JP4787711B2 - Data processing apparatus and method, computer program, information storage medium, and data processing system - Google Patents

Description

  The present invention relates to a data processing apparatus and method for generating an object program for a parallel processing device, a computer program for the data processing apparatus, an information storage medium in which the computer program stores data, and an operation corresponding to the object program The present invention relates to a parallel processing device, and a data processing system including the parallel processing device and a data processing device.

  At present, products called so-called CPUs (Central Processing Units) and MPUs (Micro Processor Units) have been put into practical use as processor units that can freely execute various data processing.

  In such a data processing system using a processor unit, various object codes in which a plurality of operation instructions are described and various processing data are stored in a memory device, and the processor unit receives operation instructions and processing data from the memory device. Data is read in order and a plurality of data processes are sequentially executed.

  For this reason, various data processing can be realized by one processor unit. However, in the data processing, it is necessary to sequentially execute a plurality of data processing in order, and the processor unit receives operation instructions from the memory device for each sequential processing. Since it is necessary to read, it is difficult to execute complicated data processing at high speed.

  On the other hand, when the data processing to be executed is limited to one, if the logic circuit is formed by hardware so as to execute the data processing, the processor unit sequentially outputs a plurality of operation instructions from the memory device. There is no need to read and execute a plurality of data processes sequentially. For this reason, it is possible to execute complicated data processing at high speed, but of course, only one data processing can be executed.

  That is, in the data processing system in which the object code can be switched, various types of data processing can be executed. However, since the hardware configuration is fixed, it is difficult to execute the data processing at high speed. On the other hand, a logic circuit made up of hardware can execute data processing at high speed, but only one data process can be executed because the object code cannot be changed.

  In order to solve such a problem, the present applicant has created a parallel arithmetic unit as a processor unit whose hardware configuration changes corresponding to software, and has filed an application as Japanese Patent Application No. 2000-04-3202. In this parallel processing device, a large number of small processing circuits and wiring circuits are arranged in a matrix, and a state management unit is arranged in parallel in the matrix circuit unit.

  A plurality of processing circuits individually execute data processing corresponding to operation instructions set individually, and a plurality of wiring circuits are connected to a plurality of processing circuits corresponding to operation instructions set individually. The relationship is controlled individually. In other words, since the hardware configuration is changed by switching operation instructions between a plurality of processing circuits and a plurality of wiring circuits in the parallel processing device, various data processing can be executed, and a large number of small-scale hardware can be executed. Since the processing circuit executes simple data processing in parallel, the data processing can be executed at high speed.

  Then, since the state management unit sequentially switches the context made up of the operation instructions of the plurality of processing circuits and the plurality of wiring circuits as described above corresponding to the object code for each operation cycle, the parallel arithmetic device corresponds to the object code. Parallel processing can be executed continuously.

Examples of parallel computing devices that sequentially switch contexts in a plurality of cycles in this way include “Introduction to the Configurable, Highly Parallel Computer” published in Non-Patent Document 1 by “Lawrence Snyder” of “Purdue University”, There are Patent Document 1, Patent Document 2, Patent Document 3, Patent Document 4, and the like, for which the applicant has applied for a patent.
JP 2000-138579 A JP-A No. 200-2224025 JP 2000-232354 A JP 2000-232162 IEEE Computer, vol.15, No.1, Jan. 1982, pp47-56

  In the parallel arithmetic device as described above, since the state management unit sequentially switches the context included in the object code for each operation cycle, a plurality of processing circuits and a plurality of wiring circuits arranged in a matrix correspond to the context. Can be operated in parallel every operation cycle.

  However, since such a parallel computing device is fundamentally different from the conventional CPU in both structure and operation, the object code cannot be simply generated from the source code by the conventional method. However, a technique for appropriately generating the object code from the source code has not been established.

  The present invention has been made in view of the above-described problems, and provides a data processing apparatus and method capable of satisfactorily generating data for an object program of a parallel processing apparatus that sequentially switches contexts, a computer program for the data processing apparatus, An object of the present invention is to provide an information storage medium in which a computer program stores data, a parallel processing device that operates in accordance with the object program, and a data processing system including the parallel processing device and the data processing device. To do.

A first data processing apparatus according to the present invention includes a plurality of processing circuits that individually execute data processing in response to first operation instructions that are individually set with data, and a plurality of processes that correspond to first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control circuit connection relations are arranged in a matrix, and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to the second operation command A context including a first operation instruction for each operation state in which a plurality of processing circuits and a plurality of wiring circuits are sequentially switched from a series of source codes in which the operation of the parallel processing device separately provided is described in a high-level language. A data processing device that generates data of a series of object codes including at least two operation instructions, at least a physical structure and physical characteristics of a parallel processing device The condition storage means in which the constraint conditions corresponding to the data are registered in advance, the source input means for inputting the source code data, and the object code for sequentially switching the context for each operation state from the data input source code, The data path corresponding to the matrix circuit part and the finite state machine corresponding to the state management part are separated, and the object generation means for generating data and the data generated object code are output as data so as to satisfy the constraints. Object output means.

A second data processing apparatus according to the present invention includes a plurality of processing circuits that individually execute data processing corresponding to first operation instructions that are individually set with data, and a plurality of processes that correspond to first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control circuit connection relations are arranged in a matrix, and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to the second operation command RTL (Register Transfer Level) of a parallel processing circuit for each operation state in which a plurality of processing circuits and a plurality of wiring circuits are sequentially switched from a series of source codes in which the operation of a parallel processing device separately provided is described in a high-level language A data processing device for generating a description of data, in which a constraint description corresponding to at least the physical structure and physical characteristics of the parallel processing device is registered in advance. Means, source input means for inputting source code data, and RTL description for sequentially switching the context for each operation state from the inputted source code, corresponding to the data path corresponding to the matrix circuit section and the state management section RTL generating means for generating data and RTL output means for outputting data of the generated RTL description are provided so as to satisfy the constraint condition by separating from the finite state machine.

A third data processing apparatus according to the present invention includes a plurality of processing circuits that individually execute data processing corresponding to first operation instructions set individually and a plurality of processes corresponding to the first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control circuit connection relations are arranged in a matrix, and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to the second operation command Is a data processing device that generates data from a RTL description of a parallel processing device that is separately provided for each operation state in which a plurality of processing circuits and a plurality of wiring circuits are sequentially switched. The condition storage means in which the constraint conditions corresponding to at least the physical structure and the physical characteristics of the parallel processing device are registered in advance, and the RTL description is the data input Separating the data path corresponding to the matrix circuit part and the finite state machine corresponding to the state management part from the RTL input means to be inputted and the netlist for sequentially switching the context for each operation state from the RTL description inputted with data, In order to satisfy the constraint conditions, a net list generating unit for generating data and a net list output unit for outputting the data generated net list are provided.

A fourth data processing apparatus according to the present invention includes a plurality of processing circuits that individually execute data processing corresponding to first operation instructions that are individually set with data, and a plurality of processes that correspond to first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control circuit connection relations are arranged in a matrix, and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to the second operation command At least a context including a first operation instruction and a second operation instruction for each operation state in which a plurality of processing circuits and a plurality of wiring circuits are sequentially switched from a net list of circuit resources included in the parallel processing device separately including A data processing device that generates data for a series of included object codes, and is a constraint corresponding to at least the physical structure and physical characteristics of the parallel processing device Matrix circuit unit including condition storage means in which data is registered in advance, netlist input means for netlist data input, and object code for sequentially switching contexts for each operation state from the data input netlist An object generation means for generating data and an object output means for outputting data of the generated object code so as to satisfy a constraint condition by separating a data path corresponding to the finite state machine and a finite state machine corresponding to the state management unit And.

Data processing method of the present invention, the condition storage means, the source input means, the object generation means, and a data processing apparatus having an object output means is a data processing method for generating object code, the condition storage means, individually A plurality of processing circuits that individually execute data processing in response to a first operation command that is set with data, and a plurality of wirings that individually switch and control connection relationships between the plurality of processing circuits in response to the first operation command At least a physical structure of a parallel processing device separately including a matrix circuit unit in which circuits are arranged in a matrix and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle corresponding to the second operation command and has physical properties and constraints corresponding to is the data registered, the source input means, serial in operation a high-level language parallel operation device A series of source code and a step of data input, the object generation means, from the source code data input from the first operation instruction for each successively cut switched operating state of a plurality of processing circuits and a plurality of printed circuit A series of object codes that include at least a context and a second operation instruction, and sequentially switch contexts for each operation state, by separating a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit The step of generating data so as to satisfy the constraint conditions, and the step of outputting the object code generated by the object output means are provided.

Data processing method, the condition storage means, a data processing method source input means, RTL generation means, and the data processing apparatus provided with RTL output means generates the RTL description, the condition storage means, are individually data set A plurality of processing circuits that individually execute data processing in response to the first operation command and a plurality of wiring circuits that individually switch and control connection relations between the plurality of processing circuits in response to the first operation command. At least a physical structure and physical characteristics of a parallel processing device separately including a matrix circuit unit arranged in matrix and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to the second operation instruction DOO and the corresponding constraint is data registered in the source input means, a set of source code that operation of the parallel arithmetic unit written in a high-level language A step of over data input, RTL generation means, from the source code data inputted, consists of a first operation instruction for each successively cut switched operating state of a plurality of processing circuits and a plurality of printed circuit context and the second A series of RTL descriptions that include at least operation instructions and sequentially switch contexts for each operation state are separated from the data path corresponding to the matrix circuit portion and the finite state machine corresponding to the state management portion, and the constraint conditions are satisfied. as such, the steps of data generation, the RTL output means, comprising the steps of data output RTL description is data generated, a.

