Massively Parallel Diagonal-Fold Tree Array Processor

These invention relate to computers and particularly to massively parallel array processors.

This continuing application claims priority and is filed before the patenting or abandonment of or termination of proceedings on the following parent applications or an application similarly entitled to the benefit of the filing date of these applications:

U.S. Ser. No. 07/526,866, filed May 22, 1990, of S. Vassiliadis et al, entitled: Orthogonal Row-Column Neural Processor (now U.S. Pat. No. 5,065,339, issued Nov. 12, 1991); and its following pending divisional applications:

U.S. Ser. No. 07/740,355, filed Aug. 5, 1991, of S. Vassiliadis et al, entitled: Scalable Nerual Array Processor (now U.S. Pat. No. 5,146,543) and,

U.S. Ser. No. 07/740,556, filed Aug. 5, 1991, of S. Vassiliadis et al, entitled: Adder Tree for a Neural Array Processor (now U.S. Pat. No. 5,146,420); and,

U.S. Ser. No. 07/740,568, filed Aug. 5, 1991, of S. Vassiliadis et al, entitled: Apparatus and Method for Neural Processor (abandoned and continued as Ser. No. 08/000,915 and issued as U.S. Pat. No. 5,251,287); and,

U.S. Ser. No. 07/740,266, filed Aug. 5, 1991, of S. Vassiliadis et al, entitled: Scalable Neural Array Processor and Method (now U.S. Pat. No. 5,148,515); and

U.S. Ser. No. 07/682,786, filed Apr. 8, 1991, of G. G. Pechanek et al, entitled: Triangular Scalable Neural Array Processor (abandoned and continued as Ser. No. 08/231,853 and issued as U.S. Pat. No. 5,309,106); and

U.S. Ser. No. 07/681,842, filed Apr. 8, 1991, of G. G. Pechanek et al, entitled: SPIN: A Sequential Pipelined Neurocomputer (now U.S. Pat. No. 5,337,395); and

U.S. Ser. No. 07/702,261, filed May 17, 1991, of G. G. Pechanek et al, entitled: A Learning Machine Synapse Processor System Apparatus (abandoned and continued as Ser. No. 08/079,697, now abandoned); and

U.S. Ser. No. 07/702,260, filed May 17, 1991, of G. G. Pechanek et al, entitled: Virtual Neurocomputer Architectures for Neural Networks (now U.S. Pat. No. 5,243,688); and

U.S. Ser. No. 07/702,262, filed May 17, 1991, of G. G. Pechanek et al, entitled: Scalable Flow Virtual Learning Neurocomputer (abandoned and continued as Ser. No. 08/080,989 and issued as U.S. Pat. No. 5,329,611); and

U.S. Ser. No. 07/702,263, filed May 17, 1991, of G. G. Pechanek et al, entitled: PLAN: Pyramid Learning Architecture Neurocomputer (abandoned and continued as Ser. No. 08/079,695 and issued as U.S. Pat. No. 5,325,464); and

U.S. Ser. No. 07/864,112, filed Apr. 6, 1992, of G. G. Pechanek et al, entitled: Massively Parallel Array Processor (now abandoned).

In addition, filed concurrently is a related application:

U.S. Ser. No. 07/881,594 filed May 15, 1992, of G. G. Pechanek et al, entitled: Scalable Massively Parallel Group Partitioned Diagonal-Fold Switching Tree Computing Apparatus. (abandoned and continued as Ser. No. 08/496,826)

These applications and the present continuation-in-part application are owned by one and the same assignee, International Business Machines Corporation of Armonk, N.Y.

The descriptions set forth in these above applications are hereby incorporated into the present application.

During the detailed description which follows the following works will be referenced as an aid for the reader. These additional references are:

1. U.S. Ser. No. 07/799,602, filed Nov. 27, 1991, by H. Olnowich, entitled: "Multi-Media Serial Line Switching Adapter for Parallel Networks and Heterogenous and Homologous Computer Systems" abandoned and continued as Ser. No. 08/390,893 and issued as U.S. Pat. No. 5,535,373.

2. U.S. Ser. No. 07/798,788, filed Nov. 27, 1991, by P. M. Kogge, entitled: "Dynamic Multi-mode Parallel Processor Array Architecture" abandoned and continued as Ser. No. 08/324,295 and issued as U.S. Pat. No. 5,475,856. (Systems which allow dynamic switching between MIMD, SIMD, and SISD.)

3. D. E. Rumelhart, J. L. McClelland, and the PDP Research Group, Parallel Distributed Processing Vol. 1: Foundations Cambridge, Mass.: MIT Press 1986. (Herein referred to as "Rumelhart 86".)

4. APPARATUS AND METHOD FOR NEURAL PROCESSOR, S. Vassiliadis. and G. G. Pechanek, U.S. Ser. No. 07/526,866, filed May 18, 1990 now U.S. Pat. No. 5,065,339, (herein sometimes referred to as "SNAP" or "Vassiliadis SNAP 90".)

