADRS router address bits. This significantly reduces the number of switching stages that are required, reducing the number of switching chips and wires interconnecting the chips. This, in turn, serves to enhance the reliability of the routing network, since the failure rate is directly related to the number of chips and, particularly, wires interconnecting- those chips.

In addition, by associating multiple wires in a terminal group with each encoding of the RTR ADRS router address bits, if a wire, or its associated driving circuitry on the chip, is defective, the switching chips will still be able to transmit over the remaining wires in the terminal group. The message traffic that can be accommodated will be -89- so ewhat reduced, but the messages can still be transferr through the routing network 30.

B. Switching Chip

The internal organization and circuitry within a switching chip 201(n) or 202(n) will now be described in connection with Figs. 9 and 10A through IOC. As will be made clear in the following discussion, the same chip as described herein can also be used as a crossbar switching chip in the crossbar stage 203.

Fig. 9 (comprising Figs. 9-1 and 9-2) depicts the general circuit organization of a portion of the switching chip constructed in accordance with the invention. With reference to Fig. 9, a switching chip includes a plurality of switching circuits each associated with one input terminal. One such switching circuit, designated by the reference numeral 210(0), is shown in Fig. 9. In one embodiment, a switching chip includes sixty-four switching circuits. All of the switching circuits are identical, an so only one is depicted in the Figure. Each of the switching circuits is associated with one of the input terminals; switching circuit 210(0) is associated with the

input (0) terminal, and generally switching circuit 210(u) ("u" is an integer) is associated with the input (u) terminal.

Switching circuit 210(0) includes an input (0) circuit 211(0), sixteen coupling groups 212(0) through 212(15) [generally designated 212(v) ] and sixteen output terminal- groups 217(0) through 217(15) [generally designated 217(v) ] . Each output terminal group includes four output circuits generally identified by the reference numeral 217(v) (w) . The input (0) circuit 211(0), the circuit of which is described below in connection with Fig. 10A, receives an IN(0) input signal from a PE chip pair if the chip is in stage 201 (see Fig. 8A) or from a preceeding switching stage if the chip is in stages 202 or 203. In response to the IN(0) input signal the input (0) circuit 211(0) transmits a ADRS(0) address signal onto a line 213(0) representing complement of the the RTR ADRS router address bits to be retired by the chip and a DAT(0) data signal onto a line 214(0) representing the remaining bits of the ^"message.

Each coupling group 212(v) includes an address decodin and control circuit 215(0) through 215(15) [generally designated by reference numeral 215(v) ] and a coupling

circuit 216(0) through 216(15) [generally 216(v) ] . One circuit for a coupling group will be described in detail below in connection with Figs. lOB-1 and 10B-2. Line 213(0) from input (0) circuit 211 couples the ADRS(O) address signal to the address circuit 215(0) in coupling group 212(0) and line 214(0) couples the DAT(0) data signal to all of the 216(v) coupling circuits in parallel. In response to the four RTR ADRS router address bits retired by the switching chip, one of the address control circuits 215(v) in switching circuit 210(0) will assert an ADRS(u)EN(v) signal. For example, if the IN(0) signal received by input (0) circuit 211(0) has RTR ADRS router address bits having the encoding "0000", address control circuit 215(0) will assert the ADRS(0)EN(0) signal.

Each coupling circuit includes four coupling modules generally designated by reference numeral 216(v)(w), with the (v) index identifying the coupling circuit and the (w) index identifying the module within the circuit. For example, the coupling modules of coupling circuit 216(0) depicted in Fig. 9 are designated by the reference numerals 216(0) (0) through 216(0) (3), the coupling modules of

coupling circuit 216(1) being designated by the reference numerals 216(1) (0) through 216(1) (3), and so forth.

The coupling modules 216(v)(0) through 216(v)(3) in each switching circuit 210(0) through 210(63) are connected to one of sixty-four data lines 220(0) (0) through 220(15) (3) [generally designated by reference numeral 220(v)(w)]. Data line 220(0) (0) is connected in parallel to all of the coupling modules 216(0) (0) in all of the switching circuits 210(0) through 210(63), data line 220(0) (1) is connected in parallel to all of coupling modules 216(0) (1) in all of the switching circuits 210(0) through 210(63)and so forth.

At any one time only one of the coupling modules connected to one data line 220(v)(w) is enabled,, as described below, to couple a data signal, corresponding to the DAT(u) data signal from its respective input (u) circuit 211(u), onto the data line 220 as a GRP(v)DAT(w) group data signal. The output terminal group circuit 217(v) receives the GRP(v)DAT(w) group data signals relating to its terminal group and couples them as the switching chip's output signals.

Each coupling module 216(v)(w) is enabled to couple the DAT(u) data signal onto the respective ^ata line 220(v)(w) -93- in response to two enabling signals. One enabling signal, the ADRS(u)EN(v) coupling group enabling signal, is generated by the address control circuit 215(v) in respons to the decoding of the RTR ADRS router address bits from t ADRS(u) signal.

The second signal which enables a coupling module 216(v)(w) in switching circuit 210(u) is a GRP(v)EN(w) IN(u enabling signal. For clarity in Fig. 9, only the GRP(0)EN(0)IN(0) through GRP(0)EN(3) IN(0) enabling signals associated with output terminal group circuit 217(0) are depicted. In an actual circuit, the other output terminal group circuits 217(1) through 217(15) would also generate corresponding enabling signals which would be coupled to coupling modules 216(n)(0) through 216(n) (3) in coupling circuits 216(1) through 216(15).