Data processing method, the condition storing means, RTL input means, the net list generating means, and a data processing method in which the data processing apparatus comprising a netlist output means generates a netlist, in the condition storage means, each data A plurality of processing circuits that individually execute data processing corresponding to the set first operation command and a plurality of wiring circuits that individually switch and control connection relations between the plurality of processing circuits corresponding to the first operation command And at least a physical structure of a parallel processing unit separately including a matrix circuit unit in which the matrix circuit unit is arranged in matrix and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle corresponding to the second operation instruction constraint condition corresponding to the physical properties are data registration, RTL input means, a series of operations of the parallel arithmetic unit written in a high-level language A step of data enter the TL description, the net list generating means, the RTL description is a data input, sequentially switches the context for each successively cut switched operating state of a plurality of processing circuits and a plurality of wiring circuits, a parallel arithmetic unit A step of generating data so as to satisfy a constraint condition by separating a data path corresponding to a matrix circuit unit and a finite state machine corresponding to a state management unit from a net list of circuit resources provided; and a net list output unit There are provided a step of data outputting netlist data generated, a.

Data processing method of the present invention is a condition storage means, the netlist input means, the object generation means, and a data processing method in which data processing apparatus having an object output means generates an object code, a condition storage means, the individual A plurality of processing circuits that individually execute data processing in response to a first operation command set in the data and a plurality of processing circuits that individually switch and control connection relations between the plurality of processing circuits in response to the first operation command. At least a physical unit of a parallel processing device comprising: a matrix circuit unit in which wiring circuits are arranged in a matrix; and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to the second operation command. structure and physical characteristics and constraints corresponding to are data registered netlist input means, the operation of the parallel arithmetic unit A step of a sequence of data input to the netlist described in grade language, the object generation means, from the netlist data input, for each successively cut switched operating state of a plurality of processing circuits and a plurality of wiring circuits first A series of object codes including at least a context including the operation instructions and a second operation instruction, and sequentially switching the context for each operation state, a data path corresponding to the matrix circuit unit, and a finite state machine corresponding to the state management unit was separated, so as to satisfy the constraint conditions, the process of data generation, the object output means, comprising the steps of data output object code data generated, a.

The computer program of the present invention has a connection relationship between a plurality of processing circuits that individually execute data processing corresponding to a first operation instruction that is individually set with data and a plurality of processing circuits that correspond to the first operation instruction. A matrix circuit section in which a plurality of wiring circuits that individually control switching are arranged in a matrix, and a state management section that sequentially switches the context of the matrix circuit section for each operation cycle in response to the second operation command A context and a second operation instruction including a first operation instruction for each operation state in which a plurality of processing circuits and a plurality of wiring circuits are sequentially switched from a series of source codes in which the operation of the parallel arithmetic device is described in a high-level language the set of object code that is at least contained to data generation, and the computer, at least the physical structure and physical JP parallel operation device Corresponding to the matrix circuit section, the condition storage means for registering the constraint conditions corresponding to the data, the source input means for inputting the source code data, and the object code for sequentially switching the context for each operation state from the data input source code Data processing comprising: an object generation means for generating data by separating a data path and a finite state machine corresponding to the state management unit, referring to a constraint condition, and an object output means for outputting the generated object code as data It is a computer program for functioning as a device .

The computer program of the present invention has a connection relationship between a plurality of processing circuits that individually execute data processing corresponding to a first operation instruction that is individually set with data and a plurality of processing circuits that correspond to the first operation instruction. A matrix circuit section in which a plurality of wiring circuits that individually control switching are arranged in a matrix, and a state management section that sequentially switches the context of the matrix circuit section for each operation cycle in response to the second operation command from a series of source code operation of the parallel arithmetic unit written in a high-level language, RTL description of a parallel processing circuit for each successively cut switched operating state of a plurality of processing circuits and a plurality of wiring circuits for data generation, co the computer, condition storage means for data registration constraint condition corresponding to at least the physical structure and physical properties of the parallel arithmetic unit, data source RTL description that sequentially switches context for each operation state from the source input means that inputs data, separates the data path corresponding to the matrix circuit part and the finite state machine corresponding to the state management part, A computer program for causing a data processing apparatus to function as an RTL generation unit that generates data and an RTL output unit that outputs data generated RTL descriptions so as to satisfy the constraint conditions.

The computer program of the present invention has a connection relationship between a plurality of processing circuits that individually execute data processing corresponding to a first operation instruction that is individually set with data and a plurality of processing circuits that correspond to the first operation instruction. A matrix circuit section in which a plurality of wiring circuits that individually control switching are arranged in a matrix, and a state management section that sequentially switches the context of the matrix circuit section for each operation cycle in response to the second operation command from RTL description of the parallel arithmetic unit, for each successively cut switched operating state of a plurality of processing circuits and a plurality of wiring circuits, a netlist of a circuit resources provided in the parallel arithmetic unit for data generation, the computer, the parallel operation Condition storage means for registering constraints corresponding to at least the physical structure and physical characteristics of the device, RTL input for inputting RTL description data Means, from the RTL description that has been input data, the netlist that sequentially switches the context for each operation state, the data path corresponding to the matrix circuit part and the finite state machine corresponding to the state management part are separated, and the constraint conditions are satisfied As described above, the present invention is a computer program for functioning as a data processing apparatus including a net list generation unit that generates data and a net list output unit that outputs data of the generated net list.

The computer program of the present invention has a connection relationship between a plurality of processing circuits that individually execute data processing corresponding to a first operation instruction that is individually set with data and a plurality of processing circuits that correspond to the first operation instruction. A matrix circuit section in which a plurality of wiring circuits that individually control switching are arranged in a matrix, and a state management section that sequentially switches the context of the matrix circuit section for each operation cycle in response to the second operation command A series including at least a context including a first operation instruction and a second operation instruction for each operation state in which a plurality of processing circuits and a plurality of wiring circuits are sequentially switched from a net list of circuit resources included in the parallel arithmetic device. to the object code data generation, control of the computer, corresponding to at least the physical structure and physical properties of the parallel arithmetic unit Condition storage means for registering condition data, netlist input means for inputting netlist data, object code for sequentially switching contexts for each operation state from the data input netlist, data paths and states corresponding to the matrix circuit section As a data processing apparatus comprising an object generation means for generating data and an object output means for outputting data of the generated object code so as to satisfy a constraint condition by separating from a finite state machine corresponding to the management unit It is a computer program for making it function.

The information storage medium of the present invention stores the above computer program as data .

The data processing system according to the present invention includes the first or fourth data processing apparatus, a plurality of processing circuits that individually execute data processing in response to a first operation command for which data is individually set, and a first processing circuit. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control connection relations of a plurality of processing circuits corresponding to the operation command are arranged in a matrix, and a context of the matrix circuit unit corresponding to the second operation command A state management unit that sequentially switches for each operation cycle; and a parallel processing unit that includes a storage unit in which the object code output by the first or fourth data processing device is stored.

  The various means referred to in the present invention need only be formed so as to realize the function. For example, dedicated hardware that exhibits a predetermined function, a data processing apparatus provided with a predetermined function by a computer program A predetermined function implemented in the data processing apparatus by a computer program, a combination thereof, or the like is permitted. Further, the various means referred to in the present invention do not have to be individually independent, and one means is allowed to be a part of another means.

  The information storage medium referred to in the present invention may be hardware in which a computer program for causing the data processing apparatus to execute various processes is stored in advance. Fixed ROM (Read Only Memory), HDD (Hard Disc Drive), CD (Compact Disc) -ROM, FD (FLEXIBLE Disc), etc. that can be exchangeably loaded in a device that includes a data processing device. Allow.

  The data processing apparatus referred to in the present invention may be any hardware that can read computer programs and execute corresponding data processing. For example, a CPU (Central Processing Unit) is mainly used as a ROM or RAM. (Random Access Memory) and hardware to which various devices such as I / F (Interface) units are connected are allowed.

  In the present invention, causing the data processing apparatus to execute various operations corresponding to the computer program also allows the data processing apparatus to control operations of the various devices. For example, storing various data in the data processing device means storing the various data in an information storage medium such as a RAM provided as a part of the data processing device or loading the data processing device in a replaceable manner. It is allowed to store various data in an information storage medium such as an FD.

  In the data processing device of the present invention, constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel arithmetic device are registered in advance, and the data is input according to the constraint conditions from the source code of the operation description of the parallel arithmetic device. Data is generated for a series of object codes that sequentially switch the context of the parallel processing device for each operation cycle, and by outputting the generated series of object codes, parallel computation is performed from the source code of the behavior description of the parallel processing device. Data can be generated for object code that sequentially switches the device context for each operation cycle.

