US 4,860,201 AGrant
Binary Tree Parallel Processor
Issue Date:1989-08-22
•9 Claims
•15 Drawing Sheets
Abstract
A plurality of parallel processing elements are connected in a binary tree configuration, with each processing element except those in the highest and lowest levels being in communication with a single parent processing element as well as first and second (or left and right) child processing elements. Each processing element comprises a processor, a read/write or random access memory, and an input/output (I/O) device. The I/O device provides interfacing between each processing element and its parent and children processing elements so as to provide significant improvements in propagation speeds through the binary tree. The I/O device allows the presently preferred embodiment of the invention to be clocked at 12 megahertz, producing in the case of a tree of 1023 processors, each having an average instruction cycle time of 1.8 .mu.s, a system with a raw computational throughput of approximately 570 million instructions per second. The I/O device communicates data and queries from the root processing element to all other N processing elements in the array in one processor instruction cycle instead of in O(log.sub.2 N) processor instruction cycles as in prior art binary tree arrays. Primitive queries are executed in parallel by each processing element and the results made available for reporting back to the root processing element. In several important cases, these results can be combined and reported back to the root processing element in a single processor instruction cycle instead of in O(log.sub.2 N) processor instruction cycles as in prior art binary tree arrays. Thus, the elapsed time for a broadcast and report operation is in effect a constant time regardless of the number of processors in the array.
Metadata
Assignee
- The Trustees of Columbia University in the City of New York
Inventors
- Salvatore J. Stolfo
- Daniel P. Miranker
Application Information
Application Number:US 9025472
Filing Date:1986-09-02
Priority Date:1986-09-02
Art Unit:237
Classifications
IPC:
G06F 1300
Field of Search:
364200 MS File;900 MS File
Patent Drawings (15 sheets)
Description
CROSS-REFERENCE TO RELATED APPLICATION
A related application is "Parallel Processing Method", U.S. Ser. No. 903,031, filed concurrently herewith, which is incorporated herein by reference.
BACKGROUND OF THE INVENTION
This invention relates generally to data processing and more particularly to a binary tree-structured parallel processing machine employing a large number of processors, each such processor incorporating its own I/O device.
Throughout the history of the computer there has been a continuing demand to increase the throughput of the computer. Most of these efforts have concentrated on increasing the speed of operation of the computer so that it is able to process more instructions per unit time In serial computers, however, these efforts have in certain senses been self-defeating since all the processing is performed by a single element of the computer leaving most of the resources of the computer idle at any time.
In an effort to avoid some of these problems, special purpose machines such as array processors have been developed which are especially designed for the solution of special classes of problems. Unfortunately, while commercially successful in the performance of certain computational tasks, such computers fall far short of adequate performance in others.
In recent years substantial efforts have been made to increase throughput by operating a plurality of processors in parallel. See, for example, Chuan-lin Wu and Tse-yun Feng, Interconnection Networks for Parallel and Distributed Processing (IEEE 1984). One such parallel processor is that in which a plurality of processors are connected in a tree-structured network, typically a binary tree. S. A. Browning, "Computations on a Tree of Processors," Proc. VLSI Conf., California Institute of Technology, Pasadena, Jan. 22-24, 1979; A. M. Despain et al., "The Computer as a Component", (1979 unpublished); A. Mago, "A Cellular Language-directed Computer Architecture," Proc. VLSI Conf., California Institute of Technology, Jan. 22-24, 1979; R. J. Swan et al., "Cm*-A Modular, Multimicroprocessor", Proc. 1977 NCC, pp. 645-655 (June 1977); J. R. Goodman et al., "Hypertree: A Multiprocessor Interconnection Topology", IEEE Trans. on Computers, Vol. C-30, No. 12, pp. 923-933 (December 1981), reprinted in Wu and Feng at pp. 46-56; J. L. Bentley and H. T. Kung, "Two Papers on a Tree-Structured Parallel Computer", Technical Report, Dept. of Computer Science, Carnegie-Mellon University, Sept. 1979.
In a binary tree computer, a large number of processors are connected so that each processor except those at the root and leaves of the tree has a single parent processor and two children processors. The processors typically operate synchronously on data flowing to them from the parent processor and pass results to descendant processors.
Important problems in the storage and retrieval of data can be analyzed following J. L. Bentley, "Decomposable Searching Problems", Information Processing Letters, Vol. 8, No. 5, pp. 244-250 (June 1978). Bentley defines a static searching problem as one of preprocessing a set F of N objects into an internal data structure D and answering queries about the set F by analyzing the data structure D. Bentley defines three functions of N that characterize the complexity of the searching function: the amount of storage S required by D to store the N objects, the preprocessing time P required to form D in S, and the time Q required to answer a query by searching D.
An illustration of a problem that can be solved by such a database is the membership problem. In this case N elements of a totally ordered set F are preprocessed so that queries of the form "Is x in F?" can be answered quickly. The common solution for serial computers is to store F in a sorted array D and binary search. Thus, the membership problem can be computed on sequential computers with the following complexity: S=N; P=O(N log N); Q=O(log N).
Bentley defines a decomposable searching problem as one in which a query asking the relationship of a new object x to a set of objects F can be written as:
where B is the repeated application of a commutative, associative binary operator b that has an identity and q is a primitive query applied between the new object x and each element f of F. Hence the membership problem is a decomposable searching problem when cast in the form:
where OR is the logical function OR and equal is the primitive query "Is x equal to f?" applied between the object x and each element f of F.
The key idea about this type of problem is its decomposability. To answer a query about F, we can combine the answers of the query applied to arbitrary subsets of F.
This type of problem is well suited to quick execution in a parallel processing environment. The set F is partitioned into a number of arbitrary subsets equal to the number of available processors. The primitive query q is then applied in parallel at each processor between the unknown x that is communicated to all processors and the locally stored set element f. The results are then combined in parallel by log.sub.2 N repetitions of the operator b, first performing b computations on N/2 adjacent pairs of processors, the b computations on N/4 pairs of results of the first set of computations and so on until a single result is obtained.
The complexity of this operation in the parallel processing environment is computed as follows. Each of the N elements of the set F must be distributed among the processors so that the N elements are distributed among all the processors. The number of time steps to do this equals the number of elements of the set. Thus, P=O(N). If each element is stored in a different processor such that S=N, the time required to answer a primitive query is a single time step; and the time required to compute the final answer is the number of time steps required to report back through the binary tree which is O(log.sub.2 N) Thus, Q=O(1)+O(log.sub.2 N). Compared with the complexity of the membership problem when executed on a serial computer, the use of a parallel processor provides substantial savings in the preprocessing time required to build the data structure since there is no need to store the data structure in an ordered array.
Bentley and Kung proposed a specific tree structure illustrated in FIG. 1 which was designed to achieve throughputs on the order described above. As shown in FIG. 1, their tree structure comprises an array of processors P1-P10 organized into two binary trees that share leaf processors P4-P7. Data flows in one binary tree from root processor P1 to leaf processors P4-P7. Data is operated on at the leaf processors and the results flow in the second tree from leaf processors P4-P7 to root processor P10. Obviously, many more processors can be used in the array if desired.
To load data into each of leaf processors P4-P7 of FIG. 1, the data for each leaf processor is provided to the root processor P1 at successive time steps and is routed through the array to each leaf processor via intermediate processors P2 and P3. Thus, it takes at least one time step per leaf processor to load data into the leaf processors.
The data structure is queried by providing the query to root processor P1 and propagating the query in parallel to each leaf processor. The results of the query are then reported out through processors P8 and P9 to root processor P10 with each of these processors computing a result from two inputs higher up in the tree. As will be apparent, propagation times of the query and the result through the binary tree introduce significant delays in overall throughput comparable to those of a serial computer.
While the time required to answer a single query in the parallel processor is comparable to that in a serial computer, queries can sometimes be processed in pipeline fashion in the parallel processor while they cannot be in the serial computer. Thus, after O(log.sub.2 N) steps, results begin to flow out of the parallel processor at a rate of one per time step. If the number of queries is large enough that the pipe filling and flushing times can be considered negligible, the complexity can be computed as: S=N, P=O(N) and Q=O(1).
There are, however, numerous instances in which pipelining cannot be used to minimize the effect of propagation delays in the binary tree. These include:
1. Decomposable searching problems where the number of "external" queries is small, or a series of queries are generated internally from processing elements within the tree. Internally generated queries would need to migrate to the root in logN steps under Bentley and Kung's scheme, and be "broadcast" down once again in logN steps. Since each query would force pipe flushing, Q=O(log N) for all queries. Artificial Intelligence production systems provide an illustration of these kinds of problems.
2. Searching problems where a single data structure D cannot be constructed. That is, for certain sets of complex (or multi-dimensional) objects, searching problems cannot be applied to a single data structure D. Consider relational databases where each element of a set is a complex record structure with possibly numerous fields or keys. To search such records, a data structure D would necessarily be needed for each field. Hence, in this case P(N)=kN for k fields in each record.
3. A set F of first order predicate logic literals, i.e., a knowledge base. We wish to compute a series of unifications of a series of "goal literals" against each of the elements of F. Since logic variables can bind to arbitrary first order terms during unification, a single query can change the entire set of elements in the knowledge base by binding and substituting new values for variable terms. Successive queries, therefore, are processed against succeedingly different sets of literals. (The simpler case involves frequent modifications to a dynamic set F, necessitating frequent reconstruction of D. Relational databases provide a convenient example.) Hence,
where F.sub.i =function (F.sub.i-1,Query (x.sub.i-1, F.sub.i-1)). (In the case of logic programming, function is substitution after logical unification, while for relational databases function may be insert or delete.)
4. Problems where a single query in a series of queries cannot be computed without knowing the result of the previous query. In dynamic programming approaches to statistical pattern matching tasks, a single match of an unknown against the set of reference templates cannot be computed without knowing the best match(es) of the previous unknown(s). Hence, for a series of unknown x.sub.i, i=1, . . . , M,
In this case, the same pipe flushing phenomenon appears as in the first case noted above. 5. Searching problems where we wish to compute a number of different queries about the same unknown x over a set, or possibly different sets. Hence,
Artificial intelligence production systems provide an illustration of this type of problem as well.
We will refer to problems of this type as almost decomposable searching problems.
Additional deficiencies of binary tree type parallel processors include efficiency and fault tolerance. Efficiency of a computation performed by the tree is often reduced since the amount of computation time required by each processor on a particular cycle may be vastly different depending on its local state. Such differences often result in unnecessary waiting and increased computation time. Additionally, it is well known that binary tree processors are inherently not very fault tolerant. Since any such fault has a tendency to ripple through the binary tree architecture, it is imperative that the fault(s) be not only detected but also compensated for so as to produce an accurate computation despite the fault(s).