Four daisy chain GRP(v)EN(w) enabling signals are originated by each output terminal group circuit 217(v) us by the coupling modules 216(v)(w) in the switching circuit 210(u). The GRP(v)EN(w) enabling signals are normally asserted by the output terminal group circuits 217(v). However, under some circumstances, such as, for example, i a communications path to the next stage is disrupted due t a broken wire or the like, the output terminal group circuit 217(v) negates the corresponding GRP(v)EN(w) enabling signal. Circuitry for detecting the presence of a disrupted communication path is described below in connection with Figs. 10A and IOC.

The signals from the output terminal group circuits 217 (v) are initially coupled to the coupling modules 216(v)(w) of the switching circuit 210(0) as a GRP(v)EN(w)IN(0) enabling signal depicted on Fig. 9. If a coupling module 216(v)(w) in switching circuit 210(0) receives an asserted GRP(v)EN(w) IN(0) enabling signal, and is enabled by the ADRS(u)EN(v) signal from the address control circuit, it transmits a negated GRP(v)EN(w)OUT(0) enabling signal to coupling module 216(v)(w) of switching circuit 210(1). In that case, the coupling module will couple the DAT(0) signal onto the 220(v)(w) data line. On the other hand, if a coupling module 216(v)(w) is not enabled by the ADRS(u)EN(v) signal from the address control circuit, it asserts a GRP(v)EN(w)OUT(0) enabling signal, which is coupled to coupling module 216(v) (w) of switching circuit 210(1). The coupling modules 216(v) (w) of that, and other switching circuits 210(2) through 210(63) operate similarly.

Thus the GRP(v)EN(w) enabling signals, originating fro the output terminal groups 217(v), are passed through the coupling modules 216(v)(w) in a daisy-chained manner. If a coupling module 216(v) (w) is enabled by the associated address control circuit 215(v) and receives the GRP(v)EN(w) enabling signal in an asserted condition [that is, its GRP(v)EN(w)IN(u) signal is asserted], the coupling module uses the associated data line 220(v)(w) and blocks the GRP(v)EN(w) enabling signal [that is, it transmits the GRP(v)EN(w)OUT(u) enabling signal to the next switching circuit in a negated condition] .

In any coupling group 216(v), only one coupling module 216(v)(w) will be enabled to couple the DAT(u) signal onto the associated data line 220(v)(w). Thus, for example, if the leftmost coupling module 216(0) (0) (as depicted in Fig. 9) is enabled by the ADRS(0)E (0) signal from the address control circuit 215(v), but is- disabled by a negated GRP(0)EN(0)IN(0) signal, it in turn enables the coupling module 216(0) (.1) to its immediate right. If that coupling module 216(0) (1), is enabled by an asserted GRP(0)EN(1)IN(0)

signal, it will couple the DAT(O) signal onto the data line 221(0) (1). However, if the GRP(O)EN(1)IN(0) signal is negated, that coupling module 216(0) (1) will enable the coupling module 216(0) (2) to its immediate right. Essentially, the ADRS(0)EN(0) enabling signal from address control circuit 215(0) is essentially daisy-chained-through the coupling modules 216(0) (0) through 216(0) (3).

For any^' received message, in any switching circuit 210(u) at most one of the ADRS(u) (EN(v) enabling signal from one of the address control circuits 215(v) will be asserted. That is, for any message only one coupling circuit 216(v) in any switching circuit 210(u) will be enabled by an address control circuit 215(v). The circuit of an address control circuit 215(v) will be described below in connection with Fig. 1QB-1. In brief, each address control circuit 215(v) includes an input terminal 221(v) and an output terminal 222(v). Each address control circuit 215(v) receives through its input terminal 221(v) an input signal which is related to the ADRS(u) address signal, which in turn is related to the RTR ADRS router address bits. - The address control circuit 215(v) also includes a latch and an inverter. As long as the latch is reset, the address

control circuit couples, through the inverter, the complement of the signal at its input terminal 221(v) to the output terminal 222(v). If the latch is set, it remains set throughout the message, and causes the inverter to be- bypassed, thus enabling the true value of the signal at the input terminal 221(v) to be coupled to the output terminal 222(v).

As depicted in Fig. 9, the address control circuits 215(v) in a switching circuit 210(u) are connected serially, so that each input terminal 221(1) through 221(15) of address control circuits 215(1) through 215(15) is connected to the output terminal 222(0) through 222(14), respectively, of the address control circuit 215(0) through 215(14) to its left. Initially, all of the latches are reset. Accordingly, if the IN(0) signal carries an initial RTR ADRS router address bit having the binary value "0", the ADRS(0) signal received by input terminal 221(0) is asserted. As a result, the inverter in the address control circuit 215(v) provides a negated output signal at its output terminal 222(0), which is coupled to the input terminal 221(1) of address control circuit 215(1).

Address control circuit 215(1), in turn receives the negated input terminal through its input terminal 221(1) and its inverter couples an asserted output signal to the output terminal 222(1), which is, in turn, coupled to the input terminal 222(2) of address control circuit 215(2). The remaining address control circuits 215(v) of switching circuit 210(0) operate similarly. Thus, when the address control circuit 215(15) receives an input signal through its input terminal 221(15), the input signals to address control circuits 215(v) having an even-numbered or zero index (v) will be asserted, and the input signals to the address control circuits having an odd-numbered index (v) will be negated.