  As another form of the present invention, a data path corresponding to the matrix circuit part and a finite state machine corresponding to the state management part are separated when generating a series of object codes from a series of source codes due to constraints. By doing so, it is possible to generate object codes corresponding to the matrix circuit unit and the state management unit of the parallel processing device separately.

  In addition, the DFG data is generated by language analysis of a series of source codes, and a CDFG in which a plurality of stages of operation states are scheduled according to the constraint condition is generated from the generated DFG. From the source code, it is possible to generate data of a CDFG that schedules a plurality of stages of operation states that sequentially transition.

  Further, the CDFG is updated so that the state management unit updates the data of the context in which the corresponding operation is executed while the corresponding operation is being executed in one of the predetermined number of contexts stored in the matrix circuit unit by data. When the parallel arithmetic unit is operated corresponding to the object code generated from the CDFG by generating the data, the matrix circuit unit stores a large number of contexts of the object code to a predetermined number and sequentially executes the corresponding operation. In the meantime, the state management unit can update the data of the context in which the corresponding operation has been executed to a new context.

  In addition, by separating the RTL description of the operation state of a plurality of stages from the generated CDFG into a matrix circuit unit and a state management unit according to the constraint condition, data is generated, so that the matrix circuit unit and the state management unit of the parallel processing device It is possible to generate RTL descriptions that individually correspond to the data.

  In addition, a plurality of stages in which the processing circuit netlist of the parallel processing device is sequentially shifted by generating data from the RTL description of the generated matrix circuit unit for each operation state in a plurality of stages according to the constraint condition. Data can be generated for each operation state.

  In addition, by generating a netlist of a plurality of types of circuit resources of the processing circuit from the RTL description of the generated matrix circuit unit for each operation state at a plurality of stages according to constraints, the netlist of the processing circuit of the parallel processing device is generated. Can be generated for each of a plurality of types of circuit resources having different numbers of bits.

  In addition, the object code of the state management unit corresponding to the operation states in a plurality of stages is registered in advance, and the RTL description of the state management unit generated by the data corresponding to the net list of the matrix circuit unit that has generated the data is stored. By converting the data into an object code and outputting the converted object code of the state management unit, the object code of the state management unit can be generated corresponding to the net list of the matrix circuit unit.

  In addition, by generating a net list of m / nb circuit resources of the processing circuit from the RTL description according to constraints, it is possible to generate a net list of processing circuits of the parallel processing device for each m / nb circuit resource.

  Also, by assigning “xb” data processing in the RTL description to Y mb circuit resources “x = Ym + Zn” and Z nb circuit resources, “xb” data processing becomes m / nb circuit resources. Appropriate number can be assigned.

Also, by calculating the integer part “Y” of the result of dividing “x” by “m”, and then calculating the integer “Z” of the result of dividing the remainder by “n”, the “xb” Data processing can be allocated as much as possible to mb circuit resources having a large number of bits, and only the remainder can be allocated to nb circuit resources having a small number of bits.

  Further, when the remainder is smaller than a predetermined threshold, an integer “Z” is calculated, and when the remainder is greater than or equal to the threshold, “1” is added to “Y” to make “Z = 0”, thereby obtaining the mb circuit. The remainder part of the data processing of “xb” allocated to resources as much as possible can be allocated to an appropriate m / nb circuit resource corresponding to the number of bits.

  In addition, since “nb” of the constraint condition in which data is registered is composed of “1 (bit)”, the nb circuit can be used regardless of how many bits are left in the data processing of “xb” allocated to the mb circuit resource. Can be allocated to resources.

  In addition, various object codes of the processing circuit are registered in advance, and a net list generated for each operation state of a plurality of stages is allocated to a plurality of processing circuits arranged in a matrix for each context of a plurality of cycles. By converting the netlist assigned to this processing circuit into the corresponding object code, the object code of the plurality of processing circuits of the parallel processing device can be generated for each context of a plurality of cycles.

  In addition, the object code of the wiring circuit is registered in advance, and by converting the net list of the wiring circuit into the corresponding object code corresponding to the object code of the processing circuit subjected to the data conversion, Data can also be generated for the object codes of a plurality of wiring circuits that individually switch and control the connection relationships of the plurality of processing circuits.

  In addition, by allocating a netlist in which processing data is linked in a continuous multiple cycle context to a register file in a fixed position in the matrix circuit unit, the processing data linked in the continuous multiple cycle context of the parallel processing unit is fixed. Data can be generated for the object code to be temporarily stored in the register file at the position.

  In addition, after assigning a netlist to a register file or the like regardless of the linkage of processing data for each context of multiple cycles, a series of register files to which a netlist to which processing data is linked is assigned by adjusting this allocation position By fixing the position in the matrix circuit part of the register file to which the netlist to which the processing data is linked in the multiple cycle context is first assigned in common, By assigning other netlists to other register files, etc. for each context of multiple cycles, and in the matrix circuit part of the register file to which the netlist to which processing data is linked in the context of multiple cycles is first assigned in common relative position The object code that temporarily holds the associated processing data in the register file at a fixed position by assigning other netlists to other register files etc. for each context of multiple cycles while maintaining this relative position after setting. Data can be generated by predetermined data processing.

  In addition, a plurality of operation states that are sequentially transitioned are detected from a plurality of operation states that are assigned to one context according to a constraint condition, and the plurality of operation states that are detected are assigned to one context, whereby the plurality of operation states are assigned to one context. As described above, data can be set in the parallel arithmetic device at one time.

  In addition, by detecting a plurality of operation states in which transition stages are continuous, it is possible to set data in the parallel computing device as a plurality of continuous operation states as one context.

  In the first parallel processing device of the present invention, the object code generated by the data processing device of the present invention is stored, and the context is sequentially switched for each operation cycle corresponding to the object code. A plurality of processing circuits arranged in a matrix corresponding to the context to be switched operate in parallel with individually configurable data processing, thereby corresponding to the object code generated and stored by the data processing apparatus of the present invention Thus, a plurality of processing circuits can be operated in parallel for each operation cycle.

  In the second parallel processing device of the present invention, externally generated data generated by the data processing device of the present invention is used, and contexts are sequentially switched for each operation cycle corresponding to the object code, and this sequentially switched context is supported. A plurality of processing circuits arranged in a matrix are operated in parallel with individually configurable data processing, so that data is generated by the data processing device of the present invention and each operation cycle corresponds to an object code inputted externally. A plurality of processing circuits can be operated in parallel.

  In the first data processing system of the present invention, the object code generated by the data processing apparatus of the present invention is externally input to the parallel processing apparatus of the present invention, so that the source code input to the data processing apparatus is externally input. Correspondingly, parallel operations can be executed in a context of a plurality of cycles that are sequentially switched to the parallel operation device.

  In the second data processing system of the present invention, the data processing device generates data by separating the object code from the source code into the data path and the finite state machine, and the data path corresponds to the matrix circuit unit of the parallel processing device. In addition, since the finite state machine and the state management unit correspond to each other, the object code can be generated by the data processing device in the software structure corresponding to the hardware structure of the parallel processing device.

  An embodiment of the present invention will be described below with reference to the drawings. First, as shown in FIG. 4 and FIG. 5A, the array processor 100 which is a parallel processing device of this embodiment includes an I / F unit 101 which is an object input unit, a code memory 102 which is a code storage unit, a state A management unit 103, a data distribution unit 104, a matrix circuit unit 105, and a data memory 106 are provided. The matrix circuit unit 105 includes a processor element 107 that is a plurality of processing circuits and a switch element 108 that is a plurality of wiring circuits. A large number of mb buses 109, a large number of nb buses 110, and the like are arranged in a matrix.

  Further, as shown in FIG. 5B, the processor element 107 includes an instruction memory 111, an instruction decoder 112, an mb circuit resource mb register file 113, an nb circuit resource nb register file 114, and an mb circuit resource. Each of the switch elements 108 includes a bus connector 121, an input control circuit 122, and an output control circuit 123. The switch element 108 includes an mb ALU 115 that is an mb arithmetic unit, an nb ALU 116 that is an nb arithmetic unit and an internal wiring resource 117, and the like. , Etc. respectively.

  The I / F unit 101 is connected to a data processing device 300 described later, and a series of object codes are externally input from the data processing device 300. The code memory 102 is composed of an information storage medium such as a RAM, and stores data of externally input object codes.

  As will be described later in detail, this object code is set as data as a context for each operation cycle in which operation instructions of the plurality of processor elements 107 and the plurality of switch elements 108 arranged in matrix in the matrix circuit unit 105 are sequentially switched. The operation command of the state management unit 103 for switching the context for each operation cycle is set as data as an operation state in which transition is made sequentially.

  The state management unit 103 reads the context data for each operation cycle from a series of object codes stored in the code memory 102, and receives instruction pointers for the plurality of processor elements 107 and the plurality of switch elements 108 from the operation instructions. appear.

  In this embodiment, since the instruction memory of the switch element 108 is also used as the instruction memory 111 of the adjacent processor element 107, the state management unit 103 corresponds to a set of instruction pointers of the generated processor element 107 and switch element 108. To the instruction memory 111 of the processor element 107.