5. A TRIANGULAR SCALABLE NEURAL ARRAY PROCESSOR, G. G. Pechanek, and S. Vassiliadis, U.S. Ser. No. 07/682,785, filed Apr. 8, 1991 now U.S. Pat. No. 5,274,097, (herein sometimes referred to as "T-SNAP" or "T-SNAP".)

6. A LEARNING MACHINE SYNAPSE PROCESSOR SYSTEM APPARATUS, G. G. Pechanek et al, U.S. Ser. No. 07/702,261, filed May 17, 1991 abandoned and continued as Ser. No. 08/079,697, now abandoned, (herein sometimes referred to at the "Learning Machine")

7. E. B. Eichelberger and T. W. Williams, "A Logic Design Structure for Testability," Proceedings 14th Design Automation Conference, IEEE, 1977. (Herein referred to as "Eichelberger 77".)

8. J. J. Hopfield, "Neurons With Graded Response Have Collective Computational Properties Like Those of Two-State Neurons," Proceedings of the National Academy of Sciences 81, pp. 3088-3092, May, 1984. (Herein referred to as "Hopfield 84".)

9. J. L. McClelland and D. E. Rumelhart, Explorations in Parallel Distributed Processing: A handbook of Models, Programs and Exercises Cambridge, Mass.: MIT Press, 1988. (Herein referred to as "McClelland 88".)

10. J. J. Hopfield, "Neural Networks and Physical Systems with Emergent Collective Computational Abilities," Proceedings of the National Academy of Sciences 79, pp. 2554-2558, 1982. (Herein referred to as "Hopfield 82".)

11. M. J. Flynn, J. D. Johnson, and S. P. Wakefield, "On Instruction Sets and Their Formats," IEEE Transactions on Computers Vol. C-34, No. 3, pp. 242-254, March 1985. (Herein referred to as "Flynn 85".)

These additional references are incorporated by reference.

In the never ending quest for faster computers, engineers are linking hundreds, and even thousands of low cost microprocessors together in parallel to create super supercomputers that divide in order to conquer complex problems that stump today's machines. Such machines are called massively parallel. Multiple computers operating in, parallel have existed for decades. Early parallel machines included the ILLIAC which was started in the 1960s. Other multiple processors include (see a partial summary in U.S. Pat. No. 4,975,834 issued Dec. 4, 1990 to Xu et al) the Cedar, Sigma-1, the Butterfly and the Monarch, the Intel ipsc, The Connection Machines, the Caltech COSMIC, the N Cube, IBM's RP3, IBM's GF11, the NYU Ultra Computer, the Intel Delta and Touchstone.

Large multiple processors beginning with ILLIAC have been considered supercomputers. Supercomputers with greatest commercial success have been based upon multiple vector processors, represented by the Cray Research Y-MP systems, the IBM 3090, and other manufacturer's machines including those of Amdahl, Hitachi, Fujitsu, and NEC.

Massively Parallel (MP) processors are now thought of as capable of becomming supercomputers. These computer systems aggregate a large number of microprocessors with an interconnection network and program them to operate in parallel. There have been two modes of operation of these computers. Some of these machines have been MIMD mode machines. Some of these machines have been SIMD mode machines. Perhaps the most commercially acclaimed of these machines has been the Connection Machines series 1 and 2 of Thinking Machines, Inc. These have been essentially SIMD machines. Many of the massively parallel machines have used microprocessors interconnected in parallel to obtain their concurrency or parallel operations capability. Intel microprocessors like i860 have been used by Intel and others. N Cube has made such machines with Intel '386 microprocessors. Other machines have been built with what is called the "transputer" chip. Inmos Transputer IMS T800 is an example. The Inmos Transputer T800 is a 32 bit device with an integral high speed floating point processor.

As an example of the kind of systems that are built, several Inmos Transputer T800 chips each would have 32 communication link inputs and 32 link outputs. Each chip would have a single processor, a small amount of memory, and communication links to the local memory and to an external interface. In addition, in order to build up the system communication link adaptors like IMS C011 and C012 would be connected. In addition switches, like a IMS C004 would be profided to provide, say, a crossbar switch between the 32 link inputs and 32 link outputs to provide point to point connection between additional transputer chips. In addition, there will be special circuitry and interface chips for transputers adapting them to be used for a special purpose tailored to the requirements of a specific device, a graphics or disk controller. The Inmos IMS M212 is a 16 bit process, with on chip memory and communication links. It contains hardware and logic to control disk drives and can be used as a programmable disk controller or as a general purpose interface. In order to use the concurrency (parallel operations) Inmos developed a special language, Occam, for the transputer. Programmers have to describe the network of transputers directly in an Occam program.

