Abstract: A manifold array topology includes processing elements, nodes, memories or the like arranged in clusters. Clusters are connected by cluster switch arrangements which advantageously allow changes of organization without physical rearrangement of processing elements. A significant reduction in the typical number of interconnections for preexisting arrays is also achieved. Fast, efficient and cost effective processing and communication result with the added benefit of ready scalability.

Abstract: Processing element to processing element switch connection control is described using a receive model that precludes communication hazards from occurring in a synchronous MIMD mode of operation. Such control allows different communication topologies and various processing effects such as an array transpose, hypercomplement or the like to be efficiently achieved utilizing architectures, such as the manifold array processing architecture. An encoded instruction method reduces the amount of state information and setup burden on the programmer taking advantage of the recognition that the majority of algorithms will use only a small fraction of all possible mux settings available. Thus, by means of transforming the PE identification based upon a communication path specified by a PE communication instruction an efficient switch control mechanism can be used.

Abstract: A SIMD machine employing a plurality of parallel processor (PEs) in which communications hazards are eliminated in an efficient manner. An indirect Very Long Instruction Word instruction memory (VIM) is employed along with execute and delimiter instructions. A masking mechanism may be employed to control which PEs have their VIMs loaded. Further, a receive model of operation is preferably employed. In one aspect, each PE operates to control a switch that selects from which PE it receives. The present invention addresses a better machine organization for execution of parallel algorithms that reduces hardware cost and complexity while maintaining the best characteristics of both SIMD and MIMD machines and minimizing communication latency. This invention brings a level of MIMD computational autonomy to SIMD indirect Very Long Instruction Word (iVLIW) processing elements while maintaining the single thread of control used in the SIMD machine organization.

Abstract: A multi-processor array organization is dynamically configured by the inclusion of a configuration topology field in instructions broadcast to the processors in the array. Each of the processors in the array is capable of performing a customized data selection and storage, instruction execution, and result destination selection, by uniquely interpreting a broadcast instruction by using the identity of the processor executing the instruction. In this manner, processing elements in a large multi-processing array can be dynamically reconfigured and have their operations customized for each processor using broadcast instructions.

Abstract: A hierarchical instruction set architecture (ISA) provides pluggable instruction set capability and support of array processors. The term pluggable is from the programmer's viewpoint and relates to groups of instructions that can easily be added to a processor architecture for code density and performance enhancements. One specific aspect addressed herein is the unique compacted instruction set which allows the programmer the ability to dynamically create a set of compacted instructions on a task by task basis for the primary purpose of improving control and parallel code density. These compacted instructions are parallelizable in that they are not specifically restricted to control code application but can be executed in the processing elements (PEs) in an array processor. The ManArray family of processors is designed for this dynamic compacted instruction set capability and also supports a scalable array of from one to N PEs.

Abstract: An array processor includes processing elements arranged in clusters which are, in turn, combined in a rectangular array. Each cluster is formed of processing elements which preferably communicate with the processing elements of at least two other clusters. Additionally each inter-cluster communication path is mutually exclusive, that is, each path carries either north and west, south and east, north and east, or south and west communications. Due to the mutual exclusivity of the data paths, communications between the processing elements of each cluster may be combined in a single inter-cluster path. That is, communications from a cluster which communicates to the north and east with another cluster may be combined in one path, thus eliminating half the wiring required for the path. Additionally, the length of the longest communication path is not directly determined by the overall dimension of the array, as it is in conventional torus arrays.

Abstract: An array processor topology reconfiguration system and method enables processor elements in an array to dynamically reconfigure their mutual interconnection for the exchange of arithmetic results between the processors. Each processor element includes an interconnection switch which is controlled by an instruction decoder in the processor. Instructions are broadcast to all of the processors in the array. The instructions are uniquely interpreted at each respective processor in the array, depending upon the processor identity. The interpretation of the commonly broadcast instruction is uniquely performed at each processor by combining the processor identity for the executing processor, with a value in the instruction. The resulting control signals from the instruction decoder to the interconnection switch, provides for a customized linkage between the executing processor and other processors in the array.

Abstract: A plurality of processor elements (PEs) are connected in a duster by a common instruction bus to a sequencing control unit with its associated instruction memory. Each PE has data buses connected to at least its four nearest PE neighbors, referred to as its North, South, East and West PE neighbors. Each PE also has a general purpose register file containing several operand registers. A common instruction is fetched from the instruction memory by the sequencing control unit and broadcast over the instruction bus to each PE in the cluster. The instruction includes an upcode value that controls the arithmetic or logical operation performed by an execution unit in the PE on one or more operands in the register file. A switch is included in each PE to interconnect it with a first PE neighbor as the destination to which the result from the execution unit is sent.

Abstract: A parallel processing system and method is disclosed, which provides an improved instruction distribution mechanism for a parallel processing array. The invention broadcasts a basic instruction to each of a plurality of processor elements. Each processor element decodes the same instruction by combining it with a unique offset value stored in each respective processor element, to produce a derived instruction that is unique to the processor element. A first type of basic instruction results in the processor element performing a logical or control operation. A second type of basic instruction results in the generation of a pointer address. The pointer address has a unique address value because it results from combining the basic instruction with the unique offset value stored at the processor element. The pointer address is used to access an alternative instruction from an alternative instruction storage, for execution in the processor element.

Abstract: A neural array processor computing structure with a plurality of processing units which have orthogonal sets of neurons and communicating adder trees which provide a reverse communication path. Each processing unit has a weight storage unit. The processing units are arranged so that they are connected to the weight storage units in symmetric pairs. General processing cells share communication paths so that a reduced number of communicating adder trees is used.