  The instruction memory 111 stores a plurality of operation instructions for the processor element 107 and the switch element 108 in advance, and the operation instructions for the processor element 107 and the switch element 108 using two instruction pointers supplied from the state management unit 103. Is specified. The instruction decoder 112 decodes the operation instruction designated by the instruction pointer, and controls the operation of the switch element 108, the internal wiring resource 117, the m / nb ALU 115, 116, and the like.

  Since the mb bus 109 transmits mb data which is “8 (bit)” processing data and the nb bus 110 transmits nb data which is “1 (bit)” processing data, the switch element 108 is an instruction decoder. Corresponding to the operation control 112, the connection relation of the plurality of processor elements 107 by the m / nb buses 109 and 110 is controlled.

  More specifically, in the bus connector 121 of the switch element 108, the mb bus 109 and the nb bus 110 communicate with each other from four directions, and communicate with the mutual connection relationship of the plurality of mb buses 109 communicating in this way. The mutual connection relationship of the plurality of nb buses 110 is controlled.

  For this reason, in the array type processor 100 of this embodiment, the state management unit 103 sequentially switches the context of the matrix circuit unit 105 for each operation cycle corresponding to the object code set in the code memory 102, and a plurality of units are provided for each stage. The processor elements 107 operate in parallel with individually configurable data processing.

  Here, in order to simplify the description, a structure in which each part is arranged in a planar shape is illustrated. However, in practice, the above-described m / nb buses 109 and 110, the switch element 108, the processor element 107, It is also possible to form the layers in a stacked structure.

  The input control circuit 122 controls the data input connection relationship from the mb bus 109 to the mb register file 113 and the mb ALU 115 and the data input connection relationship from the nb bus 110 to the nb register file 114 and the nb ALU 116. 123 controls the data output connection relationship from the mb register file 113 or mb ALU 115 to the mb bus 109 and the data output connection relationship from the nb register file 114 or nb ALU 116 to the nb bus 110.

  The data distribution unit 104 distributes a series of externally input processing data into mb data and nb data, and a plurality of m / nb data from the m / nb buses 109 and 110 whose connection relations are controlled by the switch element 108. Are appropriately input to a specific part of the processor element 107.

  The internal wiring resource 117 of the processor element 107 corresponds to the connection relationship between the mb register file 113 and the mbALU 115 and the connection relationship between the nb register file 114 and the nbALU 116 in the processor element 107 in response to the operation control of the instruction decoder 112. Control.

  The mb register file 113 temporarily holds mb data input from the mb bus 109 or the like and outputs it to the mbALU 115 or the like corresponding to the connection relation controlled by the internal wiring resource 117. The nb register file 114 temporarily holds the nb data input from the nb bus 110 or the like and outputs it to the nbALU 116 or the like corresponding to the connection relation controlled by the internal wiring resource 117.

  The mbALU 115 executes data processing corresponding to the operation control of the instruction decoder 112 with mb data, and the nbALU 116 executes data processing corresponding to the operation control of the instruction decoder 112 with nb data. The processor element 107 reads the m / nb data with the built-in m / nb register files 113 and 114 as described above, but also reads the processing data with the external data memory 106 if necessary.

  As shown in FIG. 1, the data processing system 200 according to the present embodiment, in which the array processor 100 according to the present embodiment is a part, includes an array processor 100, a data processing device 300, a source supply unit 201, an object supply unit 202, Data input means 203 and the like are provided.

  As shown in FIG. 3, the data processing apparatus 300 according to the present embodiment includes a CPU 301 as hardware serving as a computer. The CPU 301 includes a ROM 303, a RAM 304, an HDD 305, and an FD 306 via a bus line 302. Hardware such as an FDD (FD Drive) 307 in which a CD-ROM 308 is interchangeably loaded, a CD drive 309 in which a CD-ROM 308 is interchangeably loaded, a keyboard 310, a mouse 311, a display 312, and an I / F unit 313 are connected. ing.

  In the data processing apparatus 300 of this embodiment, hardware such as ROM 303, RAM 304, HDD 305, exchangeable FD 306, exchangeable CD-ROM 308, etc. corresponds to an information storage medium, and at least one of these is a computer program for the CPU 301. And various data are stored as software.

  For example, computer programs that cause the CPU 301 to execute various data processing are stored in advance in the FD 306 or the CD-ROM 308. Such software is installed in the HDD 305 in advance, and is copied to the RAM 304 and read by the CPU 301 when the data processing apparatus 300 is activated.

  As described above, when the CPU 301 reads an appropriate computer program and executes various data processing, the data processing apparatus 300 according to the present embodiment has a source input unit 211, a condition storage unit 212, Various means such as an object generation means 213 and an object output means 214 are logically provided as various functions.

  The source supply unit 201 of the data processing system 200 includes, for example, an FD 306 in which source code is stored, and a series of source codes in which the array-type processor 100 is described in a high-level language such as C language as a source of the data processing apparatus 300 Supply to the input means 211.

  The source input unit 211 of the data processing apparatus 300 corresponds to the function of the CPU 301 controlling the operation of the FDD 307 corresponding to the computer program stored in the RAM 304, and the source code is input from the source supply unit 201.

  The condition storage unit 212 corresponds to a storage area of the HDD 305 that the CPU 301 recognizes data corresponding to the computer program described above, and various constraint conditions corresponding to the physical structure and physical characteristics of the array processor 100 are preliminarily stored in the data. It is registered.

  The object generation unit 213 corresponds to a function of the CPU 301 executing predetermined data processing corresponding to a computer program. The object generation unit 213 is an array type according to the constraint condition of the condition storage unit 212 from the source code input to the source input unit 211. A series of object codes for sequentially switching the context of the matrix circuit unit 105 of the processor 100 is generated.

  As described above, since the state management unit 103 and the matrix circuit unit 105 are separately formed in the array type processor 100, the object generation unit 213 generates the object code from the source code according to the constraints reflecting this structure. When data is generated, as shown in FIG. 4, the data path corresponding to the matrix circuit unit 105 and the finite state machine corresponding to the state management unit 103 are separated.

  As will be described in detail later, the object generation unit 213 generates object code data from the source code through various multi-stage processes, and appropriately refers to necessary constraint conditions from the condition storage unit 212 in the multi-stage various processes. Therefore, various constraint conditions are set in the condition storage unit 212 with data contents necessary and sufficient for various processes of the object generation unit 213.

  The object output unit 214 corresponds to a function of the CPU 301 controlling the data output of the I / F unit 313 corresponding to a computer program, and a series of object codes generated by the object generation unit 213 are stored in the data processing system 200. Data is output to the object supply means 202.

  The object supply unit 202 corresponds to a connection connector (not shown) that connects the I / F unit 313 of the data processing device 300 and the state management unit 103 of the array processor 100, and the data processing device 300 outputs data. The object code to be input is input to the I / F unit 101 of the array type processor 100.

  Since the code memory 102 is connected to the I / F unit 101 of the array type processor 100 as described above, the object code is held in the code memory 102. The data input unit 203 includes, for example, a predetermined data generation circuit (not shown), and generates a series of processing data and inputs it to the data distribution unit 104 of the array type processor 100.

  As shown in FIG. 2, the object generation unit 213 of the data processing apparatus 300 according to the present embodiment includes a graph generation unit 221, a schedule generation unit 222, a description generation unit 223, a list generation unit 224, an object storage unit 225, and a state conversion unit 226. , Processing arrangement means 227, code conversion means 228, wiring conversion means 229, and the like.

  The graph generation unit 221 performs language analysis on the series of source codes described in the above-described C language or the like to generate DFG data, and the schedule generation unit 222 generates predetermined constraints from the DFG generated by the graph generation unit 221. According to the conditions, as shown in FIG. 6, data is generated for a CDFG that schedules a plurality of stages of operation states in which the array type processor 100 sequentially changes.

  More specifically, as shown in FIG. 7, for example, when a plurality of data processes in which the processing data is linked are described in the source code, the plurality of data processes correspond to the causal relationship as shown in FIG. 8. As a result, the DFG linked together is generated by an existing method.

  When CDFG data is generated from such DFG, the number of processor elements 107 of the array type processor 100, the delay time of the m / nb data buses 109 and 110, etc. are referred to as constraints, as shown in FIG. Thus, in response to this, a plurality of DFG data processes are sequentially scheduled in order to a plurality of stages of operating states.

  As described above, when a plurality of nodes of the DFG are scheduled in order, if the number of the plurality of data processes executed in parallel at one time in one operation state exceeds the number of processor elements 107 which is a constraint condition, Excess data processing is shifted to the next operation state.

  In addition, when the cumulative time of multiple data processes that are continuously executed at one time in one operating state exceeds the clock interval of the operating state, which is a constraint, the data processing from the data processing exceeding that clock interval to the next Transition to stage operating state. Similarly, if multiple data processes that are executed continuously in a certain operating state include data processes that cross operation cycles, such as port input / output, memory access, and multi-cycle operations, the data processing occurs. At this point, the operating state is shifted to the next stage.

  Here, the scheduling for shifting the operation state by comparing the number of operations described above with the number of processor elements 107 will be specifically described below. First, the array type processor 100 of this embodiment has only two processor elements 107, and a conventional array type processor (not shown) to be compared has one adder and one subtractor. Assume that

  In this case, the conventional array type processor can execute one addition process and one subtraction process in parallel in one operation state, but can execute two addition processes and two subtraction processes in parallel in one operation state. I can't. On the other hand, the array type processor 100 according to the present embodiment can execute two data processes in parallel in one operation state, and an addition process, a subtraction process, and the like can be freely set as the two data processes.