Some of these MP machines use parallel processor arrays of processor chips which are interconnected with different topologies. The transputer provides a crossbar network with the addition of IMS C004 chips. Some other systems use a hypercube connection. Others use a bus or mesh to connect the microprocessors and there associated circuitry. Some have been interconnected by circuit switch processors that use switches as processor addressable networks. Generally, as with the 14 RISC/6000s which were interconnected last fall at Lawarence Livermore by wiring the machines together, the processor addressable networks have been considered as coarse-grained multiprocessors.

Some very large machines are being built by Intel and nCube and others to attack what are called "grand challenges" in data processing. However, these computers are very expensive. Recent projected costs are in the order of $30,000,000.00 to $75,000,000.00 (Tera Computer) for computers whose development has been funded by the U.S. Government to attack the "grand challenges". These "grand challenges" would include such problems as climate modeling, fluid turbulence, pollution dispersion, mapping of the human genome and ocean circulation, quantum chromodynamics, semiconductor and supercomputer modeling, combusion systems, vision and cognition.

Our Massively Parallel Diagonal-Fold Tree Array Processor architecture, which is the subject of tills patent, is applicable for modeling high computational parallel data algorithms, for example matrix processing and high connectivity neural networks. To demonstrate the general processing capability of our system an example of matrix multiplication is included.

During the detailed discussion of our invention, we will reference other work including our own unpublished works, as mentioned above. These background literature references are incorporated herein by reference.

It is a problem for massively parallel array processors to attack adequately the matrix processing problems which exist.

Our newly developed computer system may be described as a Massively Parallel (MP) Diagonal-Fold Tree Array Processor which operates in a Single Instruction Multiple Data (SIMD) fashion with general purpose application capability. The MP system we prefer will have a N.sup.2 Processor Element (PE) structure in which each PE .contains instruction and data storage units, receives instructions and data, and executes instructions. The N.sup.2 PE structure should contain N communicating ALU trees, N Tree Root Processors (TRP), and a mechanism for communicating both instructions and data back to the PEs by means of the communicating ALU trees.

The preferred apparatus which will be described contains N.sup.2 PEs placed in the form of a N by N matrix, with PEs identified by column-row subscripts PE.sub.column,row =PE.sub.ij, that has been folded along the diagonal and made up of Diagonal-PEs and General-PEs.

In our preferred system, the Diagonal-PEs are comprised of single Processing Elements, PE.sub.ii, and the General-PEs are comprised of two symmetric Processing Elements, PE.sub.ij and PE.sub.ji, that are merged together and which are associated with the same PE elements of the N by N PE array prior to folding.

Our new organization of PEs and new PE architecture is described in the best way we know to implement the improvements with an example implementation for matrix multiplication and discussion concerning neural network emulation, matrix addition, and Boolean operations.

These and other improvements are set forth in the following detailed description. For a better understanding of the invention, together with advantages and features, reference may be had to the co-pending applications for other developments we have made in the field. However, specifically as to the improvements, advantages and features described herein, reference will be made in the description which follows to the below-described drawings.

FIG. 1 shows a vector matrix calculation.

FIG. 2 illustrates a general matrix multiplication.

FIG. 3A shows a diagonal cell, a general cell, and a diagonal processor.

FIG. 3B shows a general processor.

FIG. 4 shows a preferred communicating ALU tree;

FIGS. 5A-5C illustrates a 4.sup.2 PE and 4-Root Tree Processor Massively Parallel Diagonal-Fold Tree Array Processor;

FIG. 6 illustrates a Processor Element tagged instruction/data format;

FIGS. 7A and 7B illustrates a processor element example instruction set.

Our detailed description follows as parts explaining our preferred embodiments of our invention provided by way of example.

A vector martix multiplication operation utilizing a sum of product calculation especially suited for our preferred MP organization is shown in FIG. 1 where there are i columns and j rows. The input matrix z.sub.i is defined as:

This is a subset of the general case of matrix multiplications. Consider the three N.times.N matrices, Y, W, and (Y*W) result matrix z, as shown in FIG. 2, with i columns and j rows and a notation of Y.sub.(column),(row). It is assumed that a MP Diagonal-Fold Tree Array Processor is available with the following assumed capabilities:

N Root Tree Processors: each possessing a Y value memory capacity of N Y values and an additional memory capacity for N result values.

N.sup.2 PEs with the W values to be stored in internal PE registers.

The Root Tree Processor complex issues all instructions in broadcast mode.

The algorithmic procedure to be described represents only one of many possibilities and is not necessarily the "best" procedure depending upon an application. It is meant to demonstrate the capability of our MP Diagonal-Fold Tree Array Processor. The basic procedure is to have N Root Tree Processors send a row of the Y matrix and a multiplication instruction, with the Auto mode specified, to the PEs which execute the multiplication and send the results to the CATs for summation, providing a row of the result matrix back to the Root Tree Processors for storage. The Root Tree Processors then read out a new row of the Y matrix and send it back to the PEs, continuing generating a row of the result matrix at the output of the CATs and storing the row in memory until all result rows have been calculated. The W value matrix, once initialized in the PEs, remains fixed, internal to the PEs throughout the matrix multiplication operations.