SUMMARY OF THE INVENTION
The present invention comprises a plurality of parallel processing elements connected in a binary tree configuration, with each processing element except those in the highest and lowest levels being in communication with a single parent processing element as well as first and second (or left and right) child processing elements. Illustratively, 1023 processing elements are arranged in the tree in ten levels.
Each processing element comprises a processor, a read/write or random access memory, and an input/output (I/O) device. The I/O device provides interfacing between each processing element and its parent and children processing elements so as to provide significant improvements in propagation speeds through the binary tree. The I/O device allows the presently preferred embodiment of the invention to be clocked at 12 megahertz. The average processor instruction cycle time is 1.8 microseconds, producing in the case of a tree of 1023 processors a system with a raw computational throughput of approximately 570 million instructions per second, very little of which is required for communication overhead.
To minimize propagation delays in the binary tree computer of the present invention, the I/O device communicates data and queries from the root processing element to all other N processing elements in the array in one processor instruction cycle instead of in O(log.sub.2 N) processor instruction cycles as in prior art binary tree arrays.
Primitive queries are executed in parallel by each processing element and the results made available for reporting back to the root processing element. In several important cases, these results can be combined and reported back to the root processing element in a single processor instruction cycle instead of in O(log.sub.2 N) processor instruction cycles as in prior art binary tree arrays. This mode of operation is called Broadcast/Match/Resolve/Report.
The result is that the elapsed time for a broadcast and report operation with the apparatus of the present invention is in effect a constant time regardless of the number of processors in the array. In the present embodiment of the invention, this time is approximately 1.5 microseconds/byte. As a result, each of the above noted almost decomposable searching problems is efficiently run by the present invention. Internally generated queries, as in case 1, can be reported and broadcast in a constant number of machine cycles. Similarly, case 4 is efficiently handled by the same constant time broadcast/report cycle. In case 2, multiple data structures are not needed for the processor array of the present invention. Queries about different fields within a record can be implemented by broadcast of the field location prior to the execution of the query. In case 3, the successively different sets are computed on the fly and left intact in each processing element while successive queries are continuously broadcast, each in a constant number of machine cycles.
Another key capability of the present invention is to provide direct hardware support for quickly computing a range of commutative and associative binary operators b. In the membership problem defined above, the binary operator OR is repeatedly applied to all of the results of the primitive query: equal (x,f). In a sequential environment, this operation may require linear time to compute. In a parallel environment it can be computed in log time. With the present invention, it is computed in constant time.
The I/O circuit of the present invention also provides a high speed function resolve and report function that determines a most favorable value of a set of values stored in the processing elements and reports this value. Illustratively, this value is a minimum value and the function is called min-resolve. Min-resolve calculates in hardware in one instruction cycle the minimum value of a set of values distributed one to a processing element. Not only is the minimum value reported, but the processing element with the minimum value is set to a "winner state," providing an indication to all processing elements of the losers and the single winner in the computation, which can subsequently be reported in constant time. (Ties are arbitrated in hardware according to a fixed processing element ordering scheme.) The membership problem can thus be computed by applying min-resolve to zeros and ones (distributed throughout the tree after the application of the complemented equality operator) to compute OR. Similarly data generated within the processor array can be read in sorted order in O(N) time by sequential enumeration of min-resolved winners.
For example, in the case of the membership problem, the min-resolve function is implemented as follows:
1. Preprocess N objects of set F by distributing each element in turn to one processing element of the array. Note S=N for the entire machine and P=N.
2. Broadcast the unknown object x to all processing elements in constant time.
3. Apply the query equal (x,f)? in parallel at each processing element. In parallel, each processing element sets the min-resolve value to 0 if x=f, else sets to 1.
4. Min-resolve in one instruction cycle.
The overall computation time is Q=O(1), the sum of steps 2, 3 and 4. In cases where the primitive query q is more complex than equality, the running time is proportional to the slowest of the entire set of concurrent executions of q.
Thus, if four data elements are to be stored in the processor array of the present invention, it will take four instruction cycles to store such data elements in four processing elements. At any time thereafter a query may be processed by presenting the query to the root processing element. Within one instruction cycle, the query is propagated to each processing element, is tested against the information stored at each processing element and an answer is reported back.
Other illustrations of problems which are well suited to solution on the apparatus of the present invention are artificial intelligence production systems. Examples of such production systems are described in Salvatore J. Stolfo, "Five Parallel Algorithms for Production System Execution on the DADO Machine", Proc. of National Conference on Artificial Intelligence, AAAI, University of Texas at Austin, pp. 300-307 (Aug. 6-10, 1984).
Logic programming, relational database operations and statistical pattern recognition can be similarly defined as almost decomposable searching problems. Each of these problems shares the same common programming paradigm on the processor array of the present invention:
Distribution of an initial set of data.
Broadcast of data in constant time.
Local execution of some primitive query against a prestored element of the reference set (unification of first order logic literals, relational selection of tuples, dynamic time warp match of template of a speech utterance, respectively).
Resolution to find the best answer in constant time.
Report of the final answer.
From the foregoing, several key principles of the present invention can be identified:
The form of the data manipulated by the apparatus of the present invention can be of any type or size and hence can process the same wide range of data that any conventional computer system in existence today can manipulate.
The primitive query q can be any computation that is programmable on any computer in existence today. Hence, the apparatus of the present invention is not limited to any particular class of applications.
The inventive apparatus can quickly execute arbitrary storage and retrieval functions definable as an almost decomposable searching problem for any application area definable including AI, signal processing, database processing, text processing, numeric processing, etc.
For example, the primitive query q can be more complex than simple equality testing. In the speech recognition area we may wish to define a primitive query q which calculates the degree of match between the unknown and each of the reference set elements (set F in our example). In this case q can be a function which calculates 0 if the unknown exactly matches the set element, infinity (the largest number representable in the computer's memory) if there are no similarities at all, and some integer representing a distance metric, i.e., roughly how close the unknown matches the set element. A variety of programmable techniques can be employed in defining this function. The present invention supports them all since each processing element is a programmable computer.
Thus, in the case of a computer match of names a suitably defined primitive query q might rate the partial match of "Smith" to "Smythe" as closer (to 0) than would the evaluation of the match of "Jones", "Mason" and "Hamershold". If we name q in this context as distance, then to determine whether "Smith" is an element of our set, or whether "Smith" closely matches some element of F, we define
The key point is that the principles set forth above for fast execution of almost decomposable searching problems are generic and hence applicable to a wide range of problem areas--indeed, to all problem areas. In this brief example, the distance function can pertain to partial matching of character strings, Manhattan distances of a set of points in the plane, partial matches of AI symbolic expressions (definable for example in LISP expressions, or First Order Logic expressions or PROLOG expressions), or a set of vectors representing measured parameters of some signal, as for example LPC coefficients of a speech or acoustic waveform. The present invention supports all of these data types and others which may be defined, represented and stored on any computer system.
Accordingly, it is a principle object of the present invention to provide new and improved parallel processing systems.
Another object of the invention is to provide a plurality of processing elements arranged in a binary tree configuration wherein each processing element has associated with it its own I/O device capable of communication with that processing element as well as other processing elements.
A further object of the invention is to provide a parallel processor arranged in a binary tree configuration having improved efficiency and fault tolerance capabilities.
A still further object of the invention is to provide a parallel processor which essentially guarantees proper operation of a binary tree machine after two successive faults, as well as a 50% chance of proper operation after a third successive fault.
Another object of the invention is to provide a system which improves the efficiency of parallel processing machines by reducing the computation time deviation of each processor from the mean computation time.
BRIEF DESCRIPTION OF THE DRAWINGS
These and other objects, features and advantages of the invention will become more readily apparent with reference to the following description of the invention in which:
FIG. 1 is a schematic illustration of a prior art binary tree structure;
FIG. 2 illustrates the general configuration of a preferred embodiment of the invention comprising a binary tree of 1023 processing elements;
FIG. 3 depicts, in block diagram form, a single processing element comprising an 8751 processor, an I/O device, and a random access memory;
FIGS. 4A and 4B depict a detailed schematic of the single processing element of FIG. 3;
FIG. 5 is a functional flow chart of the underlying model for a single computation cycle;
FIGS. 6A and 6B are a schematic diagram of the I/O device comprising broadcast, report-resolve, debug, parity check, instruction decoder, memory support and address latch blocks;
FIG. 7 is a general schematic of the broadcast block of FIGS. 6A and 6B;
FIG. 8 is a detailed schematic of the broadcast block of FIG. 7;
FIG. 9 is a general schematic of the resolve portion of the resolve-report block of FIGS. 6A and 6B;
FIGS. 10A and 10B are a detailed schematic of the resolve-report block of FIGS. 6A and 6B;
FIGS. 11 through 16 are detailed schematics of the block 660 of FIGS. 10A and 10B;
FIGS. 17A 17B are a detailed schematic of the instruction decoder block of FIGS. 6A and 6B;
FIG. 18 is a general schematic of a fault control system; and
FIG. 19 is a flow chart of the software kernel.
DETAILED DESCRIPTION OF THE INVENTION
As shown in FIG. 2, the binary tree parallel processor of the present invention comprises a plurality of processing elements identified as PE1 through PE1023 arranged in the form of a binary tree. Each PE except the root PE1 which has no parent processing element and the leaf processing elements PE512-PE1023 which have no children processing elements communicates with a parent processing element as well as two children processing elements For example, PE2 communicates with its parent PE1 over communication line C2, with a first (or left) child PE4 over communication line C4 and with a second (or right) child PE5 over communication line C5. Similarly, PE4 communicates with its parent PE2 over communication line C4 as well as with its left and right children PE8 and PE9 (not shown) over communication lines C8 and C9.
Root processing element PE1 communicates with a host coprocessor 3 via board to board interface circuitry 7 over line 5. In the preferred embodiment, a Digital Equipment Corp. VAX 11/750 (TM) serves as the host coprocessor and is the only device that would be apparent to a user of the present invention. The presently preferred embodiment may thus be viewed as a transparent back-end processor to the VAX 11/750.
Referring to FIGS. 3, 4A and 4B, a single processing element of the preferred embodiment is shown in block diagram form. This processing element comprises an Intel 8751 microcomputer chip functioning as processor 30, a read/write or random access memory (RAM) 40 comprising two 8k.times.8 static random access memory chips 42, 44 and a 16k.times.1 parity RAM 46, and an I/O device 50.
Processor 30 is an 8-bit microcomputer incorporating a 4K erasable programmable read only memory (EPROM) and a 256-byte RAM on a single silicon chip. The EPROM may advantageously be used for subsequent debugging, modifications or the like. Processor 30 is provided with four parallel, 8-bit ports which simplify interconnection between processors and greatly assist the I/O procedure. RAMs 42, 44 function as the local memory for processor 30.