  Then, as shown in FIG. 8, when two addition processes exist in parallel in the DFG and two subtraction processes using the output data of these addition processes as input data exist in parallel, the conventional array processor is used. In the scheduling process, as shown in FIG. 9B, the first addition process is executed in the first state, the first subtraction process and the second addition process are executed in parallel in the second state, CDFG data is generated so as to execute the second subtraction process in the three states.

  On the other hand, in the scheduling process to the array type processor 100 according to the present embodiment by the data processing apparatus 300 according to the present embodiment, the first addition is performed by the two processor elements 107 in the first state as shown in FIG. CDFG data is generated so that the processes are executed in parallel and the first and second subtraction processes are executed in parallel in the second state.

  Further, the case where the CDFG is scheduled corresponding to the port input as a constraint condition and scheduled according to the delay time as the constraint condition will be briefly described below. First, when there is only one port input in one operation state as a constraint condition, if a plurality of port inputs are described in the source code as shown in FIG. 10, the source code is shown in FIG. The first row and third row port inputs are scheduled to separate operating states in the CDFG.

  Also, if the delay time allowed for one operating state is “30 (nsec)”, the time required for addition processing is “10 (nsec)”, and the time required for multiplication processing is “20 (nsec)”, As shown in FIG. 10, when one addition process and two multiplication processes are described in the second line of the source code, as shown in FIG. 11, the one addition process and the first multiplication process Is scheduled in one operating state, but the second multiplication process is scheduled in the next operating state.

  In the array type processor 100 of this embodiment, the processor element 107 actually includes only the 8-bit ALU 115 and the 1-bit ALU 116 as the arithmetic unit as described above. Due to the constraints, four 8-bit ALUs 115 are executed in parallel.

  In this case, four are subtracted from the number of 8-bit ALUs 115 that are the constraint conditions corresponding to one 32-bit data process scheduled in one operation state. The data processing for one bit is shifted to the operation state of the next stage.

  In this embodiment, each of the processor elements 107 includes the m / nb ALUs 115 and 116, so that the arithmetic processing is distributed and scheduled to the mb ALU 115 and the nb ALU 116 according to the constraint conditions corresponding thereto. Furthermore, when there are circuit resources for executing data processing in addition to the m / nb ALUs 115 and 116 in the processor element 107, these are also incorporated into the constraint conditions and used for scheduling.

  The description generation unit 223 generates data by separating RTL descriptions of a plurality of stages of operation states into the matrix circuit unit 105 and the state management unit 103 from the CDFG generated by the schedule generation unit 222 according to predetermined constraints. At this time, the processing capabilities of the processor element 107 and the switch element 108 are referred to as constraints, and as shown in FIGS. 9A and 12, m / nb ALUs 115 and 116 for executing data processing and data transfer are allocated. It is done.

  However, the allocation at this point is not limited to which process is executed by which type of circuit resource of the array processor 100, and which data process is executed by which m / nb ALU 115, 116. In other words, the positions of the m / nb register files 113 and 114 that hold processing data to be linked in a plurality of stages of operation states are not determined. Further, in this allocation, a combination that minimizes the circuit resources used for each of the plurality of operation states is generated by trial and error.

  As shown in FIG. 13, the list generation unit 224 generates a net list of the processor element 107 for each of the operation states in a plurality of stages according to a predetermined constraint condition from the RTL description of the matrix circuit unit 105 generated by the description generation unit 223. , Data is generated for each m / nb circuit resource such as m / nb ALU 115,.

  More specifically, when there are more “xb (x− (bit))” data processes than “mb” in the RTL description, the list generation unit 224 obtains the result of dividing “x” by “m”. After calculating the integer part “Y”, the integer “Z” obtained by dividing the remainder by “n” is calculated, and natural numbers “Y” and “Z” satisfying “x = Ym + Zn” are calculated. The data processing of “xb” is allocated to Y mbALU 115 and mb register file 113 and Z nbALU 116 and nb register file 114.

  The object storage means 225 corresponds to a storage area of the HDD 305 that the CPU 301 recognizes data, and various object codes corresponding to a plurality of operation states of the state management unit 103, and the processor elements 107 and switch elements of the matrix circuit unit 105 108 various object codes are registered in advance.

  The state conversion unit 226 converts the RTL description of the state management unit 103 generated by the description generation unit 223 into data corresponding to the object code corresponding to the net list of the matrix circuit unit 105 generated by the list generation unit 224. Convert.

  The processing arrangement unit 227 assigns the net list of the matrix circuit unit 105 generated by the list generation unit 224 for each of the plurality of stages of operation states to the plurality of processor elements 107 arranged in a matrix for each of a plurality of cycles of context.

  At this time, as shown in FIG. 14, the processing placement unit 227 first assigns the netlist to the m / nb register files 113, 114, etc. regardless of the linkage of the processing data for each context of a plurality of cycles. As shown in FIG. 15, by adjusting the positions and matching the positions in the contexts of a plurality of consecutive cycles of the m / nb register files 113 and 114 to which the netlist to which the processing data is linked is assigned, A netlist in which processing data is linked in the context of a cycle is commonly assigned to m / nb register files 113 and 114 at fixed positions in the matrix circuit unit 105.

  In this case, the netlists of the m / nb register files 113 and 114 in which the processing data cooperate in the context of a plurality of consecutive cycles are expressed by giving the same name. In addition, the arrangement of the netlist for each context described above and the arrangement of the netlist for a plurality of contexts are optimal in an iterative process by applying an existing iterative improvement algorithm such as a simulated annealing method. It becomes.

  However, the placement of the netlist by the processing placement unit 227 is executed for each context, and the processing for moving the netlist that cannot be placed in a certain context to another context is not executed. For this reason, when a situation occurs in the processing arrangement unit 227 that cannot arrange all the netlists necessary for a certain context, for example, the processing arrangement unit 227 notifies the schedule generation unit 222 of the data, and the notification data is also restricted. As a condition, the schedule generation unit 222 generates the CDFG again.

  Further, here, in order to simplify the explanation, a case where one m / nb register file 113, 114 temporarily holds only one m / nb data is illustrated, but for example, the m / nb register file 113, 114 In the case where a plurality of m / nb data can be independently read at different addresses 114, if data writing and data reading do not occur simultaneously in one context, one m / nb register file 113, 114 is stored. It is possible to read / write multiple m / nb data in multiple contexts.

  The code conversion unit 228 converts the netlist assigned to the processor element 107 by the processing arrangement unit 227 into a corresponding object code, and the wiring conversion unit 229 converts the object of the processor element 107 that has been converted by the code conversion unit 228. Corresponding to the code, the netlist of the switch element 108 is converted into an object code.

  The various means of the data processing apparatus 300 described above are realized by using hardware such as the FDD 307 and the I / F unit 312 as necessary. The main body corresponds to software stored in an information storage medium such as the RAM 304. This is realized by the function of the CPU 301, which is the hardware of the data processing device 300.

  Such software includes, for example, processing for storing various constraints corresponding to the physical structure and physical characteristics of the array processor 100 in the HDD 305 and the like, and causing the FDD 107 to read data stored in the FD 306 and the like. Processing for inputting the source code of the operation description of the processor 100, processing for generating the object code of the array type processor 100 from the input source code according to the constraint condition, and the generated object code as the I / F unit It is stored in an information storage medium such as the RAM 304 as a computer program for causing the CPU 301 or the like to execute data processing such as data output from 313 or the like.

  Further, the data processing software for generating the object code from the source code described above is described in a process for storing the object code of the state management unit 103 and the object code of the matrix circuit unit 105 in the HDD 305 or the like, C language, or the like. A process of generating a DFG data by linguistically analyzing the series of source codes, and generating a CDFG that schedules a plurality of stages of operation states in which the array processor 100 sequentially transitions according to a predetermined constraint condition from the generated DFG. Processing, processing for generating data by separating RTL descriptions of a plurality of stages of operation states from the generated CDFG into a matrix circuit unit 105 and a state management unit 103 according to a predetermined constraint, and the data generated matrix circuit unit 105 RTL descriptions The process of generating data for each m / nb circuit resource such as m / nb ALUs 115 and 116, the net of the matrix circuit unit 105 in which the data is generated, according to predetermined constraints Processing for converting the RTL description of the state management unit 103 into a corresponding object code corresponding to the list, and a plurality of processors in which the net list of the matrix circuit unit 105 generated for each operation state in a plurality of stages is arranged in a matrix A process of assigning to the element 107 for each context of a plurality of cycles, a process of converting the netlist assigned to the processor element 107 into a corresponding object code, and an object code of the processor element 107 after the data conversion Stored netlists switch element 108 processes data into an object code, an information storage medium RAM304 such as a computer program for executing a data processing CPU301, etc. etc. in response to.

  Note that the computer program as described above is stored in an information storage medium such as the CD-ROM 108 and supplied to the data processing apparatus 300, and the information storage medium includes the constraint conditions of the condition storage unit 212 and the like. The object code of the object storage means 225 can also be stored as structure data.