The TSNAP structure as described in --T-SNAP--required the HOST processor to provide update operations on PE stored data as required by a neural network model since T-SNAP--did not provide any PE update modification mechanisms. Additionally, T-SNAP does not provide the Root Tree Processor function capability provided by our present architecture. The Learning Machine structure as described in --Learning Machine--specifically addressed the capabilities available by a machine specifically tailored for neural network emulation as evidenced by the examples for Back-Propagation learning, and the high connectivity Hopfield neural networks. In order to accommodate "general" update rules on parallel stored data in the multiple PEs and provide "general" Root Tree Processor capabilities, major modifications to both T-SNAP and the Learning Machine are required. These modifications provide capabilities beyond those normally associated with the neural network paradigm. Instead of the fixed multiplication function provided in T-SNAP and limited instruction set architecture of the Learning Machine, a more general processor architecture is put in place with an expanded instruction set architecture and expanded data storage is allowed for. Both word parallel and bit serial operational modes available depending upon problem specific needs, as neither is precluded.

Internally, the T-SNAP structure utilized two types of "cell" structures, the Diagonal-Cells and the General-Cells for the direct emulation of the neural sum of products function and did not address the processing of locally stored data, for example as required by learning algorithms see--Rumelhart 86. The basic multiplier element structures, without programmability are repeated in FIG. 3A, and new "cell" processing structures, with the additional local data manipulation capabilities provided by an instruction set architecture, are shown in and FIGS. 3A and 3B. The term "element", used in the following discussion, refers to the minimum architected functional unit required TSNAP neural emulation capabilities, namely a weight register, a Y value register, and a multiplier symbolically indicated as W.sub.ij Y.sub.j. The register terminology of "Y" and W.sub.ij Y.sub.j is kept through out the general processor discussion with out limiting the register usage to the neural emulation paradigm. The first "cell", FIG. 3A, is associated with the diagonal elements, W.sub.ij .times.Y.sub.j, and the second "cell", General-Cell, FIG. 3A, is associated with the rest of the elements W.sub.ij Y.sub.j and contains two elements placed in the General-Cell, shown in a top and bottom arrangement. The new form of processor cell, in general referred to as the Processor Element (PE), consists of the Diagonal-PE and the General-PE. The modifications to the basic processing structure, FIG. 3A, are shown in FIGS. 3A and 3B and consist in the addition of a tag compare function, an optional instruction buffer, an instruction (INSTR) register, an expanded register file, Conditional-Execution Bits (CEB) in each data register, a data path register, an instruction path bit, selector and distributor control structures, and expanded functions, such as division, square root, etc. that may be application specific in addition to multiplication, as represented by the EXecution Unit (EXU) block. These modifications are required for multiple reasons. First, since a processing element receives both instructions and data from the same source path, namely from the cell attached communicating ALU tree (CAT), a method of differentiating instructions from data must be used. It is assumed that a fixed format is used for the instructions and data allowing a fixed single bit field be used to differentiate instructions from data. Second, by utilizing programmable data path and instruction path registers in conjunction with a front end decoding and distribution mechanism, the destination for a received instruction or data word can be specified. Third, a tag compare function allows further capability in specifying instruction and data destination points. Fourth, since an instruction may specify multiple functions, an instruction register and instruction decoding, operand selection, function selection, destination selection, and execution mechanisms must be provided. Fifth, to provide flexibility in algorithmic capability, a register file with Conditional Execution Bits are provided along with capability of manipulating the CEBs. The execution of an instruction, whose result destination is a local PE register, is conditional based upon the state of the destination register's CEB. The CEB indicates whether a register can be modified or not. The PE programmability is obtained through the instructions which are decoded at a PE's instruction register received from either an optional instruction buffer or with no instruction buffer from the attached communicating ALU tree that is in a communications mode. Each PE upon receipt of an instruction in the instruction register will execute the operation specified by that instruction. The instruction types include a data/instruction path specification, data movement, arithmetic, and logical instructions. Each PE contains an instruction register for each processing element specifying the source and destination paths and EXU functions; a Diagonal-PE contains one instruction register and the General-PE contains two instruction registers.