I/O device 50 serves as the primary means for inputting and outputting instructions and data to and from the processing element. I/O device 50 contains memory support circuitry and also computes parity for both memory and communication, facilitating fault detection. I/O device 50 preferably is an IC implemented in gate array technology to allow the parallel processor to be clocked at 12 megahertz, the maximum speed of processor 30 and RAMs 42, 44, 46. The average machine instruction cycle time is 1.8 microseconds, producing a system with a raw computational throughput of roughly 570 million instructions per second. As will be detailed below, very little of this computational resource is required for communication overhead.
I/O device 50 is accessed as a memory mapped device, that is, processor 30 accesses the I/O device by reading and writing to special locations in its address space. I/O device 50 contains six writable locations and eight readable locations. The four eight-bit parallel ports of processor 30 are designated ports P0 through P3. Port P0 is connected to an address/data bus 51 that connects processor 30 to port P1 of its left and right children processors, that serves as a data bus between processor 30 and memory 40 and also serves as the address bus for the low order eight bits of address which are stored in a latch in I/O device 50. Port P1 serves as the connection between processor 30 and its parent. The connection between port P0 processor 30 and port P1 of its children serves as the data path between the processors and may be a direct processor chip to processor chip connection or may be an off the board connection via a bus transceiver.
Five bits of eight-bit port P2 represented on pins A8 to A12 provide the high order five bits of address to memory 40. One bit of port P2 represented on pin A13 of processor 30 is used as the direction control for the bus transceiver while another bit represented on pin A14 of processor 30 serves as an incoming handshake line.
Five bits of the lines from port P3 and one bit of the lines from port P2 are used as handshake lines when accessing neighboring processing elements, and two bits from port P3 are used as the read and write strobes for memory. Referring to FIGS. 4A and 4B, pins LACK of port P2, LINT of port P3 on processor 30 are connected to pins PACK, PINT of its left child while pins RACK, RINT of port P3 on processor 30 are connected to pins PACK, PINT of its right child. Similarly, pins PACK, PINT of port P3 on processor 30 are connected to either LACK, LINT (if left child) or RACK, RINT (if right child) of its parent. Signals on these six pins represent handshaking to/from the children/parent processors. An asynchronous four cycle handshake protocol is used to move data when transfers are limited to neighboring processors (tree neighbor communication).
Referring again to FIGS. 3, 4A and 4B, lines 52, 53, 54 comprise the data lines connected from processor 30 of the PE to its parent, left child and right child, respectively. As will be apparent, data line 52 of a PE is connected to either data line 53 or data line 54 of its parent on the next upper level. Similarly, lines 56, 57, 58 comprise I/O lines connected from I/O device 50 of the PE to its parent, left child and right child, respectively; and I/O line 56 of a PE is connected to either I/O line 57 or I/O line 58 of its parent on the next upper level.
Referring to FIG. 5, there is shown a functional flow chart of the underlying model for a single computation cycle as performed by a plurality of PEs in accordance with the present invention. In essence, a single computation cycle comprises a broadcast phase 60, a compute phase 65, a resolve phase 70 and a report phase 75. The broadcast phase 60 comprises the steps of broadcasting data and/or instructions to a subset of PEs in the binary tree. Following this broadcast, a computation is performed by a second subset of PEs during the compute phase 65. Subsequent to this computation, the resolve phase 70 determines the identity of a third subset of PEs which will report the result of the computation during the report phase 75. Illustratively the data and instructions are broadcast by the host coprocessor.
For this model, four software dependent modes of operation are distinguishable, depending on the subset of PEs selected during each phase and the type of data and/or instructions which is broadcast. These four modes are: SIMD for single instruction stream multiple data stream, MIMD for multiple instruction stream multiple data stream, MSIMD for multiple-single instruction stream multiple data stream and SPMD for single program multiple data.
More specifically, in SIMD mode, each PE is first loaded with its own data and then a single stream of instructions is broadcast to all PEs. Depending on its local state, which is the result of the originally broadcast data and the preceding instructions, each SIMD PE determines which instructions from the stream to execute.
In MIMD mode, each PE is first broadcast its local program and data and then each PE is logically disconnected from its neighbor PEs and executes independently. Logical reconnection to the rest of the binary tree occurs when the PE has finished its computation.
In MSIMD mode, the binary tree of PEs is partitioned into a number of subtrees which maintain the full functionality of the ordinal tree. The root of each such subtree which is operating in MIMD mode is functionally identical to the root of the ordinal tree except for the I/O capabilities with a host. The system operating in MSIMD mode can therefore be viewed as a set of binary trees.
In SPMD mode each PE is first broadcast a copy of the common program and then broadcast its local data. Unlike MIMD mode, the PE does not logically disconnect itself at this point, but rather continues to receive the data stream that is broadcast from above. The computation of each PE is thus determined by three components, namely, the common program that each PE executes, the local data with which the PE was originally loaded and the continuous data stream being broadcast from above.
The I/O device must know whether its PE and its children's PEs are in the MIMD or SIMD mode because a PE in MIMD is logically disconnected from its parent and therefore no I/O communication may occur between them. Three bits of information corresponding to the MIMD/SIMD status of a PE, its left child and its right child are used by the I/O device.
To maintain these bits a writable bit in each I/O device is used to set itself to MIMD or SIMD. Another pair of bits MIMD.sub.-- RC and MIMD.sub.-- LC are used to indicate if that PE's children are in MIMD or SIMD mode. The pair of bits MIMD.sub.-- RC and MIMD.sub.-- LC are writable in two ways. One is directly by the parent processor. The second way is by propagation of an MIMD/SIMD status bit from the I/O device of a child to its parent. This bit is readable directly by the parent or it may be read and simultaneously latched into the MIMD.sub.-- RC or MIMD.sub.-- LC bits of the parent.
As indicated above, processor 30 and RAM 40 of each PE are off-the-shelf integrated circuits. The I/O device is not. The I/O device of a PE essentially performs five operations, namely the I/O functions of broadcast, resolve and report as well as functions of memory support and debugging. These functions are accomplished by the apparatus of FIGS. 6A and 6B comprising a broadcast block 100, a debug block 200, an instruction decoder block 300, a memory support block 400, a parity check block 500, a report-resolve block 600 and an address latch block 700.
Instruction decoder 300 controls the I/O device by controlling the addressing of the broadcast block 100, debug block 200, parity check block 500 and resolve-report block 600, as well as the MIMD status of its children and itself, supervises transmission of interrupts to the processor and records the address in the address latch 700 at the time of an interrupt.
Memory support block 400 handles all external memory operations. This block determines which 8K.times.8 RAM to select, enables the 16K.times.1 parity RAM.
Each broadcast block 100 handles system wide broadcasts of data and instructions. It is responsible for receiving a byte from a parent PE and transmitting a byte to children PEs within two 8751 processor instruction cycles and then re-synchronizing so another byte can be transmitted from the MIMD root processor.
Debug block 200 handles system wide interrupts from itself, its parent and its children and generally propagates the interrupt signal throughout the entire tree.
Parity check block 500 checks the byte communication among the numerous PEs as well as memory operations for parity errors.
Resolve-report block 600 is responsible for determining which single PE is to be enabled to report its result by transmitting it to the host coprocessor.
Address latch 700 provides te I/O device with addresses and/or data. Address latch 700 is controlled by the address latch enable on line ALE from processor 30. When ALE is high, the address on line AD(7...0.) or the data on line IO.sub.-- BUS7...0. is transmitted through the latch onto line AD.sub.-- LATCH 7...0. to instruction decoder 300.
Before discussing the individual components of I/O device 50 it is helpful to set forth the register organization of this device. As indicated above, the I/O device is memory mapped to the external RAM address space of the processor 30, illustratively to the first 256 bytes of an Intel 8751 microprocessor. In the present embodiment of the invention, there are 14 defined addresses, consisting of data, control, and I/O processor status registers, organized into three functional blocks of control/status block, broadcast block and resolve/report block. The description of these registers is set forth in Table I below:
Unused registers in each of these blocks, and other registers between hex addresses 30 to 18FFFF are reserved for future expansion of the I/O device, and referencing these registers, as well as any registers beyond the upper limit of external RAM, will yield indeterminate results The function of these registers is as follows.
The Control/Status Block
The SYNC, CTRL, STAT, PCAR, RESET and STATE registers in the Control/Status Block are used to provide processor 30 with information about the current state of the I/O device and its neighbors, as well as allowing the processor 30 to save and restore its I/O device's context (during an interrupt handler, for instance) and providing explicit (though limited) control over the global communication paths.
The SYNC register allows processor 30 to specify the MIMD status flags used by the I/O device to control the handling of Broadcast and Resolve/Report operations. In normal operation, writing to the SYNC register causes the signals present on the I/O device's dynamic MIMD status lines to be stored in the internal SYNC latches, which values control the behavior of the Broadcast and Resolve/Report logic. In kernel mode, however, the internal SYNC latches are loaded from the source operand rather than the external MIMD status lines. This provides a way for processor 30 to explicitly control the behavior of Broadcast and Resolve/Report operation with respect to neighboring I/O devices, regardless of their actual state. Resetting the I/O device clears the SYNC register, indicating SIMD operation for the local I/O) device as well as its descendants. Reading the SYNC register yields the contents of the I/O device's internal SYNC latches.
The CTRL (Control) register may be used to specify SIMD and MIMD operation for the local I/O device, specify which events will cause the I/O device to interrupt processor 30, enable or disable kernel mode operation, or initiate global interrupts.
The CTRL register contains bits corresponding to the I/O device's MIMD-self latch, the kernel mode enable latch, mask bits for the three sources of I/O device interrupts to processor 30 and a global interrupt trigger. With the exception of the global interrupt trigger, each of these bits is a simple latch, and may be read or written in either user mode or kernel mode. Writing a one bit to the global interrupt trigger initiates a global interrupt to the entire tree, and interrupts the I/O device regardless of the setting of the global interrupt mask bit. Resetting the I/O device places the I/O device in SIMD mode, disables parity and global interrupts, and disables kernel mode.
The STAT (Status) register provides processor 30 with various information about internal and external I/O device events. The presence of a one bit in any of parity.sub.-- err, refresh.sub.-- err, or global.sub.-- err indicates that the corresponding event has occurred. The I/O device can also interrupt processor 30 whenever any of these events occurs, depending on the values of the interrupt mask bits in the CTRL register. When a parity error has been detected, the mem.sub.-- parity bit will be one if the source of the bad parity was external RAM, zero if the source was the communications logic between processing elements. Resetting the I/O device clears these bits.