At this point, the latches in the address control circuits 215(v) latch the input signal. If the input signal to an address control circuit 215(v) is asserted, the latch remains cleared, otherwise the latch is set. After a, latch is set, it enables its address control circuit's inverter to D by-passed, and so the signal received at it's -input terminal 221(v) will be coupled directly to the output terminal 222(v). If the latch is set, the address control circuit 215(v) is also inhibited from asserting the ADRS(u)E (v) coupling circuit enabling signal.

The input (0) circuit then transmits an ADRS(O) signal related to the second RTR ADRS router address bit to the input terminal 221(0) of address control circuit 215(0). The address control circuits 215(v) ("v" an even number or zero) whose latches are cleared operate in the same manner as described above and transmit a signal at their output terminals 222(v) which is the complement of the signal at their input terminals 221(v). At the same time, the address control circuits 215(v) ("v" an odd number) whose latches are set pass the signal which they receive at their input terminals 221(v) to their output terminals 222(v). Thus, with a second RTR ADRS input negated, the input signals at the address control circuits 215(0) and others 215(v) whose indexes (v) are divisible by four are asserted. Contrariwise, the input signals to the other address control circuits 215(v) whose latches are not already set [that is, circuits 215(v) where (v) is an even number but not divisible by four] are negated.

At this point, the latches again latch the input signals. This process is repeated for the third and fourth RTR ADRS router address bits which are received. It will be appreciated that in response to each RTR ADRS router address

bit, the latches in one-half of the address control circuits that are still clear, 215(v) are set. Accordingly, after four RTR ADRS router address bits are received and processed by the circuit depicted in Fig. 9, the latch of only one address control circuit will be cleared, and the others will all be set. In the example above in which the RTR ADRS router address bits have the encoding "0000", only the latch of address control circuit 215(0) will be cleared, and that address control circuit will assert the ADRS(0)EN(0) signal. The operations in response to other encodings of the RTR ADRS router address bits are similar, except that each different encodings enable latches of a different address control circuit 215(v) to remain cleared, enabling that address control circuit to assert its ADRS(u)EN(v) signal.

C. Description of Specific Circuits

With this background, specific circuits.for various portions of the switching chip depicted in Fig. 9 will be described. Fig. 10A depicts a circuit of an input (u) circuit. Figs. lOB-1 and 10B-2, together comprising Fig. 10B, depict a circuit of an address control circuit 215(v) and its related coupling group 216(v). Fig. 10C depicts a circuit relating to a data line 220(v)(w) that

comprises part of an output terminal group 217(v).

In the following description, the index (u) , which is used in the description of Fig. 9, has not been used^' in -^• connection with the signal names and reference numerals. ^"'

i. Input Circuit 211

With reference to Fig. 10A, an input circuit 211 receives an IN input signal on an input line 230. The IN input signal is coupled to two latches, namely a miss latch 231 and a break latch 232, to an inverter 233 and to one input terminal of an AND gate 234. In response to the IN input signal, the inverter 233 generates an ADRS signal which is coupled to the address control circuit depicted in Fig. lOB-1.

As has been described above in connection with Fig. 8B, the first signal bit received by the switching chip, and thus the first input signal received by the input circuit 211, is a. protocol bit, labeled P in Fig. 8B. If the bit has a binary value of one, that is, if the IN signal is asserted during the P bit time, a message is being received through the input circuit 211. If, on the other hand, the bit has a binary value of zero, that is, if the IN signal is

negated during the P bit time, no message is being transferred through the input circuit. Miss latch 231 latches the condition of the IN signal during the P bit time, which is defined by an LD MISS LTH load miss latch timing signal which derived from a RTR CTRL router control signal (not .shown) from the array control unit 10 (Fig. 1)^*.

In particular, the RTR CTRL router control signal is transmitted to the switching chips of the first stage 201 (Fig. 8A) of the routing network 30 (Fig. 2)-. When the switching chips 201(0) through 201(63) of that stage have retired all of the bits of the first router control field 205 (Fig. 8B) and are prepared to transmit to the next stag 202, they transmit a RTR CTRL NXT router control next stage signal to the switching chips 202(0) through 202(63). This process is repeated as between switching chips of stage 202 and crossbar stage 203, and as between chips of crossbar stage 203 and the input terminals of PE chips 20 (Fig. 2) . Thus, the array control unit 10 may initiate a transfer through the routing network with the single RTR CTRL router control signal coupled only to the first switching stage, with each stage controlling timing of the next stage. This arrangement simplifies expansion of the routing network, as

the array control unit does not have to be informed of any minimum or maximum number of switching stages.

The miss latch 231 includes two inverters 228 and 229, with the output of inverter 228, which provides a MISS signal, being connected to the input of inverter 229. The output of inverter 229 is connected through a pass transistor 235 to the input of inverter 228. The IN input signal is also coupled to the input of inverter 228 through a pass transistor 236. The pass transistors are, in turn, controlled by the LD MISS LTH load miss latch timing signal. Transistor 236 is on, that is, conducting, when the LD MISS LTH load miss latch timing signal is asserted, and otherwise is off. Transistor 235, on the other hand, is controlled by the complement of the LD MISS LTH load miss latch timing signal, and is on, that is, conducting, when the LD MISS LTH load miss latch timing signal is negated, and otherwise off.

Thus, when the LD MISS LTH load miss latch timing signal is asserted, transistor 236 ^"is turned on and transistor 235 is turned off. Transistor 236 couples the IN input signal to the input of inverter 228. The MISS signal from inverter 228 is the complement of the received IN input signal while the LD MISS LTH load. miss latch timing signal

is asserted. Inverter 229 receives the MISS signal from inverter 228 and complements it again. When the LD MISS LTH load miss latch timing signal is again negated, at the end of the time defining the P protocol bit (Fig. 8B) , transistor 236 turns off, isolating the input of inverter 228 from the IN input signal.