The modification to the T-SNAP cells must preserve the functional capabilities provided by the original cells, inorder to support neural emulation as well as other applications requiring similar capabilities. An essential, novel, and general purpose functional capability provided by the T-SNAP multiplier cells, that must be maintained in the new processor cell structure, concerns the emulation of completely connected processors, for example neuron processor as used for completely connected networks such as Hopfield 82 and Hopfield 84. This important function is briefly reviewed using the original T-SNAP cells FIG. 3A. For example, with a neural network model in an execution mode, implying a multiplication operation in each processing cell, the diagonal cell multiplies its stored weight with its stored Y value and supplies the multiplied result to the attached add tree. In the communications mode for the diagonal cells, a Y value is received from the attached add tree and stored into the Y value register. The "General-Cells" of the structure also generate a weight times Y value and supply the product to their attached add trees. In the communications mode for these "General-Cells", a Y.sub.j value received from the bottom multiplier add tree is stored into the top Y value register and likewise a Y.sub.i value received from the top multiplier add tree will be stored into the bottom Y value register. This switch in storing the Y values is an essential characteristic supporting complete connectivity. For the modified processing cells, FIGS. 3A and 3B, this path switch is programmable allowing further unique architectural features for processing, as will be described in the Processor Element Instruction Set section of this Chapter. To preserve the internal path switch function of the original T-SNAP cells, the new processor cells require that the data path registers be specified (loaded) in advance of receiving data from a tree. The data path register specifies the destination of the Yj data received from the bottom tree to be the top Yj register and the destination of the Yi data received from the top add tree to be the bottom Yi register thereby preserving the complete connectivity function.

The symbolic summation tree is shown on the left of FIG. 4 with ALUs at each stage designated by the letter A. The mole detailed representation of the communicating ALU tree structure that will be used is shown on the right-hand side of FIG. 4. Pipeline latches have been left out for more clarity. For specific applications, the ALU function might be as simple as a bit-serial adder or provide more complex programmable functions requiring an instruction set architecture. For the purposes of describing the function execution and communications operations a summation operation may be referred to in this text. The use of the summation function is for simplicity of explaination and not intended to imply a limit to the functionality the communicating ALU tree can provide. In addition, the tree nodes' control mechanism, that determines the nodes operational mode and function, can use separate control lines or tagged tree node instructions. For a single node function such as addition and two operational modes, namely communictions and function execution, a single control line implementation is feasible. If more extended functions are to be supported in a tree node, then not only would additional control mechanisms be required but storage elements may be required in a tree node. In addition, if multiple functions are provided in the tree nodes then a method of synchronistically controlling tree operations must be utilized. If varying function execution timings are to be allowed in each tree node then an asynchronous interfacing method must be provided between the tree stages. For simplicity of implementation that guarantees the synchronization control, a restriction could be enforced that the same operation be specified for each tree stage. In FIG. 4 three ALU elements are shown in a 2 stage pipelined tree arrangement. The ALU element has a SWitch 1, SW1, block on its output and two SWitch 2s, SW2, blocks bypassing the ALU. The communicating ALU tree can be placed into one of two modes, namely a function execution mode and a communications mode, also termed a bypass mode. A common control signal is used at each ALU element in order to guarantee that all nodes of the tree provide the same mode of operation. One of the functions specified by the tree control signal, an accompanying tag signal or common distributed signal, is the ALU bypass. Both switches, SW1 and SW2, have an on/off control which, when in the "off" state, keeps the switch open, i.e. in a high impedance state and when in the "on" state bypasses the ALU (node function) via a low impedance path. When SW1 is enabled SW2 is disabled and vice versa. In this manner the ALU tree can provide the summation function, for example, in one direction, SW1's on-SW2's off, while essentially acting as a communication path in ALU bypass mode, SW1's off-SW2's on. The ALU tree using 2 to 1 functional elements, such as 2-1 adders, will require log.sub.2 N stages. Alternatively, the ALU function and communication mode can be implemented with 3-1, 4-1, . . . , N-1 functional elements, such as 3-1, 4-1, . . . , N-1 adders, and their bypass switches, utilizing all the same element types or in combination, to produce the specified function. It should be noted that the Communicating ALU, FIG. 4, represents its logical function since, for example, depending upon technology, the SW1's function could be incorporated in the gate devices used in the last internal stage of each ALU element, thereby adding no additional delay to the ALU function. Alternatively, a separate communications tree path could be provided, thereby allowing communications to occur while an ALU function is in progress.

A 4 Root Tree Processor example is shown in FIGS. 5A-5C which connects the sixteen PEs with four CATs and four Root Tree Processors with a Host interface to provide a complete picture of the machine organization used in the Massively Parallel Diagonal-Fold Tree Array Processor. The CATs are assumed to provide a summation function in ALU execution mode. An example of the elements involved in a sum of "W*Y" register products calculation for the third Root Tree Processor RTP.sub.3 is written here and highlighted in FIGS. 5A-5C.