Reading ext.sub.-- mimd.sub.-- lc and ext.sub.-- mimd.sub.-- rc will return the current value of the signals on the external MIMD pins connected to the left and right descendant I/O devices. Since these values are not latched, write operations have no effect on them. (These are the values loaded into the internal SYNC latches by user mode writes to the SYNC register.)
The I/O device latches the low-order address byte from the address bus of processor 30 into the PCAR (Parity Check Address) register whenever a global interrupt occurs, provided that the PCAR register has been read by processor 30 (or a reset operation has been done) since the occurrence of the previous global interrupt. This interlock is provided so that the data in the PCAR register is not overwritten by new data if subsequent global interrupts occur before processor 30 is able to examine the PCAR register during processing of the initial interrupt. Read and reset operations write-enable the PCAR register internally. Resetting the I/O device also clears it to zero. Processor 30 is allowed to write data into this register only when the I/O device is in kernel mode.
The RESET register is written by processor 30 in order to perform a software reset of the I/O device, an operation that may only be done in kernel mode. The source operand of processor 30 is ignored during write operations, and read operations yield indeterminate data.
The STATE register contains status flags describing the state of the resolve-report and broadcast logic, as well as the parity error and global interrupt indicators.
Broadcast Block
The broadcast block contains the B.sub.-- IN, B.sub.-- OUT and B.sub.-- STATUS registers accessed by processor 30 to retrieve global broadcast data in SIMD mode, and to broadcast data to the locally-rooted subtree in MIMD mode.
Reading the B.sub.-- IN (Broadcast Input) register yields the most recent byte of broadcast data and enables the I/O device's broadcast ready pin, indicating that the I/O device is ready to receive the next byte from its MIMD ancestor. The B.sub.-- IN register can be written by processor 30 only when the I/O device is in MIMD mode.
The B.sub.-- OUT (Broadcast Output) register is written by processor 30 in MIMD mode in order to broadcast a byte of data to the MIMD subtree rooted at the local processing element. The data appears in the B.sub.-- IN register of each descendant SIMD processing element, as well as the B.sub.-- IN register of the local I/O device. (The semantics of the broadcast operation require that the MIMD root processing element read its own broadcast data, since MIMD mode processing elements will also function as SIMD processing elements in order to provide a complete binary tree.) If the B.sub.-- OUT register is written in SIMD mode, the data will be transmitted when the I/O device is put in MIMD mode and the subtree is ready to receive it. This is not recommended, however, since a simultaneous broadcast from an ancestor processing element could either overwrite the data while the local I/O device is in SIMD mode or cause an I/O parity error while the local broadcast is in progress.
The B.sub.-- STATUS (Broadcast Status) register contains three kinds of information: bits b.sub.-- ready and b.sub.-- dav indicate whether the broadcast logic is ready to accept the byte to be broadcast in MIMD mode, and whether new broadcast data is available in SIMD mode, respectively; tree.sub.-- ready indicates whether or not the locally-rooted subtree is ready to receive data; input.sub.-- parity and output.sub.-- parity contain the dynamic parity bits for the B.sub.-- IN and B.sub.-- OUT registers, respectively.
Flag b.sub.-- dav is set to one when the I/O device receives broadcast data in SIMD mode from an ancestor processing element, or broadcast data in MIMD mode (since processing elements always receive their own broadcasts). This flag is cleared whenever processor 30 reads the contents of the B.sub.-- IN register or resets the I/O device. Flag b.sub.-- ready is set to one whenever the I/O device is in MIMD mode and there is no outstanding data to be transmitted to descendant processing elements. This flag does not, however, indicate that the descendant processing elements have retrieved previous data from their I/O devices. It simply means that the local I/O device is ready to buffer another byte of output. Sub.sub.-- tree.sub.-- ready can be used in MIMD mode processing elements to determine whether a broadcast byte can be delivered immediately because it indicates that no processing elements have outstanding data in the B.sub.-- IN registers of their local I/O device. Flag tree.sub.-- enable can be cleared in order to disable all serial communication between the local I/O device and all of its children. This has the side-effect of clearing the sub.sub.-- tree.sub.-- ready flag in all ancestor processing elements.
Only b.sub.-- ready, tree.sub.-- enable, and b.sub.-- dav are affected by write and reset operations, and then only if the I/O device is in kernel mode. Reset turns on b.sub.-- enable, tree.sub.-- enable, and b.sub.-- ready, and turns off b.sub.-- dav, indicating that no SIMD input data is available and no MIMD mode is pending.
Resolve/Report Block
The Resolve/Report block contains the RR.sub.-- IN, RESOLVE, REPORT, RR.sub.-- STATE.sub.-- 0 and RR.sub.-- STATE.sub.-- 1 registers accessed by processor 30 to perform the Min-Resolve and Report operations.
When flag rr.sub.-- dav of the RR.sub.-- STATE.sub.-- 0 register is set, the RR.sub.-- IN register contains the next available byte of resolve/report data. Reading the RR.sub.-- IN register clears this flag. The RR.sub.-- IN register is normally read only by a MIMD mode processor 30 or when the state of the I/O device must be saved.
The RESOLVE register accepts a byte of resolve/report data and initializes the resolve/report logic. Both MIMD and SIMD mode processing elements write resolve/report data to this register at the beginning of a resolve/report sequence. Each SIMD I/O device propagates the largest value calculated among itself and its two descendants up to its parent, until the MIMD processing element at the root of the subtree receives the largest such byte in its RR.sub.-- IN register.
The REPORT register accepts successive bytes of resolve/report data following an initial byte written to the RESOLVE register.
A processing element in MIMD mode must read flag rr.sub.-- dav of RR.sub.-- STATE.sub.-- 0 register to determine whether or not resolve/report data is available in the RR.sub.-- IN register. The other bits of this register are internal status registers and exist only to allow the state of the resolve/report logic to be preserved across interruptions in the resolve/report operation. In SIMD mode, bit.sub.-- rr ready indicates whether or not the resolve/report logic is ready to accept a new byte of data in the RESOLVE or REPORT registers; bit rr.sub.-- dav indicates that resolve/report data is available in the RR.sub.-- IN register. The remaining bits in this register are used only to save or restore the context of the resolve/report logic during a context switch.
The RR.sub.-- STATE.sub.-- 1 register contains additional resolve/report state information, including flags to indicate whether the local processing element is still eligible to compete in the resolve/report operation (i.e. it has not been "killed"), which are monitored by the resolve/report firmware.
Resetting the I/O device has the following effects on its internal registers:
in the SYNC register, the internal MIMD latches are cleared, indicating that the local I/O device and its immediate descendants are in SIMD mode;
in the CTRL register the internal MIMD-self latch is cleared, the parity interrupts and global interrupts are disabled and the kernel mode is disabled;
in the STAT register, parity and global interrupt event flags are cleared;
in the PCAR register, the register is cleared and enabled to receive the low-order address byte recorded at the next global interrupt;
in the STATE register, the broadcast and resolve/report logic are reset to their idle state, and parity error and global interrupt flags are cleared;
in the B.sub.-- STATUS register, all broadcast logic is enabled and the broadcast data-available flag b.sub.-- dav is cleared; and
in the RR.sub.-- STATE.sub.-- 0.sub.-- and RR.sub.-- STATE.sub.-- 1 registers, all resolve/report logic is enabled and the resolve/report data-available flag rr.sub.-- dav is cleared.
Instruction decoder 300 of FIGS. 6A and 6B is depicted in greater detail in FIGS. 17A and 17B. Instruction decoder 300 comprises six functional blocks, namely primary decoder 320, sub-decoder 330, control register 340, status register 350, synchronous register 360 and parity error address register 370. The four registers 340-370 are the CTRL, STAT, SYNC and PCAR registers described above.
Primary decoder 320 decodes the high order bits address latch 700 of FIGS. 6A and 6B over line AD.sub.-- LATCH 6..4 and pin A14 of the 8751 processor of FIG. 4A to initially determine whether the I/O device is being addressed and then, if indeed it has been addressed, to select the appropriate block with either a read or write operation.
Sub-decoder 330 is a block which may be selected by primary decoder 320. Once selected, sub-decoder 330 will read the lowest 3 bits of the address on line ADLATCH(2...0.) to sub-decoder 330 to select either an I/O device reset or control register 340 via control register write and control register read lines CNTREGWR, CNTREGRD, status register 350 via status register write and status register read lines STREG.0.WR, STREG.0.RD, synchronous register 360 via synchronous register write and synchronous register read lines SYNCREGWR, SYNCREGRD, or parity address register 370 via parity address register read line PADDRREGRD.
Control (CTRL) register 340 contains mask bits for interrupt handling, a kernel enable bit and a set global interrupt bit on line SET.sub.-- GLOBAL.sub.-- INT, all of which are used by status register 350. Control register 340 also provides a bit status on line MIMD.sub.-- S for MIMD/SIMD operation. On reset all masks are enabled, the kernel bit is cleared and the device is configured for SIMD operation.
Status (STAT) register 350 allows for dynamic reading of the MIMD status of the left and right children PEs on lines MIMD.sub.-- LC, MIMC.sub.-- RC and receives information from the I/O device concerning the internal state of its children and itself. Status register 350 essentially serves to process parity errors and global interrupts and then to interrupt the processor if the masks are disabled. Once the processor is interrupted, it will read the status register to check if the interrupt was caused by a global interrupt or by a parity error. If the interrupt was caused by a parity error, the status register returns a bit which if set, indicates that the error was caused by a memory read operation. If this bit is not set, the parity error was caused by a transmission error in the Broadcast or Resolve-Report block. On reset, all registers are cleared.
Synchronous (SYNC) register 360, upon receiving a write command on line SYNCWR, latches the state information on the MIMD status lines MIMD.sub.-- LD, MIMD.sub.-- RC from the children and the MIMD status information MIMD.sub.-- S in control register 340 so that this data is accessible to the other blocks of the I/O device. The kernel enable bit must be set to explicitly write to this register from the data bus. On reset, synchronous register 360 clears all bits causing all PE's to be in the SIMD mode.
Parity error address (PCAR) register 370 records the address in the address latch 700 of FIGS. 6A and 6B over line AD.sub.-- LATCH7...0. at the time of an interrupt. Parity error address register 370 is used primarily to record hardware memory chip failures.
Memory support block 400 of FIGS. 6A and 6B handles all external memory operations. This block determines which 8K.times.8 dynamic or static RAM memory chip 42,44 of FIG. 3 to select, enables the 16k.times.1 static parity memory chip 46 of FIG. 3 and controls the refreshing of two 8K.times.8 RAMs. Memory support block 400 is controlled by address latch enable ALE on line A14 of processor 30 and the low order bit of the address latch of processor 30 of FIG. 4A. Line A14 is examined to determine whether or not a memory operation may take place. Similarly, the low order bit of the address latch is examined to determine which of the two 8K.times.8 RAMS should be selected for a memory operation while the other RAM is refreshed. The parity chip is enabled over line PC.sub.-- EN of memory support block 400 whenever an external memory operation occurs.