At the same time, however, transistor 235 turns on, coupling the output signal from inverter 229 to the input of inverter 228» Since the output signal from inverter 229 has the same condition as the IN input signal while the LD MISS LTH load miss latch timing signal was asserted, the MISS signal from inverter 228 remains in the same condition; that is, the IN signal is latched in the miss latch 231. The MISS signal represents the complement of the P bit, that is, the complement of the IN input signal when the LD MISS LTH load miss latch timing, signal was last asserted. When the LD MISS LTH load miss latch timing signal is next asserted, transistor 235 is turned off, isolating the input of inverter 228 from the output of inverter 229 and permitting thi IN input signal to be coupled to the input of inverter 228 through transistor 236. If the MISS signal is asserted, the IN input signal during the P protocol bit time was negated, and so no message is being received through the input circuit 211. If, on the other hand, the MISS signal is negated, the IN signal was asserted during the P protocol bit time, and a message is being received through the input circuit 211.

The break latch 232 is constructed in a manner similar to miss latch 232 and operates in the same way in response to a LD BRK LTH IN load break latch in timing signal from the array control unit 10 (Fig. 1) . The break latch is used to latch the IN input signal while the communication path from the preceeding chip is being tested to determine if it has been disrupted due to a broken wire or otherwise, as described below. If the communication path is disrupted, the break latch 232 asserts a BREAK signal and negates its complement, a NO BRK no break signal. The BREAK signal is coupled to an input of OR gate 240 which generates a BRK OR OUT break OR out signal. Or gate 240 receives a BRK OR IN break OR input signal from a similar circuit in an input circuit 211 on the switching chip or, as described below in connection with Fig. 10C, an output terminal group circuit 217(v). The OR gates 240 are connected in an OR chain which drives an illumination device (not shown) so that, if any communications path to which the chip is connected is disrupted, the illumination device is energized.

The NO BRK no break signal is coupled to a NAND gate 241, to which the complement of the MISS signal is also coupled. If either a miss condition occurs, or the communication pathway to input terminal 230 is disrupted, NAND gate 240 asserts a BLOCK signal which is coupled to the address control circuits 216(v), as described below in connection with Fig. lOB-1. In brief, the BLOCK signal, when asserted, inhibits the address control circuits connected to the input circuit 211 from asserting the ADRS(u)EN(v) coupling circuit enabling signals.

As described below in connection with Fig. 10B-2, a CHP MISS chip miss signal is asserted if no^' coupling circuit in a switching circuit 210(u) is enabled to couple the DAT(u) data signal onto a data output line 220. An OR gate 242 is energized to assert a DAT BLK data blocking signal when either the BLOCK signal or the CHP MISS chip miss signal is asserted. The complement of the DAT BLK data blocking signal is generated by an inverter 243 whose output is coupled to AND gate 234. Thus, if the DAT BLK data blocking signal is asserted, the AND gate 234 is disabled, inhibiting the AND gate 234 from coupling the IN input signal received on line 230 to line 214 as the DAT signal.

AND gate 234 is also disabled by an inverter 245 when an REV TIME reverse time signal is asserted, which occurs when the ACK acknowledge bit of a message (see Fig. 8B) is expected to be transferred from the destination PE to the originating PE (or from where it was blocked in the router) or when the RVM bit in the control microword is set enabling a reverse data transfer through the routing network 30. The REV TIME reverse time signal is provided by the array control unit 10. Thus, when the REV TIME reverse time signal is asserted, the AND gate 234 isolates line 214 from the signal on line 230. Conversely, the AND gate 234 is enabled to couple the IN signal on line 230 onto line 214 as the DAT data signal both when the REV TIME reverse time signal is negated, which occurs at times other than during the acknowledgement bit time or reverse transfers, and when the DAT BLK data block signal is not asserted, indicating that a coupling module is enabled to couple the DAT signal onto an output data line 220(v)(w).

The remaining circuitry on Fig. 10A performs two functions. In particular, during the time for transferring the message acknowledgement bit or during a reverse direction transfer, which occurs when the REV TIME reverse time signal is asserted, the signal representing the ACK bit or the reverse direction data is received by the input circuit over line 214. Circuitry is provided to couple the signal onto line 230 during this time. In addition, circuitry is provided to perform the communications path continuity test, which occurs in response to a PR RTR IN(L) precharge router in (asserted low) signal and a DR LOW RTR IN(L) drive low router in signal, both of which are provided by the array control unit 10 (Fig. 1) .

During the reverse direction time, the REV TIME reverse time signal is asserted, which disables AND gate 234. The PR RTR IN(L) precharge router in (asserted low) and the DR LOW RTR IN(L) drive low router in (asserted low) signals are both negated, and thus are at a high voltage level, thus enabling one input each of NAND gates 246 and 247 and two inputs of a NAND gate 250. Since the REV TIME reverse time signal is asserted, the inverter 245 couples a low input signal to the third input of NAND gate 250. The output of

NAND gate 250 thus is at a high voltage level, which enables an AND gate 251, whose output is connected to line 230.