The Host interface represent a central control point for the array of PEs allowing the Host to have access to the Root Tree Processors' internal storage possibly containing, for example, the initial parameters W, Y, etc., calculated values, and traced values. There is assumed to be a Root Tree Processor for each communicating/function execution tree and their N attached PEs. Each Root Tree Processor issues instructions and data to the N tree attached PEs through the communications mode of tree operation. Additional functions the Root Tree Processor and Host interface include the following:

1. All processor initializations

2. Starting the system

3. Stoping the system

4. communicating ALU tree control

5. PE instruction and data issuing

In operation, the N.sup.2 PE structure might require an initialization of certain registers. Even though specific registers could be initialized by sending uniquely PE tagged values to individual processors, N.sup.2 operations would be required. An alternative scheme would be to connect the PE registers in a serial daisy chained fashion for LSSD scanning--see Eichelberger 77. LSSD scanning supports testing and register initialization. Each Root Tree Processor and its attached N PEs may have one or up to multiple scan strings depending upon an implementation technology. For example, a separate scan path containing only the "W" registers for each Root Tree Processor can be provided resulting in N "W" register scan paths which can be used for initialization purposes. Other initialization mechanisms are clearly possible and it will be assumed the appropriate method can be chosen during the implementation design process. Parameter values common to a Root Tree Processor's N attached PEs can be loaded through the communicating ALU tree.

An example instruction set providing the previously discussed capability will be reviewed in this section, beginning first with a presentation of an example format for the instructions and data and continuing with an example instruction set description.

A set of decisions must be made to determine an architecture for a processing element. One major set of decisions concerns the format of instructions and data--see Flynn 85 for a review of the universe of format options available and their impact to cost and performance) and the other major set of decisions concerns the functions to be executed by the architecture. Since each Root Tree Processor issues instructions and data to N Processors Elements, a desirable capability for enhanced programming flexibility would be to extend the instruction and data format to include a processor group identification field. Consequently, the instruction format has the requirements to identify processors or groups of processors, separate instructions from data, identify operands, and result destinations. Even though other formats are clearly possible, we choose the example format depicted in FIG. 6 on page 6. Both specific Processor Elements and groups of PEs can be identified with the addition of tag bits and a Broadcast, "B", bit added to all communications, and tag and "B" bit compare functions in each PE. A "B" bit is added as the first bit in the field indicating a broadcast instruction/data for B=1 and a tagged instruction/data for B=0. A broadcast message/data goes to all N Processor Elements linked to a particular Y.sub.j Root Tree Processor, independent of the TAG. The TAG field must be at least log.sub.2 N bits long, if specific identification of all Processor Elements belonging to a Root Tree Processor is to be accomplished. Alternatively, groups of Processor Elements can utilize the same tag value, thereby uniquely identifying groups of PEs. The received tag is bit by bit compared with a stored tag in each PE. After the last tag bit compare is completed, it is known whether the following INSTR/DATA is to be received by that particular PE. A tag match results in a INSTR or Data being received, while a no match situation prevents the reception of a INSTR or Data. A parity bit or error correction bits denoted by a P can also be included in the tag field, as shown in FIG. 6, for error handling reasons.

The communicated instructions or data also contain a single bit (INSTR) indicating whether the bit string is data or instruction, and for instructions additional fields for specifying an automatic execution mode (AUTO), instruction opcode (INSTR), operand selection (SOURCE1 and SOURCE2), and result destination (DESTINATION). Error correction/detection bit/s (ECC) can be included on both instructions and data for error handling reasons. It is assumed that the instruction and data bit lengths are the same. FIG. 7 lists the present instruction set functions.

The instruction set may contain, arithmetic operations, for example, add, subtract, multiply, divide, square root, etc., logical operations, for example, AND, OR, EX-OR, Invert, etc., Compare, shift, and data storage movement operations. The instruction set is primarily determined from an application specific perspective.

A fairly standard instruction format is used with the unique addition of the AUTO bit as representing an automatic execution mode. The auto execution mode represents a capability that switches the execution mode of the PEs from an instructions execution only mode to a data dependent mode of execution. The control of the switch from a control flow execution mode to a data flow execution mode is programmable by use of the AUTO bit to engage the data flow mode and a rule that allows the return to control flow instruction execution mode. An instruction with the AUTO bit active is executed first due to normal instruction control flow execution sequencing and then it is executed each time valid data is received in the processing unit. The data flow execution continues until a new instruction is received which stops the previous "AUTO" instruction from executing and begins the execution of the newly received instruction, which may also be another AUTO instruction.

To demonstrate the importance of the AUTO mode for processing, a simple example using the Hopfield neural network will be presented. For this discussion, instruction mnemonics, as presented in the example instruction set architecture of FIGS. 7A and 7B, are used. Assume the Hopfield neural network model--see Hopfield 84--is used as an example, for the direct emulation of the network neurons sum of connection weight times connecting neuron output values. Each network update cycle consists of weight times Y value multiplication operations, summation of multiplication results, the generation of the nonlinear sigmoid neuron output Y values, and the communication of the generated Y values to the processing elements. The network updates continue until a network minimum is reached. For simplicity of discussion, assume that network convergence is not tested for on every cycle, but only after some multiple cycles have been executed. For the network emulation using the Processor Elements, an automatic mode can be specified where, instead of requiring the repeated sending of a Multiply instruction to the PEs after each network execution cycle in order to initiate the next network cycle, the automatic mode would begin the next update cycle automatically after receipt of the newly calculated Y values. This automatic mode is initiated by setting the AUTO bit to a "1" in the instruction desired, such as Multiply (MPY) for use in the Hopfield network example, which sets an automatic mode flag in the PEs. The first operation is initiated with the receipt of the instruction with the AUTO bit set to a "1" and the instruction would be repeatedly executed upon receipt of the new Updated data continuing until a new instruction is received which terminates the automatic mode, such as receipt of a NOP instruction. A capital A is appended to an instruction mnemonic to indicate that the auto bit is to be set to a "1", for example MPYA.