More specifically, if a memory operation is to occur, memory support block 400 of FIGS. 6A and 6B latches the low order eight bits of address that appear on port P0 of the processor 30 of FIGS. 3 and 4A and provides these low order eight bits to the external memory. Memory support block 400 must also determine whether the memory access is for the I/O device or for the external memory. This is accomplished by a bit fed to the I/O device from line A14 of port P2 of processor 30 of FIGS. 3 and 4A and subsequent memory support block generated control strobes sent to the external memory as well as the I/O device. With respect to error detection, memory support block 400 also generates and checks parity for both memory operation and I/O operations. Upon finding a parity error, the low order eight bits of the address are saved in the PCAR register by memory support block 400.
A general schematic of the broadcast operation is shown in FIG. 7. Broadcast block 100 comprises a broadcast input register 82, a broadcast output register 80, a multiplexer 84, AND gates 86, 92 and a D-type flip flop 88. The broadcast input and broadcast output registers are the B.sub.-- IN and B.sub.-- OUT registers described above in conjunction with Table I. Broadcast is performed as a synchronous bit serial operation and depends on the mode of operation, i.e. SIMD, MSIMD, MIMD or SPMD. Broadcast comprises the transmission of instructions and/or data to/from the individual PEs. Illustratively, data and instructions are broadcast by the host coprocessor to a subset of PEs which then proceed to execute at least some of the transmitted instructions on at least some of the transmitted data.
The broadcast input register 82 is a shift register with serial input terminal S.sub.in, a shift enable, and a parallel output terminal If the PE is in SIMD mode, incoming broadcast data illustratively originating from the host coprocessor and transmitted through higher level PEs is shifted serially into the input cell via AND gate 86. When a 0 start bit is shifted to the end of the shift register, it generates on line 90 a bin-ready signal that is routed to shift enable and disables shifting When the incoming data is read in parallel from the input register, 1-bits are parallel loaded into the input shift register to keep it from reading a data bit as a start bit. The parallel output from register 82 illustratively is connected to processor 30 via bus 51 and port 0.
The bin-ready signal provides a control signal to prevent output data from being lost if the output cell is loaded before all the children read the previous broadcast. To check for valid incoming data in input register 82, receiver PEs simply have to poll this control signal. To check whether a byte may be broadcast out, a broadcast-ready signal P.sub.-- READY is generated on line 94 by each I/O device to be used by the PE and to be propagated up the tree to the I/O device of the parent PE. Broadcast-ready is generated by AND gate 92 of a PE by gating together its bin-ready signal on line 90 and the broadcast-ready signals from its two children and the MIMD/SIMD status bits for its children.
D-type flip-flop 88 also receives the incoming data from the parent PE via AND gate 86 and propagates the data through multiplexer 84 and down the tree to the children PEs delayed by one clock cycle. This data may be broadcast to all the PEs of the array with a delay of only one clock cycle for each level of the array. Thus, it takes only nine clock cycles for data to reach the lowest level (PEs 512-1023) of the array of FIG. 2 and another ten clock cycles to clock a start bit, eight bits of data and a parity bit into the broadcast input shift register 82. Since a processor instruction cycle typically requires over twenty clock cycles, there is sufficient time in one processor instruction cycle to broadcast data to all the PEs.
The broadcast output register 80 is a 10 bit shift register, 8 bits for data, 1 bit for parity and 1 bit for a 0 value start bit. The shift register has a serial input terminal S.sub.in, a 10 bit parallel input terminal, a load terminal and a serial output terminal S.sub.out. The serial input terminal S.sub.in is wired to logic 1. The output shift register shifts continuously. In a broadcast out operation, eight data bits which illustratively are the result of a computation by processor 30 or have been received from another PE, one parity bit and one start bit are parallel loaded into the shift register through the load line. These ten bits are then shifted out through the S.sub.out line and into a multiplexer 84 to provide a multiplexed serial output data.
Referring now to FIG. 8, broadcast block 100 of FIGS. 6A, 6B and 7 is depicted in greater detail The broadcast block functionally comprises broadcast input shift register 110 which receives and holds a byte sent to the I/O device, broadcast output shift register 140 which can transmit a byte throughout the lower subtree, and a broadcast control block 170 which performs synchronization and supervises the passage of data down the tree. As indicated above, broadcast input register 110 and output register 140 are both accessible through memory mapping, thus permitting the processor 30 to read or write a byte. The processor can determine whether a byte has been received by checking the b.sub.-- dav status bit or whether a byte is available to be transmitted by checking the b.sub.-- ready status bit. These bits may be read from address 12H (bits 4 and 0, respectively) or address 05H (bits 2 and 1, respectively).
Input shift register 110 and output shift register 140 of FIG. 8 correspond to input register 82 and output register 80 of FIG. 7. Broadcast control block 170 of FIG. 8 uses signals broadcast left child B.sub.-- LC, left child ready LC.sub.-- Ready, broadcast right child B.sub.13 RC and right child ready RC.sub.-- Ready to determine when a byte may properly be broadcast. Similar broadcast parent and parent ready signals are provided for the parent processor on lines B.sub.-- P and P.sub.-- Ready of broadcast control block 170. The lines B.sub.-- P and P.sub.-- Ready are two of the eight lines 56 that connect the I/O device of a PE to the I/O device of its parent PE and these lines are received at the parent PE as two (B.sub.-- LC, LC.sub.-- Ready) of the eight lines 57 from a left child PE or two (B.sub.-- RC, RC.sub.-- Ready) of the eight lines 58 from a right child PE as the case may be.
The broadcast block of FIG. 8 may be in one of three states, namely idle, receive broadcast or transmit broadcast. The broadcast block generally spends a substantial amount of time in the idle state. If the processor is not in MIMD mode, it is possible for an I/O device to receive a broadcast issued from above itself in the binary tree. The transition from the idle state to the receive broadcast state is accomplished as follows. In the idle state the broadcast line to the parent B.sub.-- P is held high by the parent while input shift register 110 constantly shifts in ones to prevent register 110 from reading a data bit as a 0 start bit as described in conjunction with FIG. 7. The broadcast lines to the children B.sub.-- LC, B.sub.-- RC are also held high by the broadcast control block 170. When a subtree signals the MIMD root PE that it is ready by setting the corresponding ready lines high, a broadcast is issued from the MIMD root PE above in the tree by sending a zero start bit. This start bit enables control block 170 and is passed on to each of the children. The bits are delayed through D flip-flop 88 so as to hold the transmission synchronous to the system clock. This causes a one bit clock delay per I/O device. Once 9 bits (8 bits data and 1 parity bit) are transmitted, the zero start bit disables input shift register 110 and drives the parent ready line P.sub.-- Ready to the parent low. In like fashion, he P.sub.-- Ready lines in subsequent levels of the tree go low as the zero start bit is shifted to the output of shift register 110 of that level and AND gate 92 is disabled.
At each level, the processor 30 checks its b.sub.-- dav status bit to see if input shift register 110 has received the data byte. If so, the processor can read it from the I/O device. The I/O device checks input shift register 110 and if the processor has read the byte and the children's ready lines RC.sub.-- Ready, LC.sub.-- Ready are high, then the I/O device forces the parent ready line high signaling that it is ready to receive the next byte. If the processor has not read the byte, then control block 170 keeps the parent ready line low until the byte is read by the processor.
Referring to FIG. 9, there is depicted a general schematic of the resolve portion of the resolve-report operation. Essentially, the resolve operation involves determining the identity of the single PE which will perform the operation of reporting the result of a computation carried out by a number of PEs. Each PE carrying out the computation is assigned a value. The single winning PE is notified that it has won and subsequently it alone reports the result. In the case of a tie, the resolve circuit designates as the winner the PE with the winning value that is encountered first in an "in order" traversal of the binary tree.
The resolve block of each I/O device computes the following values where VP is the minimum resolve value:
Where VS is the value of the number offered by the local PE, VL is the VP value computed by the left child and VR is the VP value computed by the right child. In other words, each I/O device computes the minimum of the resolve value offered locally and the minimum values computed by its two children.
KP, for kill from parent, is a Boolean signal coming into an I/O device indicating that the local PE can not be the winner of a resolve nor may any of its children be the winner. The left child is killed, KL, if there is a kill from its parent KP, or the value either at the local PE or from the right child was less than the value from the left child, VL. KR and KS, for kill right and kill self, are computed similarly. KS is the signal read locally within the PE to determine whether it is the winner of the resolve.
The minimizing part of resolve is computed bit serially as depicted in FIG. 9. Block 600 of FIGS. 6A and 6B comprises first and second FIFO serial registers 93, 94, buffer 95, multiplexer 96, comparator 97 and buffer 98. Buffer 95 and buffer 98 are the RESOLVE and RR.sub.-- IN registers described in conjunction with Table I. Serial comparator 97 computes the current winning PE by comparing one bit at a time from each of the three sources VS, VL and VR and forwarding the winning bit via multiplexer 96 to the first or second FIFO serial register 93, 95 in a parent PE. Since the subtree that is computing the resolve may not be a complete binary tree, the value bits from the children may not arrive in the same clock cycle. FIFO buffers 93, 94 are used to align the incoming bits of the children with the bits of their parent in buffer 95.
Upon receipt of a KP signal or upon completion of a comparison, comparator 97 also issues the signals KL, KR and KS, determined as indicated by the logic equations set forth above.
The foregoing arrangement permits comparisons of values to be made in pipeline fashion simultaneously on several levels of the binary tree. Thus, the first bit of a winning value may be compared with the first bits of other winning values in a grandparent PE while the second bit is being compared in a parent PE and a third bit is being compared in a child PE and still additional bits are waiting to be compared. Advantageously, there is a delay of only a single clock cycle in performing the comparison at each level in the binary tree and the comparator is able to compare one bit of each of the offered values per clock cycle. As a result, the entire resolve operation can be performed in the time it takes a number offered by a PE to be clocked through the comparator plus the propagation time through the tree of one clock cycle per level. If the numbered offered is only a byte and the processor array has only ten levels, as in FIG. 2, the entire resolve operation can be completed in less than twenty-clock cycles which is less than the average processor instruction cycle.
The report operation is implemented in the kernel program by using the resolve circuit. Report is only well defined if there is a single enabled PE in the SIMD subtree. To implement the report operations, all PEs do a resolve where the SIMD disabled PEs use maximum values (e.g., FF(hex)) and the single enabled PE offers the value to be reported as its resolve value. The value offered by the enabled PE is therefore less than or equal to all other values resolved and the correct result is reported.