The DAT signal on line 214, which at this time represents the ACK acknowledgement bit or the reverse direction data, is complemented in an inverter 252, inverted twice through the two enabled NAND gates 246 and 247, and complemented again in a second inverter 253, whose output is coupled to the second input of AND gate 251. Thus, the signal provided by inverter 253 to AND gate 251 has the same sense as the DAT signal on line 214. Since AND gate 251 is enabled by the NAND gate 250, it couples the signal from inverter 253 onto line 230. Thus, during the reverse direction time, the DAT signal, which represents the ACK acknowledgement bit or the reverse direction data, is coupled from line 214 to line 230, that is, through the input circuit 211.

At other times than the reverse direction time, however, the REV TIME reverse time signal is negated. Inverter 245 thus couples a high input signal to NAND gate 250. Since the other input signals to NAND gate 250 are normally negated and thus at a high voltage level, the output of NAND gate 250 is low, thus disabling AND gate 251.

Accordingly, as long as the PR RTR IN(L) precharge router in (asserted low) and DR LOW RTR IN(L) drive low router in (asserted low) signals are negated, during times other than reverse direction time, AND gate 251 is inhibited from coupling the signal from inverter 253 to line 230.

The communications path continuity test is performed by circuitry in both the input circuit 211 depicted in Fig. 10A and the output terminal group described below in connection with Fig. IOC, and will be described below in connection with Fig. IOC.

ii. Address Control Circuit 215 And Coupling Circuit 216

Figs. lOB-1 and 10B-2 depict the circuit diagram of an address control circuit and four coupling modules 216(0) through 216(3) [generally identified by reference numeral 216(w)] comprising one coupling circuit 216.

a. Address Control Circuit 215

As described above, the^.address control circuit 215 includes a latch 260 comprising two inverters 261 and 262. The output of inverter 261 is connected to the input of inverter 262, and the output of inverter 262 is connected to

a pass transistor 263, which, when it is turned on by an ADRS HOLD address hold timing signal, couples the output signal from inverter 262 to the input of inverter 261. Another pass transistor 264, also controlled by the ADRS HOLD address hold timing signal, couples the output signal from inverter 262 to control two pass transistors 265 and 266, and to an inverter 267 which controls a pass transistor 270.

An EN RTR enable router timing signal from the array control unit 10 (Fig. 1) , when negated, through an inverter 268 turns on a transistor 269. This places a high signal at the input of inverter 261. When the router chip is enabled, the array control unit 10 asserts the EN RTR enable router timing signal, turning off transistor 269 and isolating the input of inverter 261 from the high input. Thus, immediately after the EN RTR enable router signal is asserted, latch 260 is in the condition such that the input of inverter 261, and thus the output of inverter 262 are high. In addition, the ADRS HOLD address hold timing signal is asserted. In this condition, the high output signal from inverter 262 enables the pass transistors 265 and 266 to be in the on, or conducting, condition, and pass transistor 270 to be in the off, or non-conducting condition by inverter 267.

In response to the assertion of an ADRS TIME address time signal, which is asserted to define each RTR ADRS router address bit (see Fig. 8B) in the IN input signal (Fig. 10A) , a pass transistor 271 is turned on, allowing the ADRS address signal from inverter 233 (Fig. 10A) to be coupled through transistor 271 and transistor 265. While the ADRS TIME signal is asserted, the ADRS HOLD address hold signal is negated, turning off transistors 263 and 264. While transistor 264 is turned off, transistors 265 and 266 and inverter 267 are held in the condition they were in by residual charge which remains on their inputs. However, the ADRS signal overcomes any residual charge on the input to inverter 261, and thus inverter 261 couples, at its output, a signal corresponding to the complement of the ADRS address signal.

Since at this point pass transistor 266 is maintained on by a residual charge on its gate, it couples the output signal from inverter 261 to a line 272, which carries an ADRS NXT address next signal to the next address control circuit 215 in the switching circuit 210 (see Fig...9). It

will be appreciated that the inverter 261 included in the latch circuit 260, which comprises the latch described above in connection with Fig. 9, is also used as the inverter described above in connection with Fig. 9.

It will also be appreciated that, if the ADRS address signal is in a high condition while the ADRS TIME address time signal is asserted, the output signals from inverters 261 will be maintained in the conditions which they had when the address control circuit was initialized in response to the EN RTR enable router initialization signal. However, if the ADRS address signal is in a low condition, the input signal to inverter 261 will be low, as will the output of inverter 262. Thus, when the ADRS TIME signal is next negated, and the ADRS HOLD address hold signal is asserted, the output of inverter 262 will be in a low condition. Transistor 264, which is turned on by the ADRS HOLD address hold signal, couples the low output signal to transistors 265 and 266, turning them off, and to inverter 267, turning on transistor 270.

Since transistor 270 is turned on, the ADRS address signal is coupled onto line 272 as the ADRS NXT signal to the next address control circuit, bypassing the inverter 261

as described above in connection with Fig. 9. In addition, since transistor 265 is turned off, the ADRS address signal is blocked in the path to inverter 261 when the ADRS TIME address timing signal is next asserted, thereby isolating the latch 260 from the ADRS address signal. It will be appreciated that, once a low ADRS signal is received^"by latch 260, the latch 260 remains isolated from the ADRS address signal until the EN RTR enable router initialization signal is asserted, which enables transistor 265 to be turned on.