Abstract: Image processing for multimedia workstations is a computationally intensive task requiring special purpose hardware to meet the high speed requirements associated with the task. One type of specialized hardware that meets the computation high speed requirements is the mesh connected computer. Such a computer becomes a massively parallel machine when an array of computers interconnected by a network are replicated in a machine. The nearest neighbor mesh computer consists of an N.times.N square array of Processor Elements(PEs) where each PE is connected to the North, South, East and West PEs only. Assuming a single wire interface between PEs, there are a total of 2N.sup.2 wires in the mesh structure.

Abstract: A Neural synapse processor apparatus having a neuron architecture for the synapse processing elements of the apparatus. The apparatus which we prefer will have a N neuron structure having synapse processing units that contain instruction and data storage units, receive instructions and data, and execute instructions. The N neuron structure should contain communicating adder trees, neuron activation function units, and an arrangement for communicating both instructions, data, and the outputs of neuron activation function units back to the input synapse processing units by means of the communicating adder trees. The apparatus can be structured as a bit-serial or word parallel system. The preferred structure contains N.sup.2 synapse processing units, each associated with a connection weight in the N neural network to be emulated, placed in the form of a N by N matrix that has been folded along the diagonal and made up of diagonal cells and general cells.

Abstract: Image processing for multimedia workstations is a computationally intensive task requiring special purpose hardware to meet the high speed requirements associated with the task. One type of specialized hardware that meets the computation high speed requirements is the mesh connected computer. Such a computer becomes a massively parallel machine when an array of computers interconnected by a network are replicated in a machine. The nearest neighbor mesh computer consists of an N.times.N square array of Processor Elements(PEs) where each PE is connected to the North, South, East and West PEs only. Assuming a single wire interface between PEs, there are a total of 2N.sup.2 wires in the mesh structure.

Abstract: A folded memory array structure that can be used with a folded processor array structure to make cosine transform calculations faster and more efficient. The memory elements of the folded memory array are interconnected is if an M by M element square array of single memory elements were logically folded or cut and overlain to form an N by N node subarray where each node has multiple interconnected memory elements. The subarray is than logically folded and the nodes connected by buses and switches as if the resultant array were folded along the diagonal, producing a triangular memory array structure. This transposed folded memory array is enhanced with computational elements disposed at the nodes of the array. A computational array processor is thereby formed which can perform cosine transform operations efficiently for multimedia, video, and other applications.

Abstract: A triangular scalable neural array processor unit for use in a neural network has an array of weight registers, multipliers, communicating adder trees, sigmoid generators, and a reverse feedback loop for communicating the output of a sigmoid generator back to input multipliers of selected neurons. The communicating adder trees provide the selectable feedback path.

Abstract: A Neural synapse processor apparatus having a neuron architecture for the synapse processing elements of the apparatus. The apparatus which we prefer will have a N neuron structure having synapse processing units that contain instruction and data storage units, receive instructions and data, and execute instructions. The N neuron structure should contain communicating adder trees, neuron activation function units, and an arrangement for communicating both instructions, data, and the outputs of neuron activation function units back to the input synapse processing units by means of the communicating adder trees. The apparatus can be structured as a bit-serial or word parallel system. The preferred structure contains N.sup.2 synapse processing units, each associated with a connection weight in the N neural network to be emulated, placed in the form of a N by N matrix that has been folded along the diagonal and made up of diagonal cells and general cells.

Abstract: A triangular scalable neural array processor unit for use in a neural network and having multipliers, communicating adder trees, sigmoid generators, and a reverse feedback loop for communicating the output of a sigmoid generator back to input multipliers of selected neurons.

Abstract: A Neural synapse processor apparatus having a neuron architecture for the synapse processing elements of the apparatus. The apparatus which we prefer will have a N neuron structure having synapse processing units that contain instruction and data storage units, receive instructions and data, and execute instructions. The N neuron structure should contain communicating adder trees, neuron activation function units, and an arrangement for communicating both instructions, data, and the outputs of neuron activation function units back to the input synapse processing units by means of the communicating adder trees. The apparatus can be structured as a bit-serial or word parallel system. The preferred structure contains N.sup.2 synapse processing units, each associated with a connection weight in the N neural network to be emulated, placed in the form of a N by N matrix that has been folded along the diagonal and made up of diagonal cells and general cells.

Abstract: A neural network architecture consisting of input weight multiplications, product summation, neural state calculations, and complete connectivity among the neuron processing elements. Neural networks are modelled using a sequential pipelined neurocomputer producing high performance with minimum hardware by sequentially processing each neuron in the completely connected network model. An N neuron network is implemented using multipliers, a pipelined adder tree structure, and activation functions. The activation functions are provided by using one activation function module and sequentially passing the N input product summations sequentially through it. One bus provides N.times.N communications by sequentially providing N neuron values to the multiplier registers. The neuron values are ensured of reaching corresponding multipliers through a tag compare function. The neuron information includes a source tag and a valid signal.

Abstract: A scalable flow virtual learning neurocomputer system and appratus with a scalable hybrid control flow/data flow employing a group partitioning algorithm, and a scalable virtual learning architecture, synapse processor architecture mapping, inner square folding and array separation, with capability of back propagation for virtual learning. The group partitioning algorithm creates a common building block of synapse processors containing their own external memory. The processor groups are used to create a general purpose virtual learning machine which maintains complete connectivity with high performance. The synapse processor group allows a system to be scalable in virtual size and direct execution capabilities. Internal to the processor group, the synapse processors are designed as a hybrid control flow/data flow architecture with external memory access and reduced synchronization problems.