Resolve-report block 600 of FIGS. 6A and 6B is depicted in more detail in FIGS. 10A and 10B. Resolve-report block 600 comprises four blocks, namely subdecoder 610, output register 620, comparater 640 and control block 660. Subdecoder 610 performs decoding and transmits read/write enabling signals to output register 620, comparator 640 and control block 660. Output register 620 which is the RESOLVE register of Table I receives a byte from processor 30 which is to be transmitted up the binary tree on line BUS7...0. of register 620. Comparator 640 is used primarily in the resolve operation and is a specific implementation of the serial comparator 97 of FIG. 9. Comparator 640 determines the winner of a resolve and issues KILL commands to children who do not have the proper value. Control block 660 performs transmitting, synchronizing and receiving functions. Each PE is responsible for receiving a byte from each of its children and transmitting the winning byte up to its parent as fast as possible. To do this as fast as possible, two FIFOs are used. Once this resolve is done, a data path is set up from the winning PE to the MIMD root PE.
In the practice of the invention, the report-resolve block of FIGS. 10A and 10B performs two types of transmissions, namely report and resolve. The resolve transmission utilizes two specific addresses. The first byte of a resolve is written into a first address which resets comparator 640. Following bytes are written to a second address but do not reset the comparator. If comparator 640 of the resolve-report block does a compare on the three bytes from itself and its two children and determines a winner, then that winner reports the result. However, if there is a tie involving the left child, comparator 640 will call the left child the winner; and if there is a tie between itself and the right child, the comparator determines itself to be the winner. In any case the I/O device will transmit up the winning byte with a 0 start bit. Upon encountering the last byte of a resolve, processor 30 reads the KILL signals from comparator 640 on lines kill right child KILLRC, kill left child KILLLC and kill parent KILLP in order to determine whether that byte was the winning byte of the resolve.
By this means successive levels of the binary tree select a single winning value, pass it on to the next level and disable all processing elements except the processing element which has the winning value.
The report function can then be implemented by having all the disabled processing elements report a maximum value such as FF(hex) when they are disabled and broadcasting an instruction to the winning processing element to report whatever of its data is of interest. Since all the other processing elements are set to a maximum value, operation of the resolve one more time will cause this data from the winning processing element to be selected consistently as the winning value and passed to the root processing element.
FIGS. 11-16 depict the elements of control block 660 of FIGS. 10A and 10B in greater detail. More specifically, input register 680 of FIG. 11 which is the RR.sub.-- IN register of Table I receives and holds the winning byte transmitted from the child to an MIMD parent on line BUS(7...0.) of input register 680. Syncing and data enabling signals SYNC, XMTDTA are input to NAND gate 681 whose output serves as the clock to input register 680.
Two first-in first-out devices (FIFOs) 690, 695 of FIG. 12 receive bytes transmitted from SIMD children to their parent. These devices are depicted as FIFOs 93, 94 in FIG. 9. If FIFO 690 of the PE receives a winning byte from the PE's right child, it will transmit this winning byte up to its parent as fast as possible on line FIRCDO. Similarly, if FIFO 695 of the same PE receives a winning byte from the PE's left child, it will transmit this winning byte up to its parent as fast as possible on line FILCDO. More specifically, FIFO 690 is supplied with data from the right child over input line RRDRC of FIG. 12. Control signals PRENXMT applied to NAND gate 692 and XMTDTA applied to D-flip flop 691 permit the output of data to the PE. Similarly, FIFO 695 is supplied with data from the left child over input line RRDLC. Control signals PRENXMT applied to NAND gate 697 and XMTDTA applied to D-flip flop 696 permit the output of data to the PE. Reset signals are applied to lines FIFORESET and R/I of NOR gate 693 to reset FIFO 690 for the right child and to NOR gate 694 to reset FIFO 695 for the left child.
Input register 680 of FIG. 11 and output register 620 of FIGS. 10A and 10B are memory mapped to permit processor 30 to read or write a byte. Processor 30 may check the status of control block 660 to determine whether a byte has been received or needs to be transmitted.
FIG. 13 depicts the GO circuitry to produce the status signal RRCLGO. Inputs to this circuit include signals RRPARDY and RRPAGO.
FIG. 14 depicts the EMPTY logic to produce the signal PREMPTY. Inputs to this circuit include signals ROUT.0.WR and RROUT1WR.
FIG. 15 depicts the READY logic to produce the signal RRPARDY. Inputs to this circuit include signals RRRCRDY, SYCRC RRLCRDY, SYNCLC and the RREMPTY signal produced by the EMPTY logic of FIG. 14.
FIG. 16 depicts the SEND logic to produce the signals XMTDTA, ENXMT and RSTFLL. Inputs to this circuit include FILC, FIRC (both produced by the RECEIVE FIFO circuit of FIG. 12), SYNCLC, SYNCRC and RRCLGO (produced by the GO logic of FIG. 13).
The resolve-report block may be in any one of several states, namely, idle, transmit resolve-report byte, receive resolve-report byte and start resolve-report.
If the PE is not in MIMD root mode the idle state is defined as follows: a status signal on line RRCLGO of control block 660 of FIGS. 10A and 10B is held logic high by the MIMD Root parent; the right child and left child data lines RRDRC, RRDLC respectively, of control block 660 are held logic high; and the parent ready line RRPARDY also of control block 660 is held logic low by the child.
The I/O device leaves the idle state and enters the transmit resolve-report state when the SIMD children's processor 30 writes a resolve byte into the I/O chip's report-resolve output register 620 of FIGS. 10A and 10B. If the SIMD child is a leaf PE in the binary tree, the parent ready line RRPARDY of control block 660 of FIGS. 10A and 10B enters the logic high state. If the SIMD child is not a leaf PE, the resolve-report ready status bit, i.e, the bin-ready signal on line 90 of FIG. 7 is ANDed by AND Gate 92 of FIG. 7 with each of the children's ready lines to determine the value of the parent broadcast ready line, as shown in FIG. 7. Once a MIMD root PE determines that the SIMD sub tree is ready, the PE will cause the status signal on line RRCLGO of control block 660 of FIGS. 10A and 10B to enter the low state. When this occurs, the leaf PEs transmit up the value in output register 620 of FIGS. 10A and 10B. If the PE is not a leaf and in SIMD, then it must wait until both of its children have begun to transmit their resolve byte up the binary tree. FIFOs 690, 695 which receive bytes transmitted from SIMD right and left children to their parent permit comparator 640 of FIGS. 10A and 10B to compare the bits as soon as all three first bits from the left child, right child and their parent are received Once an SIMD PE has finished a one bit compare, it forces the status signal on line RRCLGO high, even if the parent still has a low line, so as to enable the next byte to be written without any delays.
In addition to the idle, transmit resolve-report byte and start resolve-report states, a MIMD root PE may also be in the receive resolve-report state. When a subtree signals to a MIMD root PE that it is ready to transmit up a byte, the MIMD root PE's I/O device initially checks that it has written a byte to output register 620 of FIGS. 10A and 10B as if it were an SIMD PE. At this point, the MIMD root PE's I/O device drops the status signal on line RRCLGO of control block 660, signalling the lowest portion of the tree to begin transmission. Signal RRCLGO is depicted in the GO Logic and the SEND logic of FIGS. 13, 16, respectively. The I/O device now enters the Receive resolve-report mode. Each child transmits up either its resolve byte or report byte, depending on the operation taking place. Once the children of the MIMD root PE transmit up the data, the MIMD root PE does a one bit compare on the data and, like any other SIMD PE, the winning byte is transmitted to resolve-report block's input register 680 of FIG. 10 over line BUS(7...0.), where it may be read by processor 30.
Referring again to FIGS. 6A and 6B, parity check block 500 checks the byte communication among the numerous PEs as well as memory operations for parity errors. Parity check block 500 essentially reads the parity of the eight-bit data line IO.sub.-- BUS7...0. connecting broadcast block 100, instruction decoder block 300, report-resolve block 600 and address latch block 700 and checks this parity against a parity bit to determine whether a parity error has occurred.
More specifically, if processor 30 writes to the broadcast block 100 or report-resolve block 600, a parity bit is placed on internal parity bit bus IPB of parity check block 500 and written as the ninth bit to the data registers of report-resolve block 600. Line IPB connects parity check block 500, broadcast block 100 and report-resolve block 600. When a memory write operation takes place, parity check block 500 generates a parity bit which is stored for future reference by being written through data line PB.sub.-- OUT of parity check block 500 into 16k.times.1 parity RAM 46 of FIGS. 3, 4B.
Whenever processor 30 reads the data input registers in broadcast block 100 or resolve-report block 600, the parity of the data is checked against the parity bit transmitted with the data. If they fail to match, a parity error signal PAR.sub.-- ERR is sent to instruction decoder block 300, which, if the parity error mask is disabled, will interrupt the processor. When an external memory read operation occurs, the parity of the data in the 8K.times.8 RAMs 42, 44 of FIG. 3 is checked against the parity originally written into the 16K.times.1 RAM 46 of FIG. 3. The line which brings this parity bit from the 16K.times.1 RAM to the I/O device is PB.sub.-- IN of parity check block 500 of FIGS. 5A and 5B. If the parity of data bus IO.sub.-- BUS7...0. of parity check block 500 does not match PB.sub.-- IN, then the parity error PAR.sub.-- ERR line goes low. Parity error line PAR.sub.-- ERR of parity check block 500 is connected to instruction decoder block 300 and informs the instruction decoder that a parity error has occurred. In both read cases, there is a delay of two clock cycles from when the read line toggles low until the parity bit is checked against the data bus IO.sub.-- BUS7....0.. A bit in instruction decoder block 300 knows if the parity failure was local to the I/O device or was memory failure
Referring back to FIGS. 6A and 6B, there is shown debug block 200 which handles system wide, or global, interrupts. These global interrupts may be generated by either processor 30 initiating its own global interrupt (known as debug) or by another I/O device.
Global interrupts permit any PE in the tree to initiate an interrupt in every other PE as well as the host coprocessor. Debug block 200 may be used to signal a hardware error or may be used to enter a debugging and diagnostic supervisor program. The I/O device is constructed so the processor may read and write the I/O device's state in order for context switching to be performed. Context switching involves saving the state of the I/O device in external memory. Upon a global interrupt, the state of the processor and the I/O device are saved and a supervisor program entered. Upon exiting the supervisor program the state of the PE and the I/O device may be restored and the tree may resume processing from where the interrupt occurred.
Debug block 200 is essentially a multidirectional bus repeater. Debug block 200 may receive an interrupt signal from four different sources; the local PE, as a result of writing a particular bit to the PE, or any of its three nearest neighbors (parent PE or left and right children PEs) in the tree. Upon receipt of an interrupt, debug block 200 will propagate the interrupt in the other three directions until the entire tree is interrupted. Upon removal of the interrupt signal from the interrupt source the block will stop propagating the signal in the other three directions.