The output of inverter 261 is coupled to one input of a NOR gate 273. If the latch 260 receives four consecutive asserted signals (that is, signals at a high voltage level) through transistors 265 every time the ADRS TIME signal is asserted, which corresponds to the four RTR ADRS bits that are retired by the switching chip on which the circuit is resident (see Fig. SB) , the output signal from inverter 261 is in a low voltage condition. If the BLOCK signal from input circuit 211 (Fig. 10A) is negated (in a low voltage condition) , NOR gate 273 is energized to apsert (high) an ADRS EN enabling signal, which corresponds to the ADRS(u)EN(v) enabling signal described above in connection with Fig. 9. -115- Thus, if the address control circuit 215 receives a asserted ADRS address signal during the four consecutive asserted ADRS TIME address time enabling signals which define the four RTR ADRS bits, and if the P protocol bit (see Fig. 8B) has been received, which ensures that the BLOCK signal will be negated, the ADRS EN address enabli signal will be asserted. Otherwise, the ADRS EN address enabling signal will be negated.

b. Coupling circuit 216

The coupling circuit 216 includes four coupling mod 216(0) through 216(3) which are depicted in Figs. lOB-l 10B-2. Coupling module 216(0) is shown in Fig. lOB-l an the other coupling modules are depicted on Fig. 10B-2. Since the four coupling modules are similar, only coupli module 216(0) will be described in detail.

With reference to Fig. lOB-l, coupling module 261(0) receives the ADRS EN address enabling signal from the address control circuit 215 and a GRP EN 0 IN (L) group enable (asserted low) signal [which corresponds to the GRP(v)EN(w) group enable signal described above in connection with Fig. 9]. If the ADRS EN address enable -

signal is asserted (high) but the GRP EN 0 IN (L) group enable signal is asserted (that is, also low) , an AND gate 280 is asserted, which turns on a transistor 281, which in turn couples the DAT signal from Fig. 10A onto output data line 220(v) (0) .

The asserted (low) GRP EN 0 IN (L) group enable in signal is complemented by an inverter 282 to enable one input of a NAND gate 283. The second input of NAND gate 283 is the complement of the ADRS EN address enable signal from an inverter 284. If the ADRS EN signal is asserted, the output signal from inverter 284 is negated. As will be described below, with the input signals to NAND gate 283 in that condition, the NAND gate 283 generates a high (negated) GRP EN 0 OUT (L) group enable out (asserted low) output signal, which inhibits the succeeding coupling modules connected to the data output line 220(v)(0) from coupling data signals onto the data line.

When the ADRS EN address enable signal from address control circuit 215 is negated, the GRP EN OUT (L) group enable out signal has the same condition as the GRP EN 0 IN (L) group enable in signal. ^'If the ADRS EN address enabling signal is negated, the AND gate 280 is disabled and

transistor 281 is maintained in the off condition. Thus, the transistor 281 blocks the DAT signal from being couple onto data line 220(v)(0). At the same time, the negated ADRS EN signal is complemented to a high voltage condition by inverter 284, which enables one input of NAND gate 283. If the GRP EN 0 IN (L) group enable in signal is asserted (low) , inverter 282 couples a high signal to the other inp of NAND gate 283. With the input signals in that conditio NAND gate 283 asserts the GRP EN 0 OUT (L) group enable ou signal, allowing another coupling module 216(0) in another switching circuit 210 (Fig. 9) to use the data line 220(v) (0) .

If, on the other hand, the GRP EN 0 IN (L) group enab in signal is negated (high) , the inverter 282 couples a lo input signal to the NAND gate 283, enabling the NAND gate transmit a high, or negated, GRP EN 0 OUT (L) group enable signal. Thus, when the ADRS EN address enable signal is negated, the asserted or negated condition of the GRP EN 0 OUT (L) group enable out signal to the succeding coupling module connected to the data line 220(v)(0) is the same as the condition of the received GRP EN 0 IN (L) group enable in signal. /00456

-118- If the ADRS EN address enable signal is asserted, but the GRP EN 0 IN (L) group enable signal is negated (high) , an AND gate 285 is energized to assert an EN 1 enabling signal, which is coupled to coupling module 216(1) and is used by that coupling module in the same way that the coupling module 216(0) uses the ADRS EN address enable signal. If the GRP EN 0 IN (L) group enable signal is asserted (low) , the AND gate 285 is disabled and the EN 1 enabling signal is negated. Thus, it will be appreciated that if a coupling module 216(w) is enabled by the ADRS EN address enable signal from the address control circuit 215 or by the EN W ("W" being 1 or 2) , if it is inhibited from coupling the DAT data signal onto its respective data line 220(v) (w) by the negated GRP EN W IN (L) group enable signal, it will assert the EN W (W being 1, 2, or 3) to enable the next coupling module to its right. However, if coupling module 216(w) is enabled by both the ADRS EN address enable or EN W enabling signals and the GRP EN W I (L) group enable signal, it will negate the EN W signal to the next coupling module to its right.

Also depicted in Figs. lOB-l and 10B-2 is circuitry f terminating the data lines 220(v)(0) through 220(v) (3) to -119- ensure that the voltage level of the signals on those line does not float if no coupling module couples its associate DAT data signals onto the data lines. If the GRP EN W OUT (L) group enable out signal from the last coupling module 216(0) associated with a data line 220(v)(0) is asserted, that is, in a low voltage condition, an inverter 290 energizes a transistor 291, which causes a ground level signal to be placed on the data line 220(v)(0). This grou level signal is coupled to the next switching chip in the routing network, or to the routing network input terminal a PE chip. The ground level signal provides a negated signal during the P protocol bit time (Fig. 8B) , enabling the miss latch 231 to be set in the input circuit 211.