The source and destination addresses specified in FIGS. 7A and 7B are relative to the instruction register where the instruction is received. The relative addressing is shown in FIG. 3B where the top instruction register INSTR TREG relative addresses are shown in columnar fashion, located to the right of the register blocks, while the relative addressing for the bottom instruction register INSTR BREG is shown in columnar fashion, located to the left of the register blocks. For "k" temporary or working registers, it should be noted for example, that the bottom instruction register R2 is the same as the top instruction register R(2+k+1). A bit string received from the ALU tree, if it is an instruction, is serialized into one of the two INSTR registers in each General-cell, as directed by the INSTR PATH BIT, and the single INSTR register of a Diagonal-Cell. A data bit string received from the ALU tree, is serialized to one of the k+4 other registers available in a General-cell and one of the k/2+2 other registers available in a Diagonal-Cell as specified by the DATA PATH register. It is assumed that for a symmetrical structure the Diagonal-Cells contain half the number of instruction and data registers as compared to the General-Cells. In the Diagonal-PEs a source or destination address of R(2+k/2+1) through R(2+k+2) and CR2 are mapped as follows:

R(2+k/2+1).fwdarw.R(2+k/2)

R(2+k/2+2).fwdarw.R(2+k/2-1)

continuing

R(2+k/2+k/2+2)=R(2+k+2).fwdarw.R(2+k/2-k/2-1)=R1

CR2.fwdarw.CR1

For example, assume a k=2 working registers in the General-Cells and three bit source or destination address then having the General-Cells use all three bits and the Diagonal-Cells use only the 2 Isb bits, the proper mapping can be provided by:

000.fwdarw.CR1

001.fwdarw.R1

010.fwdarw.R2

011.fwdarw.R3

100.fwdarw.CR2

101.fwdarw.R6

110.fwdarw.R5

111.fwdarw.R4

The PATH instruction is treated differently from the other instructions, since it controls the instruction path selection mechanism. The PATH instruction is decoded prior to the distributor logic, FIGS. 3A and 3B and the DATA PATH register or INSTR PATH BIT loaded according to destination field. A possible format for the PATH instruction destination field would be to use the first bit position for CR1/2 and the other bits for register path selection, others formats are clearly possible. The PATH instruction must be reissued if a different path is desired. A default path is specified by the architecture for initialization purposes, for example the DATA PATH registers could be initialized to R5, the Y value register to support a completely connected network, and the INSTR PATH BIT set to a "1" indicating the instructions switch path also. All PE data registers are (D=L+1) bits in length, due to the conditional execution bit in each data register, see FIG. 6 showing the instruction and data formats. If a CEB is set to a "zero" in an instruction's destination register, that instruction will be treated as a NOP instruction, ie. the destination register's contents will not be changed and "zeros" will be fed to the Add tree. If the CEB is set to a "one" the register's contents can be modified. For example, this bit is used on the W registers to identify the presence or absence of a value since a zero value is not always sufficient to accomplish this. "Zeros" are always fed to the Add tree if the Add tree is not a destination point or a NOP situation is occurring. The CEBs can be set upon initialization through the chips scanning facility, the path instruction, or in a data value received from the tree.

By allowing the instruction and data paths to be programmable, two operational modes for network emulation become possible. In the first mode, termed YIN mode, for all processors, the instruction received from the bottom Adder tree is sent to INSTR BREG (CR1) and the instruction received from the top Adder tree is sent to INSTR TREG (CR1). Then for YIN mode each PE will function as specified in the instruction field. In this fashion, each Root Tree Processor can be specified with a different input PE processing function, common across all PE inputs to that Root Tree Processor. For example, referring to FIGS. 5A-5C, all inputs to Root Tree Processor 4 may be specified with a (W register value*Y value) multiply operation while all inputs to Root Tree Processor 2 may be specified with an Temp. register value*Y value operation. Since all communicating ALU trees are independent of each other, each PE input function can have different execution times in YIN mode for different Root Tree Processors implying that the summation results would occur at different times, which can cause synchronization problems in the interface to the Root Tree Processors, if not handled correctly. YIN mode can be used to asynchronously select a Root Tree Processor and its set of PEs for processing. In the second mode, termed YOUT mode, for all Root Tree Processors, the instruction received from the bottom Adder tree is sent to INSTR TREG (CR2) and the instruction received from the top Adder tree is sent to INSTR BREG (CR2). Consequently, for YOUT mode all Root Tree Processor value outputs will have the same function applied at their input destination PE. In this way each Root Tree Processor can have multiple functions at its input PEs. For example, referring to FIGS. 5A-5C, all Y4 destination PEs may be specified with a (W register value*Y value) multiply operation while all Y2 destination PEs can be specified with a (Temp. register value*Y value) operation. All functions specified at a PE input must execute in the same length of time even though the functions are different. In general, YIN mode and YOUT mode can not be interchanged among the Root Tree Processors within a single model or problem structure as conflicts could result. For the simple Hopfield network emulation example the Root Tree Processors functioning as neurons specified the YOUT mode with all Root Tree Processors issuing the same instruction to all PEs.