Functionally, debug block 200 of FIGS. 6A and 6B is a finite state machine. Input signal IO.sub.-- GLOBAL.sub.-- INT to debug block 200 comes from instruction decoder 300 and is generated by a global interrupt status bit, which is set high or low by processor 30. Receive interrupt signal INT.sub.-- IO from debug block 200 feeds instruction decoder 300 to generate an I/O device initiated interrupt. Input/output lines, global interrupt by parent G.sub.-- INT.sub.-- P, global interrupt by left child G.sub.-- INT.sub.-- LC, and global interrupt by right child G.sub.-- INT.sub.-- RC of debug block 200 may also be used to communicate interrupts. If the global interrupt was generated by the parent, left child or right child, and not by the I/O device, then whichever of these three lines was forced low indicating an interrupt will also force receive interrupt INT.sub.-- IO low. If the global interrupt is caused by input signal IO.sub.-- GLOBAL.sub.-- INT to debug block 200, then signals G.sub.-- INT.sub.-- P, G.sub.-- INT.sub.-- LC and G.sub.-- INT.sub.-- RC will also be forced low. Once an interrupt occurs, processor 30 reads a status byte in instruction decoder block 300 which signifies whether the interrupt was local to the I/O device or was generated by a parent or child and also what should be done.
In accordance with the invention, once an interrupt occurs, all or part of the state of the I/O device should be saved in external memory to be restored at some later time. The state of instruction decoder block 300 of FIGS. 6A and 6B should be saved first, then the state of broadcast block 100, and finally the state of resolve-report block 600 in order to allow for any global communications which may have begun before the interrupt to finish before the state is saved. Following this order, no data will be lost as all broadcasts will have been completed by the time instruction decoder block 300 is read out, and all resolve-reports will have settled down by the time the broadcast block is finished being read out. After the state has been saved, a soft reset of the I/O device can be performed or a pre-determined state can be written to the I/O device. The state must be written in the same order as discussed above, since the state in instruction decoder block 300 is used in both the broadcast 100 and resolve 600 blocks.
Once the debugging is done, the state of the tree can be restored. The state of the I/O device is, again, restored in the same order as it was read out, except for one case. The broadcast input shift register must be read out by processor 30 first in order to re-initialize the register. Only after it has been read can data be written into the register. This case only applies to I/O devices which are to be restored into an MIMD status condition.
In the practice of invention, efficiency is maintained by controlling the amount of computation time required by each processor on a particular cycle. Since the amount of time required by each processor may be vastly different depending on its local state, balancing the load by copying a constrained subset of the data of a processor with a skewed computation load to other processors has been found advantageous. The data may be constrained and thereby partitioned based on any suitable ordering or hashing scheme, thus reducing the mean computation time as well as the standard deviation of the mean.
Additionally, the present invention may advantageously be practiced such that operation of the processors will essentially be assured after two successive faults as well as a 50% chance of proper operation after a third successive fault. Although common parity errors can be handled with conventional error correcting methods, complete and irrevocable non-operation of an integrated circuit poses a much more serious problem. However, this can be overcome by replicating the computation four times in a binary tree parallel processor of size 4N+3, where N=1023 for the preferred embodiment.
Referring to FIG. 18, three additional nodes form the uppermost part of the N-size tree, namely T, L and R. These three nodes essentially act as arbitrators assuring agreement between the four identical and concurrent processes in subtrees 1-4. Nodes L and R work concurrently to assure agreement between their descendant subtrees T assures agreement between L and R and aids in the isolation of faults if they occur. T, L and R themselves must be hardened against faults using conventional replicated voting logic.
More specifically, assuming a fault occurs in subtree 2, the results communicated to L by subtrees 1, 2 will differ at some point in the computation. Immediately upon noticing this discrepancy, L notifies T by setting a pin which T reads continually. As R has noticed no faults of its own, it communicates valid results to T. T proceeds to transmit R's results to the external host as well as to L.
Node L, using this value supplied by R via T verifies that subtree 2 has failed. Subsequent operation of L simply passes values from its operational subtree 1 directly to T. Sub-tree can now either work independently to isolate its own fault for direct manual repair, or remain disconnected.
If another fault occurs in one of valid subtrees 1, 3, 4, the other two computations will remain valid and isolate the faulty subtree in the same manner as discussed above. Thus, T is assured to respond with valid data through two faults. If a third fault occurs, T can choose randomly from the two remaining subtrees with a 50% probability of success. Such probability may be increased with the aid of hardware for fault detection.
Advantageously, such a fault control system can be implemented in accordance with the present invention without the need of additional PEs. Nodes T, L and R preferably perform relatively simple functions and do not require a dedicated PE.
Software control of each PE is preferably provided through a kernel system resident within EPROM of each PE. The kernel provides four functions: powerup diagnostics, low level I/O routines, high level language PPL/M support and high level language PPSL support depicted in the flow diagram of FIG. 19. Appendices I and II, respectively, list the machine code and assembly language versions of the kernel.
With respect to the powerup diagnostic function of the kernel, the PEs reset by initializing their memory to 00 upon the application of power and then examining the global interrupt lines. If the global interrupt lines are low, the PE will execute an EPROM resident diagnostic program to generally check the integrity of the PE. If the global interrupt lines are idle (i.e., high, 5 volts), the I/O device is reset, the interrupt masks are disabled, low level status bits are initialized and a timer which is activated at the beginning of initialization is checked for time-out. If the timer times-out because the initialization did not complete, the PE will execute the EPROM resident diagnostic program. Alternatively, if initialization is complete, the PE enters the PPL/M mode.
With respect to the low level I/O function of the kernel, the low level status bits located in the bit address space include bits to indicate whether a PE is a root PE, whether a PE is a right child, whether a PE is in MIMD and its parent is logically disconnected, whether a left child is logically connected, whether a right child is logically connected, whether a left child physically exists regardless of MIMD/SIMD state and whether a right child physically exists regardless of MIMD/SIMD state. Initialization software preferably examines external connections and assigns appropriate status to the low level status bits.
Low level I/O routines are generally byte oriented, except for the I/O switch context routines, which are word oriented. If a routine returns a value, the return value is left in the accumulator.
The only routines that may access the processor ports that are connected to parent and child ports of processor 30 are as follows:
Write left child port.
Write right child port.
Write parent port.
Write host (parent port, callable from root only).
Write to both children at once.
Write host one byte but use proper protocol (parent port, callable from root only).
Read left child port.
Read right child port.
Read parent port.
Read host (parent port, callable from root only).
Read host one byte but use proper protocol (parent port, callable from root only).
These tree neighbor I/O routines are responsible for executing proper four cycle handshake protocol with the parent and children PEs that are logically connected to a local PE. If any such PE is not logically connected, a write is a "no operation" and a subsequent read will return a zero. While these routines test the status bits, the PPL/M high level kernel maintains the correctness of the status bits.
Since ports of a PE are generally configured for communication in a downward direction, the low level routines which move a byte upward in the binary tree must reconfigure their ports in an upward direction for the upward transmission and then return the ports to the downward direction before exiting
The basic low level operation for performing tree neighbor I/O is standard four cycle handshake which requires a ready line and an acknowledge line (Ack). A typical handshake between a Master and a Slave comprises the following steps:
The following illustrative list of routines are performed through the I/O device:
Read a broadcast byte;
Send out a broadcast byte (MIMD PE only);
Send out the first byte of a resolve;
Send out subsequent bytes of a resolve;
Send out bytes of a report;
Read a resolved or reported byte (MIMD PE only);
Set the local processor's I/O device to the children's current MIMD/SIMD state;
Set the local processor's I/O device into an MIMD state;
Set the local processor's I/O device into an SIMD state.
Set the local processor's I/O device to have the left child in MIMD.
Set the local processor's I/O device to have the left child in SIMD.
Set the local processor's I/O device to have the right child in MIMD.
Set the local processor's I/O device to have the right child in SIMD.
Set global interrupt.
Reset global interrupt.
Mask off global interrupt.
Allow global interrupt in interrupt CPU.
Mask off parity interrupt.
Allow parity interrupt in interrupt CPU.
Predicate returns dynamic status of left child MIMD line. Nonzero means MIMD. (Affects carry bit.)
Predicate returns dynamic status of right child MIMD line. Nonzero means MIMD. (Affects carry bit.)
Predicate returns the status of the I/O device's left child state. Nonzero meand MIMD. (Affects carry.)
Predicate returns the status of the I/O device's right child state. Nonzero means MIMD. (Affects carry bit.)
Predicate returns the status of the I/O device's children's state. Nonzero means both are in MIMD. (Affects carry bit.)
Predicate returns the value of a winner bit in the carry.
Write I/O device control register.
Read I/O device control register.
Write I/O device status register.
Read I/O device status register.
Context save passed a pointer to an area of 12 bytes on an off chip RAM. This routine will write out the current context information of the I/O device and leave the I/O device reset and in SIMD mode. (Affects DPTR and carry.)
Restore context passed a pointer to an off device RAM containing a previously written context. This routine will restore the context to an I/O device. (Affects DPTR and carry.)
In general, the I/O device comprises a number of memory mapped data and status registers. Communication routines have the form, busy wait on a status bit, when the status bit indicates that the data register is ready and that the data may therefore be read or written. The predicates read and mask the appropriate status bits.
Referring back to FIG. 19, and with respect to the PPL/M high level kernel, the present invention powers up in PPL/M mode, with all PEs except the root starting to execute an SIMD interpretive loop SIMDLP. If a PE is determined to be a root PE, it will execute the bootstrap loader to download code and load in a user program.
The bootstrap loader is a PPL/M program itself. It has an MIMD part in the root and a SIMD part in each of the PEs. Input files are illustratively in Intel 8080 object module format. The bootstrap loader then decides if a particular record of an input file is part of the program image. If so, header information in the record dictates where the image part is to be loaded in an off chip RAM. This is done in all PEs so each PE has an identical program image in its memory. At the end of the file, control is passed to the user program, according to the start address dictated by the load module.
The parallel behavior of the present invention is achieved by placing identical program images of the user program in all PEs combining with the use of runtime routines to control execution and perform the communications. To illustrate this, a portion of PPL/M code before and after it has been transformed by a preprocessor is presented and the actions of the runtime routines are explained as follows: A segment of the PPL/M code:
The same segment of PPL/M code after preprocessing:
Notice the "do SIMD" statement has been changed to "call SIMD". Code runs conventionally in a MIND processor until a "call SIMD" is encountered. SIMD is a kernel function that looks at the return address stack of processor 30. The return address points to the instruction right after the call SIMD, which is 2 less (due to the jump 10:) than the first address of the SIMD block. Since the program image in all PEs is identical, this address will point to the same SIMD block in all PEs. The address is broadcast to all the SIMD PEs. The SIMD procedure then jumps to the SIMD block itself (since every MIMD PE must also behave like an SIMD PE). At the end of the SIMD block the return instruction acts like the return instruction of the initial SIMD call. This returns the program to the initial "goto" after the call SIMD.