As has been mentioned, the switching chips described connection with Figs. 9 and 10A through IOC can also perfo crossbar switching, and thus are useful in the crossbar switching stage 203 (Fig. 8A) . This is accomplished if on the GRP EN 0 IN (L) group enabling signal is asserted by t output terminal group and the other signals, namely the GR EN 1 IN (L) through GRP EN 3 IN (L) group enabling signals are negated.

c. Output Terminal Group

Each output data line 220(v)(w) has an associated output circuit 217(v)(w) in the output terminal group 217(v) (Fig. 9) , which also generates the GRP EN W (L) group enable signal. The circuit in an output terminal group associated with one output data line will be described in connection with Fig. IOC.

With reference to Fig. IOC, the output data line is connected to an inverter 300, which complements the DATA OUT signal and couples it to one input of a NAND gate 301. The NAND gate 301 is controlled by a PR RTR OUT (L) precharge router out (asserted low) signal which is used in connection with the communication path continuity test described below. The PR RTR OUT (L) precharge router out signal is normally negated, that is, at a high voltage level, and in that condition, the output signal from the NAND gate 301 is the same as the DATA OUT signal.

The output of NAND gate 301 is coupled to a second NAND gate 302, which is controlled by a DR LOW RTR OUT (L) drive low router out (asserted low) signal, which also is used in connection with the communication path continuity test. The

DR LOW RTR OUT (L) drive low router out (asserted low) signal is also normally negated, thus providing an output signal which is the complement of the output of NAND gate 301. The output of NAND gate 302 is again complemented in an inverter 303 which couples a DATA NXT CHP data next chi output signal onto an output line 304, which is connected the next switching stage or to the router input of a PE ch pair. It will be appreciated that, as long as the PR RTR OUT (L) precharge router out and DR LOW RTR OUT (L) drive low router out signals are negated, the DATA NXT CHP data next chip output signal provided by inverter 303 correspon to the DATA OUT signal input to inverter 300.

The output line 304 is also connected to a break latc 305, which is constructed in the same way as break latch 23 (Fig. 10A) and operates in response to a signal from line 304 and a LD BRK LTH OUT load break latch out timing signa from array control unit 10 (Fig. 1) . Break latch 305 is used in connection with the communications path continuity ^' test described below. If there is no disruption of the communications path, break latch 305 asserts a NO BRK OUT (L) no break out (asserted low) signal, which enables one input of an OR gate 306. OR gate 306 is also enabled by a

DRV DIS output driver disable signal from other circuitry (not shown) in the array processor which controls the communications paths through routing network 30 (Fig. 2) . If the DRV DIS driver disable signal is negated (low) and the NO BRK OUT (L) no break out signal is asserted (low) , O gate 306 generates the GRP EN W (L) group enable signal, which is coupled to a coupling module as depicted on Figs. lOB-l and 10B-2. The NO BRK OUT (L) no break out signal is also coupled to an OR gate 307, which performs th same function as OR gate 240 described in connection with the input circuit 211 (Fig. 10A) .

During times other than the ACK acknowledgement bit time (Fig. 8A) or reverse direction transfers, the REV TIME reverse time signal is negated. When the REV TIME reverse time signal is negated, an AND gate 310 is disabled, isolating the data, line 220(v)(w) from the output line 304. However, when the REV TIME reverse time signal is asserted, AND gate 310 is enabled to couple the signal received fro downstream circuitry, either a switching chip in a succeeding switching stage (Fig. 8A) or a PE chip, to the data line 220(v)(w). With reference to Fig. lOB-l, it will be appreciated that the transistor 281, when enabled by

energized AND gate 280 to couple the DAT data signal from line 214 (Fig. 10A) onto data line 220(v)(0), is also capable of coupling the signal representing the ACK acknowledgement bit or the reverse direction data signal i the reverse direction from line 220(v)(0) onto line 214.

If the GRP EN N (L) group enable signal is negated (that is, in a high voltage condition) , a transistor 313 i turned on. While the REV TIME reverse time signal is negated, an inverter 312 also turns on another transistor 311. Transistor 311 and 313 are connected in series betwe line 220(v)(w) and a ground level signal represented by th legend Vss. Thus, when the two transistors are turned on, ground level signal in placed on data line 220(v)(w) so th the voltage level of the signal on the data line does not float while the GRP EN N (L) group enable signal inhibits the coupling modules 216(v)(w) from coupling DAT data signals onto the data line. When the REV TIME reverse tim signal is asserted, inverter 312 turns off the transistor 311, which isolates the data line 220(v)(w) from the groun level signal.

d. Communications Path Continuity Test

As has been mentioned the input circuit 211 depicted in Fig. 10A and the output circuit 217(v) (w) depicted in Fig. IOC includes circuitry for performing a communications path continuity test between switching chips and for inhibiting transfers in the event of a disruption in the continuity of the communications path. With reference to Figs. 10A and IOC, the test is performed in three stages, each initiated by means of a timing signal. First, the communications line, that is, line 230 or line 304 (it will be appreciated that line 304 corresponds to line 230 in the switching chip of the next stage) , is precharged to a high voltage level by either the input circuit 211 or the output circuit. This occurs in response to the PR RTR IN (L) precharge router in (low) signal, if the precharge operation is performed by the input circuit 211, or the PR RTR OUT (L) precharge router out (low) signal if the precharge operation is performed by the output circuit.