Many instructions specify a destination which is local to the individual Processor Element. This local processing can cause synchronization problems if not handled correctly. Instead of proliferating synchronization mechanisms throughout the structure the local processing synchronization problem can be localized to the Root Tree Processors. For example, if no notification of local processing completion is generated from the PEs, a fixed hardware mechanism can be provided at the Root Tree Processor to guarantee safeness of the operations. It is also not desirable to "solve" the problem via means of queues in the Processor Elements as this increases the size of the PE limiting the number which could be placed on a single chip. Rather, the instruction issuing point should be used to resolve and avoid all hazards. Any local processing instruction to the same PE must be separated from the next instruction to that same PE by the specified processor instruction's execution time. For example, if the multiply executed in 2L clocks, a 2L time out must be ensured prior to sending the next instruction. This is necessary so that an instruction buffer register is not required, thereby allowing each instruction to remain constant in a PE during the operation of the function instructioned. Each Root Tree Processor can then be set up with a synchronization mechanism to safely issue instructions to each PE--at a maximum rate. Non-local instructions, i.e. those instructions where the destination is the ADD TREE, provide notification of operation completion when the converged tree result reaches the Root Tree Processors. For non-local instructions the Root Tree Processors wait until a result is received before sending a new instruction to the PEs attached to that tree.

As a final note, a compiler would be required to ensure no destination conflicts occur in programs using the described instruction set.

The following detailed procedure will be followed, see PE example instruction sets: (Note that the PE instructions are indicated by PE-Instruction Mneumonic and don't care states indicated by (x).)

1. CATS placed into communication mode

2. Initialize W matrix into PE registers

3. Each Root Tree Processor memory is initialized as follows:

Root Tree Processor 1 initialized with Y11, Y12, . . . , Y1N.

Root Tree Processor 2 initialized with Y21, Y22, . . . , Y2N.

Root Tree Processor N initialized with YN1, YN2, . . . , YNN.

4. Initialize the Root Tree Processor and PE PATH registers:

Set PE INSTR PATH Bit to CR2 indicating YOUT mode

Set PE DATA PATH to R2

5. All Root Tree Processors are active

6. Root Tree Processors send the first row of Y values to the PEs.

7. Root Tree Processors send PE-MPYA R1*R2.fwdarw.ADD TREE, after sending the instruction to the PEs, the Root Tree Processors place the CATs into the summation mode.

8. Root Tree Processors receive the summation results from the CAT roots.

9. Root Tree Processors send the second row of Y values to the PEs.

10. While the PEs and CATs are calculating the next row of the result matrix, the Root Tree Processors can store the first row of the result matrix, in this example to the additional storage capacity assumed present.

11. Since the Auto mode was specified in the PE-MPYA instruction, the PEs upon receipt of the second row Y values will automatically execute a PE-MPY R1*R2.fwdarw.ADD TREE with CAT results to be received in the Root Tree Processors.

12. The Root Tree Processors send the third row of Y values and stores the second row of the result matrix. The process continues until

13. Root Tree Processors store the last row of the result matrix.

At completion of operation, the original Y and W matrices are intact and the result matrix is located in the Root Tree Processors' additional storage area which can then be further operated upon by the Root Tree Processors or Host system.

Matrix addition and Boolean operations can also be supported by the structure. Assuming matrices of the same form as given in FIG. 2, both Y and W matrices can be loaded into the PE array since there are N.sup.2 unique Y and W registers in the structure. Local addition or Boolean operations on the Y and W registers can be done within the structure with the result sent to the temporary registers. At completion of the operation, the original Y and W matrices will remain intact in the structure and the temp regs will contain the result matrix. The result can be scanned out or individually read out from the Processor Element cells or used for further operations (chaining or linking of instructions).

While we have described our preferred embodiments of our invention it will be understood that those skilled in the art, both now and in the future, upon the understanding of these discussions will make various improvements and enhancements thereto which fall within the scope of the claims which follow. These claims should be construed to maintain the proper protection for the invention first disclosed.