The SIMD interpretive loop SIMDLP in the SIMD PE receives the two bytes of address and does an indirect subroutine call to the SIMD block. When the SIMD PE reaches the return address at the end of the SIMD block it causes the PE to return to the SIMD interpretive loop.
All PEs may now execute the same block of code, simulating SIMD behavior. Notice at the beginning of the SIMD block and right after each communication instruction and assignment to EN1 there is a statement: "if (not EN1) then goto 11;". This is how SIMD enable/disable state is simulated. Each time a communication instruction must execute it must be executed in every PE. In other words, every PE must execute precisely the same sequence of communication instructions.
The conditional jumps on EN1 jump from communication instruction to communication instruction. Other instructions are skipped only if the PE is disabled. This together with disallowing other conditionals in SIMD blocks as well as communication instructions in slice procedures insures discipline over the sequence of communicatlons instructions.
In addition to the above, the following is a list of special locations used by the PPL/M high level kernel:
Referring again to FIG. 19, DoMIMD is the primitive that will cause the partitioning of the binary tree and pass control to the roots of each of the subtrees in a three step process. First, the address of the MIMD routine to which control is being given is broadcast. The partitions and status bits are then setup. Last, the enabled P(s call the MIMD routine. In PPL/M mode the root is always in MIMD mode although if it is SIMD disabled, it will not do a routine called by DoMIMD.
An exit is placed at the end of the MIMD code to cause that subtree to reconnect above. In effect, the I/O device switches back into SIMD and updates the local MIMDflag. The last return from the MIMD procedure will unwind the return stack and leave the PE in the SIMD interpretive loop, SIMDLP.
Sync is a command that will cause a wait until all PEs below the PE executing sync have returned to SIMD mode and causes that subtree to completely reconnect.
Thus it can be seen that there is provided a new and improved system for fast, efficient and fault tolerant data processing through the use of plurality of processor elements arranged in a binary tree configuration wherein each processor element has associated with it an individual I/O device and memory. ##SPC1##
Claims
Certificate of Correction present
The claims shown below may not include correction changes. Use the corrected PDF for authoritative text.
What is claimed is:
1. A parallel processor array comprising: a plurality of processing elements, each comprising: a processor having an arithmetic logic unit, control store, program sequences and instruction decoder; a read/write memory associated with said processor; an input/output means associated with said processor and read/write memory; means for interconnecting said processing elements in a binary tree in which each processing element except those at extremities of the binary tree is connected to one parent processing element and at least first and second child processing elements; said input/output means comprising: means for broadcasting information received from a parent processing element to said child processing elements, such that common information is distributed to each processing element of the binary tree or a subtree thereof without direct control of the processors of the processing elements; and means for determining a priority among respective values of information received from said child processing elements and information received from the processor with which said input/output means is associated without direct control of the processors of the processing elements; wherein the input/output means of the processing elements connected in said binary tree cooperate so that information is broadcast from a first parent processing element to the child processing elements in said binary tree or subtree that are most remote from said first parent processing element, and a priority is determined among values of information at each processing element in said binary tree or subtree, each in a time on the order of the logarithm of the number of processing elements in said binary tree or subtree multiplied by the time for the broadcasting of information from a parent processing element to child processing elements connected thereto, and the time required to determine priority amoung values of information received from the processor of a processing element and the child processing elements connected thereto, respectively.
2. The apparatus of claim 1 wherein said broadcasting means comprises: a register having a serial data input and an output to said processor, a flip-flop, first means for applying data signals simultaneously to said serial data input and said flip-flop, and second means for applying data signals from said flip-flop to said first means in first and second child processing elements, whereby data signals are propagated through the binary tree or subtree by means of the flip-flops of the processing elements of the binary tree or subtree.
3. The apparatus of claim 2 further comprising: means for generating a register full signal when said register is full and means for disabling the transmission of additional data to said plurality of processing elements while said register full signal is being generated.
4. The apparatus of claim 1 wherein the determining means comprises: first, second and third register, said first register storing data associated with said processing element and said second and third registers storing data associated with said first and second child processing elements, a comparator for comparing the data stored in said first, second and third registers to select a winning data in accordance with a predetermined priority, and means for reporting said winning data to one of said second or third registers in a parent processing element.
5. The apparatus of claim 4 further comprising means for disabling each processing element whose data is not selected a winning data, as a result of which every processing element except one is disabled by the time the winning data is reported to the first parent processing element in the binary tree or subtree.
6. The apparatus of claim 5 further comprising means for reporting to the first parent processing element in the binary tree information stored by the processing element that is not disabled.
7. The apparatus of claim 1 further comprising means for subdividing the binary tree into a plurality of sub-trees.
8. The apparatus of claim 7 wherein each sub-tree is operated in a single instruction multiple data mode and the plurality of sub-trees are operated in a multiple instruction multiple data mode.
9. The apparatus of claim 7 wherein at least two of the sub-trees execute identical programs on identical data and the results of such program execution are compared to detect faults in the sub-trees.
Patent Citations (11)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US4101960 | 1978-07-01 | Stokes et al. | |
| US4247892 | 1981-01-01 | Lawrence | |
| US4251861 | 1981-02-01 | Mago | |
| US4344134 | 1982-08-01 | Barnes | |
| US4388683 | 1983-06-01 | Beifuss et al. | |
| US4435758 | 1984-03-01 | Lorie et al. | |
| US4466060 | 1984-08-01 | Riddle | |
| US4543630 | 1985-09-01 | Neches | |
| US4583164 | 1986-04-01 | Tolle | |
| US4622632 | 1986-11-01 | Tanimoto et al. | |
| US4633431 | 1986-12-01 | Bar |
Non-Patent Literature (40)
- D. P. Miranker, 11/15/83, "Performance Analysis of Two Competing DADO PE Designs".
- S. Taylor, 7/29/83, "Prolog on the DADO Machine: A Parallel System for High-Speed Logic Programming".
- S. J. Stolfo, Mar. 1982, "A Tree-Structured Machine Architecture for Production Systems".
- S. J. Stolfo, Aug. 15, 1983, "The DADO Parallel Computer".
- S. J. Stolfo, 9/15/80, "Learning Control of Production Systems".
- M. van Biema, 2/6/84, "PSL: A Parallel Lisp for the DADO Machine".
- M. D. Lerner, 2/4/85, "LISP Compiler for the DADO Parallel Computer".
- S. J. Stolfo, 2/29/84, "PPL/M: The System Level Language for Programming the DADO Machine".
- S. Alterman, 4/15/85, "The Application of the AI and DADO Parallel Processor Technology to Future Unmanned Vehicle Systems".
- P. Waldes, 4/16/85, "Are Maintenance Expert Systems Practical Now?".
- S. J. Stolfo, 5/5/84, "Is CAD/CAM Ready for AI?".
- S. J. Stolfo, 5/5/83, "Knowledge Engineering: Theory and Practice".
- S. J. Stolfo, 10/15/83, "ACE: An Expert System Supporting Analysis and Management Decision Making".
- M. D. Lerner, 6/12/85, "An Overview of the DADO Parallel Computer".
- S. J. Stolfo, 6/6/85, "DADO: A Parallel Computer for Artificial Intelligence".
- S. J. Stolfo, 10/29/84, "The DADO Production System Machine".
- S. J. Stolfo, 8/15/84, "DADO: A Parallel Processor for Expert Systems".
- S. J. Stolfo, 1/18/83, "Architecture and Applications of DADO: A Large Scale Parallel Computer for Artificial Intelligence".
- S. J. Stolfo, Apr., 1981, "Specialized Hardware for Production Systems".
- S. J. Stolfo, 8/15/84, "Five Parallel Algorithms for Production System Execution on the DADO Machine".
- S. J. Stolfo, 4/16/85, "More Rules May Mean Faster Execution".
- S. J. Stolfo, 4/16/85, "A Simple Preprocessing Scheme to Extract and Balance Implicit Parallelism in the Concurrent Match of Production Rules".
- T. Ishida, 10/25/84, "Towards the Parallel Execution of Rules in Production System Programs".
- D. P. Miranker, 4/9/84, "Performance Estimates for the DADO Machine: A Comparison of TREAT and RETE".
- A. Lowery, 10/15/84, "LPS Algorighms".
- A. Lowery, 3/15/84, "LPS Algorithms: A Detailed Examination".
- A. Lowery, 3/15/84, "LPS Algorithms: A Critical Analysis".
- S. Taylor, 2/1/84, "Logic Programming Using Parallel Associative Operations".
- S. J. Stolfo, 1/1/85, "On the Design of Parallel Production System Machines: What's in a LIP?".
- T. Ishida, 3/28/84, "Simultanieous Firing of Production Rules on Tree Structured Machines".
- S. J. Stolfo, 3/15/82, "DADO: A Tree-Structured Machine Architecture for Production Systems".
- Chuan-Lin Wu and Tse-vun Feng, "Interconnection Networks for Parallel and Distributed Processing", (IEEE 1984).
- S. A. Browning, "Computations on a Tree of Processors", Proc. VLSI Conf., California Institute of Technology/Pasedena, Jan. 22-24, 1979.
- A. M. Despain et al., "The Computer as a Component", (Unpublished 1979).
- G. A. Mago, "A Cellular Language-Directed Computer Architecture", Proc. VLSI Conf., California Institute of Technology, Jan. 22-24, 1979.
- R. J. Swan et al., "The Structure and Architecture of Cm*: A Modular Multi-Microprocessor," Proc. 1977 NCC, (Jun., 1977).
- J. R. Goodman, "Hypertree: A Multiprocessor Interconnection Topology", IEEE Trans. on Computers, vol. C-30, No. 12, pp. 923-933, (Dec., 1981), reprinted in the above-cited Wu and Feng reference at pp. 46-56.
- J. L. Bently and H. T. Kung, "Two Papers on a Tree-Structured Parallel Computer", Technical Report, Dept. of Computer Sciences Carnegie-Mellon University, Sep. 1979.
- J. L. Bently, "Decomposable Searching Problems", Information Processing Letters, vol. 8, No. 5, pp. 244-250, (Jun. 1978).
- Carlo H. Sequin, "Single-Chip Computers, the New VLSI Building Blocks", Proc. VLSI Conf., California Institute of Technology, Pasedena, Calif., Jan. 1979.