After the communications line has been precharged, the other circuit (that is, the output circuit if the input circuit 211 precharged the line, or the input circuit 211 if the output circuit precharged the line) places a signal on -125- the communications line which drives the line to a low voltage condition. This occurs in response to the DR LOW IN (L) drive low router in (low) or DR LOW RTR OUT (L) dr low router out (low) signal. Finally, the break latch 23 or 305 in the circuit which initially precharged the line enabled by a LD BRK LTH IN or LD BRK LTH OUT signal, latc the state of the signal on the communications line. If t low signal reaches the break latch the communications lin is continuous and the break latch asserts a NO BRK IN or BRK OUT no break signal. On the other hand, if the low signal fails to reach the break latch, the communications path is disrupted, and so the NO BRK IN or NO BRK OUT sig will be negated. The test is performed by both the input circuit 211 and the output circuit at both ends of the communications path so that break latches in both circuits can be properly conditioned.

In the following description, the output line 304 is taken as the other end of input line 230; otherwise stated input circuit 211 is in the next switching stage from the output circuit depicted in Fig. IOC. With this background with reference to Figs. 10A and IOC, when the PR RTR IN (L precharge router in (low) signal is asserted (low) , the

output signal from NAND gate 246 is driven high. Since the DR LOW RTR IN (L) drive low router in (low) signal is negated (that is, high) NAND gate 246 couples a low signal to inverter 253, which inverted to a high signal and coupled to one input of AND gate 251.

Since the PR RTR IN (L) precharge router in signal is asserted (low) , NAND gate 250 also couples a high signal to the second input of AND gate 251. As a result, AND gate 251 is energized, forcing line 230 to a high voltage level.

The output circuit depicted in Fig. IOC then places a low signal on line 304. This occurs when the DR LOW RTR OUT (L) drive low router out (low) signal is asserted. When that occurs, NAND gate 302 transmits a high signal, which is complemented to a low voltage level by inverter 303.

Returning to Fig. 10A, when the signal from the output circuit depicted in Fig. 10C should have reached the input circuit, the LD BRK LTH IN signal is asserted, enabling break latch 232. If the communications line between the input circuit 211 and the output circuit is not disrupted, the signal on line 230 is at a low voltage level, which is inverted by the break latch circuit to provide the asserted

(high) NO BRK IN no break in signal. On the other hand, if the communications path is disrupted, the low signal from inverter 303 will not reach the break latch. Instead, the signal on line 230 will be high, which is inverted to form the negated (low) NO BRK IN no break in signal.

Similar operations are performed to load the break latch 305 in the output circuit depicted in Fig. IOC. In particular, when the PR RTR OUT (L) precharge router out (low) signal is asserted (low) , the output signal from NAND gate 301 is driven high. Since the DR LOW RTR OUT (L) driv low router out (low) signal is negated (that is, high) NAND gate 302 couples a low signal to inverter 303, which is inverted to a high signal to precharge line 304.

The input circuit depicted in Fig. 10A then places a low signal on line 230. This occurs when the DR LOW RTR IN (L) drive low router in (low) signal is asserted. When tha occurs, NAND gate 247 transmits a high signal, which is complemented to a low voltage level by inverter 253. The low signal from inverter 253 disables AND gate 251, causing a low signal to be placed on line 230.

Returning to Fig. IOC, when the signal from the input circuit depicted in Fig. 10A should have reached the output

circuit, the LD BRK LTH OUT signal is asserted, enabling break latch 305. If the communications line between the input circuit 211 and the output circuit is not disrupted, the signal on line 304 is at a low voltage level, which is coupled by the break latch circuit 305 to provide the asserted (low) NO BRK OUT no break out signal. On the oth hand, if the communications path is disrupted, the low signal from AND gate 251 will not reach the break latch 30 Instead, the signal on line 304 will be high, which is coupled through latch 305 to provide the negated ^'NO BRK OUT(L) no break out (asserted low) signal.

The foregoing description has been limited to a specific embodiment of this invention. It will be apparen however, that variations and modifications may be made to the invention, with the attainment of some or all of the advantages of the invention. Therefore, it is the obj ct the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.

What is claimed as new and desired to be secured by . Letters Patent of the United States is:

Claims

1. In an array processing system including a plurality of processor means each associated with a memory means, each o said memory means storing data in a plurality of addressabl storage locations, a memory interconnection circuit for coupling data between the memories and the processors in response to address signals having a low order portion identifying a corresponding location in all of said memory means and a high order portion identifying a memory, said memory interconnection circuit comprising:

A. demultiplexer means connected to one of said processor means and all of said memory "means for coupling data from a location identified by the low order portion of the address signals in one of the memories identified by th high order portion of the address signals to said one processor; and

B. a plurality of transceiver means each connected to one of said other processor means and its associated memory means and enabled by the high order portion of the address signals to block transmission of data between the memory means and the associated processor means when the high orde

portion identifies any memory means other than the memory means associated with said one processor, and otherwise allowing data to be coupled between the processors and the associated memories.

2. A routing network for transmitting messages between processors in an array processing system, each message including an address identifying the processor to receive the message, the routing network including a plurality of interconnected switching stages each of which includes mean for using a portion of the address to establish a path through the routing network, each stage being connected to the next stage over a plurality of communication lines divided into a plurality of groups, each group having a plurality of communications lines each associated with one encoding of the portion of the address used by the stage, each said switching stage including means for decoding the said address portion and for coupling messages onto the communications lines in the corresponding groups.

3. A communications -path continuity detection circuit for detecting the continuity of a communications path comprising:

A. latch means connected to said communications path and having a set condition and a reset condition, said latc means transmitting a break signal when in the set condition

B. means connected to said communications path for precharging the communications path to a predetermined high voltage level;

C. means connected to said communications path at the other end of the communications' path for driving a low voltage signal on the communications path; and

D. means connected to said latch means for enabling said latch means to latch the signal represented by the voltage level on the communications path after the driving means has been enabled.