  With the configuration as described above, in the data processing system 200 of this embodiment, when the source code of the operation description of the array type processor 100 is supplied from the source supply means 201 to the data processing device 300, the data processing device 300 The code is converted into an object code and data is supplied to the array type processor 100.

  Since this array type processor 100 stores the object code supplied with data, when a series of processing data is input from the data input means 203 in this state, the input processing data is converted into data corresponding to the object code. Process.

  More specifically, the data processing apparatus 300 according to the present embodiment includes various constraint conditions corresponding to the physical structure and physical characteristics of the array type processor 100, and various types of state management unit 103, processor element 107, and switch element 108. Object code and data are registered.

  In this state, in the data processing device 300 of this embodiment, as shown in FIG. 16, when a series of source codes in which the array processor 100 is described in C language or the like is input (step S1) As shown in FIGS. 7 and 8, the source code is subjected to language analysis to generate DFG data (step S2).

  Next, data such as the number of processor elements 107 of the array type processor 100 and delay times of the m / nb data buses 109 and 110 are read out as constraints for generating CDFG data from the DFG (step S3). As shown in FIG. 9 (a), data is generated for a CDFG that schedules a plurality of stages of operation states in which the array processor 100 sequentially transitions from the DFG according to this constraint condition (step S4).

  Next, the processing capability of the processor element 107 and the switch element 108 is read out as a constraint condition for generating the RTL description from the CDFG (step S5), and as shown in FIG. M / nb ALUs 115, 116, etc. are assigned to the data processing, and RTL descriptions of a plurality of stages of operation states are separated into the matrix circuit unit 105 and the state management unit 103 to generate data (step S6).

  Next, the number of bits of the m / nb register files 113, 114 and m / nb ALUs 115, 116, etc. are read out as a constraint condition for generating netlist data from the RTL description (step S7), as shown in FIG. In addition, due to this constraint condition, the netlist of the processor element 107 for each of a plurality of operation states is generated for each m / nb circuit resource such as the m / nb ALUs 115 and 116 from the RTL description of the matrix circuit unit 105 (step S8). .

  Thus, when the net list of the matrix circuit unit 105 is generated for each of the operation states in a plurality of stages, the RTL description of the state management unit 103 is correspondingly converted into data corresponding to the object code (step S9). Thus, the object code of the state management unit 103 is generated.

  Next, the net list of the matrix circuit unit 105 in which data is generated for each operation state of a plurality of stages is allocated to a plurality of processor elements 107 arranged in a matrix for each context of a plurality of cycles (step S10). Since the net list assigned to 107 is converted into the corresponding object code (step S11), the object code of the processor element 107 is generated.

  When the object code of the processor element 107 is generated as described above, the net list of the switch element 108 is converted into the object code correspondingly (step S12). Since the object code of the switch element 108 is also generated by this, the object code of the array type processor 100 corresponding to the state management unit 103, the processor element 107, and the switch element 108 is completed (step S13).

  Since the object code generated by the data processing device 300 in this way is set in the array type processor 100 as described above, in this array type processor 100, the state management unit 103 receives the object code input from the outside. The instruction pointer is generated for each processor element 107 and one of the plurality of operation instructions stored in the instruction memory 111 is specified for each processor element 107 by this instruction pointer.

  The designated operation instruction is decoded by the instruction decoder 112, the connection relationship of the plurality of processor elements 107 by the switch element 108, the connection relationship in the processor element 107 by the internal wiring resource 117, and the data processing of the m / nb ALUs 115 and 116 Therefore, the array processor 100 according to the present embodiment is in a hardware state for executing data processing corresponding to the object code.

  In this state, a series of processing data externally input from the data input unit 203 is distributed to the mb data and the nb data by the data distribution unit 104, and the connection relation and the data processing are controlled as described above. Each processor element 107 is subjected to data processing by the mbALU 115 and the nbALU 116.

  As described above, the array processor 100 according to the present embodiment performs data processing of mb data consisting of “8 (bit)” and nb data consisting of “1 (bit)” for each processor element 107 corresponding to the object code. Since the processing can be executed in parallel, even if the number of bits of a series of processing data inputted externally changes, it can be distributed more accurately according to the number of bits and processed in parallel with good efficiency.

  In addition, the array type processor 100 of this embodiment temporarily stores and outputs the m / nb data to which the m / nb register files 113 and 114 formed for each processor element 107 are input. The m / nb data processed by the m / nb ALUs 115 and 116 can be temporarily held by the dedicated m / nb register file 114 without waste.

  Moreover, in the array type processor 100 of this embodiment, since “nb” consists of “1 (bit)”, the nb ALU 116 can execute data processing with the minimum number of bits, and the number of bits of the processing data changes in various ways. However, data processing can be executed reliably.

  In addition, when the processing data is input to the array type processor 100, the data processing system 200 of this embodiment can supply the source code to the data processing device 300 and input the object code to the array type processor 100. The operating state of the array processor 100 can be switched in various ways in real time.

  In the data processing system 200 of the present embodiment, the array processor 100 includes the matrix circuit unit 105 and the state management unit 103 separately as described above, and the data processing device 300 includes a series of operation descriptions of the array processor 100. A series of object codes are generated by separating the data path and the finite state machine from the source code.

  Since the data path corresponds to the matrix circuit unit 105 of the array type processor 100 and the finite state machine corresponds to the state management unit 103, the software of the data processing device 300 and the array type processor 100 Good consistency with hardware.

  Since the array type processor 100 of this embodiment is fundamentally different in structure and operation from a conventional CPU or the like, its object code structure is also fundamentally different from that of a conventional CPU or the like. However, the data processing apparatus 300 according to the present embodiment can satisfactorily generate the object code of the array processor 100 from the source code in which the operation state of the array processor 100 is described by a general method as described above. Therefore, the array processor 100 of this embodiment can be operated satisfactorily.

  In particular, the data processing apparatus 300 according to the present embodiment generates DFG data from the source code, schedules it to the CDFG for each of the plurality of operation states, and assigns the netlist generated from the CDFG for each context of a plurality of cycles. Data can be generated for object code that sequentially switches the context of the array processor 100 for each operation cycle.

  Moreover, since the finite state machine and the data path are separated when the object code is generated from the source code, the object code of the array type processor 100 having the matrix circuit unit 105 and the state management unit 103 separately is obtained. Data can be generated satisfactorily.

  Further, since the net list of the m / nb circuit resources of the processor element 107 is generated from the RTL description of the matrix circuit unit 105, the object of the array type processor 100 in which the description contents are separated for mb data and nb data. The code can be generated with good data.

  Moreover, since “xb” data processing in the RTL description is assigned to Y mbALUs 115 and Z nbALUs 116 “x = Ym + Zn”, an appropriate number of “xb” data processings are assigned to the m / nb ALUs 115 and 116. be able to.

  In particular, since the integer part “Y” obtained by dividing “x” by “m” is calculated and then the integer “Z” obtained by dividing the remainder by “n” is calculated, the data of “xb” Processing can be assigned as much as possible to the mbALU 115 having a large number of bits, and only the remainder can be assigned to the nbALU 116 having a small number of bits. In addition, since “nb” described above is “1 (bit)”, it can be allocated to the nbALU 116 regardless of how many bits of the data processing portion of “xb” allocated to the mbALU 115 is.

  In addition, since the netlist in which the processing data is linked in the context of a plurality of continuous cycles is commonly allocated to the m / nb register files 113 and 114 at a fixed position in the matrix circuit unit 105, the context of the continuous multiple cycles of the array type processor 100 is allocated. The object code for holding the processing data to be linked in the m / nb register files 113 and 114 at a fixed position can be generated.

  In particular, after assigning a netlist to the m / nb register files 113, 114, etc., regardless of the linkage of processing data for each context of a plurality of cycles, the netlist to which the processing data is linked is assigned by adjusting the allocation position. Since the positions of the m / nb register files 113 and 114 in the context of a plurality of consecutive cycles are matched, an object code for holding the associated processing data in the m / nb register files 113 and 114 at a predetermined position is set as a predetermined code. Data can be generated by data processing.

  Since the computer program for generating object code from the source code of the data processing apparatus 300 of this embodiment has a one-to-one correspondence with the array processor 100, information such as the CD-ROM 108 in which the computer program is stored. The storage medium is preferably sold as a set with one or more array type processors 100.

  In that case, since the constraint condition of the condition storage unit 212 and the object code of the object storage unit 225 also correspond to the array type processor 100, the constraint condition and the object code can be stored as data in the information storage medium together with the computer program. Is preferred.

  In addition, when a plurality of types of products are sold as the array type processor 100, one computer program common to the plurality of types of array type processor 100 is stored in a single information storage medium and sold, and a plurality of types of arrays are sold. When the type processor 100 is selectively sold, it is preferable to sell a plurality of types of information storage media in which corresponding constraint conditions and object codes are stored.

  In addition, this invention is not limited to the said form, A various deformation | transformation is accept | permitted in the range which does not deviate from the summary. For example, in the above embodiment, it is exemplified that the data processing apparatus 300 executes a series of processing operations for generating object code data from source code corresponding to one computer program. However, data for generating RTL description data from source code The processing device and computer program, the data processing device and computer program for generating netlist data from the RTL description, and the data processing device and computer program for generating object code data from the netlist may be separated.

  Further, in the above embodiment, when the data processing of “xb” in the RTL description is assigned to Y mbALUs 115 and Z nbALUs 116 “x = Ym + Zn”, the integer part “x” obtained by dividing “x” by “m” By calculating the integer “Z” as a result of dividing the remainder by “n” after calculating “Y”, the data processing of “xb” is allocated to the mbALU 115 having a large number of bits as much as possible, and only the remainder part is bit. An example of assigning to a small number of nbALUs 116 has been described.

  However, in this case, when “nb” is “1 (bit)” and “mb” is “32 (bit)”, data processing of “30 (bit)” is executed by 30 nbALUs 116. Will occur. Therefore, in order to eliminate this, after setting a predetermined threshold value for assigning the data processing to the m / nb ALUs 115 and 116 and calculating the integer part “Y” as a result of dividing “x” by “m”, It is preferable to calculate an integer “Z” when the remainder is smaller than the threshold, and to add “1” to “Y” to make “Z = 0” when the remainder is equal to or greater than the threshold.

  In this case, the remainder of “xb” data processing allocated to the mbALU 115 as much as possible can be allocated to one of the m / nb ALUs 115 and 116 corresponding to the number of bits. For example, “nb” is “1”. (bit) ”and“ mb ”is“ 32 (bit) ”, data processing of“ 9 (bit) ”or more is executed by the mbALU 115, and data processing of“ 8 (bit) ”or less is executed by the nbALU 116. It becomes possible.

  Further, in the above embodiment, the array type processor 100 is exemplified to have two types of circuit resources, “mb” consisting of “8 (bit)” and “nb” consisting of “1 (bit)”. The number of types of circuit resources and the number of bits can be set in various ways.

  Even in such a case, the data processing apparatus 300 can generate an appropriate object code for the array processor 100 by registering appropriate constraint conditions as data. For example, if the array type processor has only the mbALU 115, the integer part “Y” of the result of dividing “x” by “m” is calculated, and if the remainder does not exist, the data processing of “xb” is performed. It is preferable to allocate to Y mbALUs 115 such that “x = Ym”, and to allocate (y + 1) mbALUs 115 where “x≈ (Y + 1) m” if there is a remainder.

  In the above embodiment, the processor element 107 of the array type processor 100 is exemplified to execute data processing by the m / nb ALUs 115 and 116. However, other circuit resources for executing data processing may be formed in the processor element 107. Is possible.

  For example, a circuit resource called DMU (Data Manipulation Unit) for executing data manipulation of mb data and nb data by at least a mask circuit can be formed in the processor element 107. Such an array type processor 100 is disclosed in the present application. A person has filed as Japanese Patent Application No. 2001-263804. Even in such a case, the data processing device 300 can appropriately generate the object code of the array type processor 100 by setting the constraint condition data corresponding to the circuit resource.

  Further, in the above embodiment, the array processor 100 is exemplified to execute data processing only by the processor elements 107 arranged in the matrix circuit unit 105. However, a dedicated arithmetic circuit may be formed outside the matrix circuit unit 105. Is possible.

  For example, since the circuit scale of a multiplier is generally excessive, it can be arranged outside the matrix circuit unit 105 and accessed from the processor element 107 as necessary. Even in such a case, the data processing device 300 can appropriately generate the object code of the array type processor 100 by setting the constraint condition data corresponding to the circuit resource. However, when the processor element 107 accesses an external circuit of the matrix circuit unit 105 as described above, this becomes a restriction condition for state transition when scheduling the DFG to the CDFG.

  Furthermore, in the above embodiment, the array type processor 100 including the m / nb ALUs 115 and 116 in which the processor element 107 as a processing circuit is an arithmetic unit is illustrated as a parallel arithmetic device. DRL (Dynamic Reconfigurable Logic) made up of).

  In the array-type processor 100 of the above embodiment, the hardware configuration is changed by updating the operation instruction to set data in the processor element 107, but in the DRL, the hardware configuration is changed by updating the data registered in the LUT. Therefore, it is possible for the data processing device 300 to generate data for an object code that realizes this.

  In the above embodiment, the instruction memory of the switch element 108 is also used as the instruction memory 111 of the adjacent processor element 107, and the state management unit 103 corresponds to a processor corresponding to a set of instruction pointers of the processor element 107 and the switch element 108. The supply to the instruction memory 111 of the element 107 has been exemplified.

  However, separately from the instruction memory 111 of the processor element 107, the switch element 108 has its own instruction memory, and the state management unit 103 switches the instruction pointer between the processor element 107 and the switch element 108 to the corresponding processor element 107 and switch. It is also possible to supply separately to the instruction memory 111 and the like with the element 108.

  Further, the instruction memory of the processor element 107 is also used as the instruction memory of the switch element 108, and the state management unit 103 supplies a set of instruction pointers of the processor element 107 and the switch element 108 to the instruction memory of the corresponding switch element 108. It is also possible.

  Furthermore, in the above embodiment, it is exemplified that only one state management unit 103 is formed in the array type processor 100, but it is also possible to manage a predetermined number of processor elements 107 by using a plurality of the state management units 103. In such a case, it is preferable that one representative of the plurality of state management units 103 performs integrated management, or that a higher-level management unit (not shown) that integrally manages the plurality of state management units 103 is formed.

  Further, in the above embodiment, since the netlist in which the processing data is linked in the context of a plurality of continuous cycles is assigned in common to the m / nb register files 113 and 114 at a fixed position in the matrix circuit unit 105, first, for each context of the plurality of cycles. Regardless of the linkage of the processing data, the netlist is assigned to the m / nb register files 113, 114 and the m / nb register files 113, 114 to which the assignment is adjusted and the netlist to which the processing data is linked are assigned. It is illustrated that the positions in the context of a plurality of consecutive cycles are matched.

  However, after fixing the position in the matrix circuit unit 105 of the m / nb register files 113 and 114 to which the netlist to which the processing data is linked in the context of a plurality of cycles is first allocated, the other netlist is transferred to the other It is also possible to assign to the m / nb register files 113, 114, etc. for each context of a plurality of cycles.

  In addition, the relative position in the matrix circuit unit 105 of the m / nb register files 113 and 114 to which the netlist to which the processing data is linked is assigned in the context of a plurality of cycles is set first, and then this relative position is maintained. It is also possible to assign other netlists to other m / nb register files 113 and 114 and the like for each context of a plurality of cycles.

  In these cases, the range of use of circuit resources for each context can be reduced, so that as shown in FIG. 17A, even if there are a small number of processor elements 107 arranged in the matrix circuit unit 105, a netlist is provided for each context. Can be assigned.

  Note that, as shown in FIG. 4B, in the conventional FPGA (Field Programmable Gate Array) and ASIC (Application Specific Integrated Circuit), the netlist for each of the multiple cycles of context as described above is allocated to one circuit resource. Therefore, a large number of dedicated circuit resources are required, and once formed hardware cannot be changed, there is no versatility of data processing.

  However, in the array type processor 100 of this embodiment, since the hardware can be changed by the object code, the versatility of data processing is good, and a series of data processing is executed in a time-sharing manner for each of multiple cycles of context. The circuit scale of the arranged circuit resources is reduced.

  Furthermore, in the above-described form, it has been exemplified that the operation states of the sequential stages managed by the state management unit 103 and the context for each operation cycle sequentially switched by the matrix circuit unit 105 have a one-to-one correspondence. It is also possible to detect a plurality assigned to one context from a plurality of stages of operation states according to the constraint condition, and to assign the detected plurality of operation states to one context.

  For example, as shown in FIG. 15, in two consecutive operating states assigned to two consecutive contexts, the positions of the m / nb register files 113 and 114 are the same, but the positions of the m / nb ALUs 115 and 116 are the same. If there is no overlap, as shown in FIG. 18, the two consecutive operation states can be assigned to one context.

  More specifically, as described above, the scheduling of the operation state is executed according to a plurality of types of constraint conditions, and the operation state regardless of the number of allocated m / nb ALUs 115 and 116 when a memory access or the like occurs. Therefore, as shown in FIG. 19A, there may occur an operation state in which the number of assigned m / nb ALUs 115 and 116 is small. When a plurality of operation states generated in this way are assigned to a plurality of contexts on a one-to-one basis, a state in which only a part of the m / nb ALUs 115 and 116 arranged in the array processor 100 operates in a certain context occurs. .

  Therefore, if the allocated number of m / nb ALUs 115 and 116 is sequentially accumulated in a plurality of consecutive operating states, and the context is shifted every time the accumulated number exceeds a predetermined allowable number, as shown in FIG. As many m / nb ALUs 115 and 116 as possible can be assigned to each of the plurality of contexts.

  For example, when the first to third operation states are assigned to the first context and the fifth and sixth operation states are assigned to the third context, the array type processor that operates corresponding to the object program At 100, as shown in FIG. 20, in the first operation cycle, data processing of the first operation state of the first context is executed, and in the second operation cycle, the second operation state of the first context is executed. Data processing is executed, and in the third operation cycle, data processing in the third operation state of the first context is executed.

  In the fourth operation cycle, the first context is switched to the second context and the data processing of the fourth operation state is executed, and in the fifth operation cycle, the second context is switched to the third context. Data processing in the fifth operation state is executed.

  As described above, when a plurality of operation states are assigned to one context, the time required for data processing in the array type processor 100 is the same as when one operation state is assigned to one context. Since the number of contexts to be set is reduced, the data capacity of the object code can be compressed.

  Further, the storage capacity of the code memory 102 that temporarily holds the object code of the array type processor 100 can be reduced, and the number of times of switching the context of the matrix circuit unit 105 by the state management unit 103 can be reduced. can do.

  Further, although the above has exemplified the assignment of two consecutive operation states to one context, it is also possible to assign two or more operation states, as shown in FIG. 21 and FIG. It is also possible to assign the operation state. In this case, the number of context switching of the matrix circuit unit 105 by the state management unit 103 may not be reduced, but the number of contexts set in the object code can be reduced.

  Furthermore, although the object code generated by the data processing device 300 is temporarily held in the array type processor 100 in the above embodiment, for example, the array type corresponding to the object code generated by the data processing device 300 is used. It is not impossible for the processor 100 to operate in real time.

  In the above embodiment, the state management unit 103 switches and controls the context of the matrix circuit unit 105 for each operation cycle corresponding to the object code temporarily stored in the code memory 102. However, the matrix circuit unit 105 It is also possible to store a plurality of contexts and execute a corresponding operation for each operation cycle.

  At this time, the state management unit 103 can update all of the plurality of contexts stored in the matrix circuit unit 105 every multiple cycles. However, the state management unit 103 can update the predetermined number of contexts stored in the matrix circuit unit 105. It is also possible for the state management unit 103 to update the data of the context in which the corresponding operation is executed to a new context while the corresponding operation is being executed. In this case, the matrix circuit unit 105 holds data only for a predetermined number of contexts, but can perform a context corresponding operation continuously without limitation on the number.

  For example, as shown in FIG. 23, when the matrix circuit unit 105 holds the first to sixth contexts and executes the data processing in order, when the fourth context data processing is executed, The management unit 103 can update data in the first to third contexts after the data processing is completed.

  However, as shown in FIG. 22, there is a case where one context is repeatedly used in a plurality of operation cycles. For example, as shown in FIG. 24, the first to sixth data held by the matrix circuit unit 105 are stored. In the case where the data processing of the first context to the third context is executed again after executing the data processing of the contexts in order, the state management unit 103 executes the fourth to fifth contexts in which the data processing is completed. The data can be updated.

  In order to cause the array processor 100 to execute the above-described operation, the corresponding operation is executed while the matrix circuit unit 105 is executing the corresponding operation in one of a predetermined number of contexts stored as data. If the data processing apparatus 300 generates CDFG data so that the state management unit 103 updates the data to a new context, the array processor 100 performs the above-described operation corresponding to the object code generated from the CDFG. Can be executed.

  Furthermore, in the above embodiment, it is exemplified that various means are logically realized as various functions of the data processing device 300 by the CPU 301 operating corresponding to the computer program stored in the RAM 304 or the like. However, each of these various means can be formed as unique hardware, and a part can be stored in the RAM 304 or the like as software and a part can be formed as hardware.

It is a typical block diagram which shows the logical structure of the data processing system of one Embodiment of this invention. It is a typical block diagram which shows the logical structure of a data processor. It is a block diagram which shows the physical structure of a data processor. It is a schematic diagram which shows the correspondence of the description content of a source code, and the hardware of the array type processor which is a parallel arithmetic unit. (a) is a schematic block diagram showing the overall structure of an array type processor, and (b) is a schematic block diagram showing the internal structure of a processor element as a processing circuit. It is a schematic diagram which shows the correspondence of the state transition of CDFG and the context switching of an array type processor. It is a schematic diagram which shows the specific example of a source code. It is a schematic diagram which shows the specific example of DFG. (a) is a schematic diagram showing a specific example of the CDFG of the array type processor of this embodiment, and (b) is a schematic diagram showing a specific example of the CDFG of the conventional array type processor. It is a schematic diagram which shows the specific example of another source code. It is a schematic diagram which shows the specific example of other CDFG. It is a schematic diagram which shows the state by which the circuit resource was allocated to the data processing of CDFG. It is a schematic diagram which shows the state by which the data processing of "xb" is allocated to the circuit resource of "m / nb". It is a schematic diagram which shows the state by which the netlist was allocated for every two continuous contexts. It is a schematic diagram which shows the state which the netlist of two continuous contexts cooperated. It is a flowchart which shows the data processing method by a data processor. (a) is a schematic diagram showing a state in which a netlist is assigned to two contexts of the array type processor of the present embodiment, and (b) is a schematic diagram showing a state in which a netlist is assigned to one surface of a conventional parallel processing device. It is. It is a schematic diagram which shows the state by which the net list | wrist of two continuous operation states was allocated to one context. It is a schematic diagram which shows the correspondence of the operation state by the allocation number of m / nbALU, and a context. It is a schematic diagram which shows the state by which the several continuous operation state was allocated to one context. It is a schematic diagram which shows the correspondence of the operation state by the allocation number of m / nbALU, and a context. It is a schematic diagram which shows the state by which the some continuous operation state was allocated to one context. It is a schematic diagram which shows the state in which one of several continuous contexts is active. It is a schematic diagram which shows the state where several continuous contexts are active.

Explanation of symbols

100 array type processor 103 state management unit 107 processor element 108 switch element 113 mb register resource 114 mb circuit resource nb register file 115 nb circuit resource mb ALU mb circuit resource
116 nbALU which is nb circuit resource
200 Data Processing System 201 Source Supply Unit 202 Object Supply Unit 203 Data Input Unit 211 Source Input Unit 212 Condition Storage Unit 213 Object Generation Unit 214 Object Output Unit 221 Graph Generation Unit 222 Schedule Generation Unit 223 Description Generation Unit 224 List Generation Unit 225 Object Storage means 226 State conversion means 227 Processing arrangement means 228 Code conversion means 229 Wire conversion means 300 Data processing device 301 CPU which is the main body of the computer
303 ROM as an information storage medium
304 RAM as an information storage medium
305 HDD as information storage medium
306 FD as an information storage medium
308 CD-ROM as information storage medium

Claims (18)

Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device comprising: a matrix circuit unit in which a plurality of wiring circuits are arranged in a matrix; and a state management unit that sequentially switches a context of the matrix circuit unit in response to a second operation instruction for each operation cycle. Context including the first operation instruction for each operation state in which a plurality of processing circuits and a plurality of the wiring circuits are sequentially switched from a series of source codes whose operations are described in a high-level language, and the second operation instruction A data processing device for generating data of a series of object codes including at least
Condition storage means in which constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device are registered in advance,
Source input means for inputting the source code data;
The object code for sequentially switching the context for each operation state is separated from the source code inputted with the data by separating a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit. Object generation means for generating data so as to satisfy the constraint conditions;
Object output means for outputting the data of the generated object code;
A data processing apparatus comprising: Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device comprising: a matrix circuit unit in which a plurality of wiring circuits are arranged in a matrix; and a state management unit that sequentially switches a context of the matrix circuit unit in response to a second operation instruction for each operation cycle. RTL (Register Transfer Level) description of the parallel processing circuit is generated for each operation state in which the plurality of processing circuits and the plurality of wiring circuits are sequentially switched from a series of source codes whose operations are described in a high-level language. A data processing device,
Condition storage means in which constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device are registered in advance,
Source input means for inputting the source code data;
The RTL description for sequentially switching the context for each operation state is separated from the source code inputted with the data by separating the data path corresponding to the matrix circuit portion and the finite state machine corresponding to the state management portion. RTL generation means for generating data so as to satisfy the constraint conditions;
RTL output means for outputting the data-generated RTL description;
A data processing apparatus comprising: Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device comprising: a matrix circuit unit in which a plurality of wiring circuits are arranged in a matrix; and a state management unit that sequentially switches a context of the matrix circuit unit in response to a second operation instruction for each operation cycle. A data processing device that generates a netlist of circuit resources included in the parallel processing device for each operation state in which a plurality of processing circuits and a plurality of wiring circuits are sequentially switched from an RTL description,
Condition storage means in which constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device are registered in advance,
RTL input means for inputting data of the RTL description;
The netlist for sequentially switching the context for each operation state is separated from the RTL description inputted with the data, by separating the data path corresponding to the matrix circuit portion and the finite state machine corresponding to the state management portion. Netlist generating means for generating data so as to satisfy the constraint conditions;
Netlist output means for outputting data of the generated netlist;
A data processing apparatus comprising: Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device that separately includes a matrix circuit section in which a plurality of wiring circuits are arranged in a matrix and a state management section that sequentially switches the context of the matrix circuit section in response to a second operation instruction for each operation cycle A series including at least a context including the first operation instruction and the second operation instruction for each operation state in which the plurality of processing circuits and the plurality of wiring circuits are sequentially switched from a net list of circuit resources provided. A data processing device for generating object code data,
Condition storage means in which constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device are registered in advance,
Netlist input means for inputting the netlist data;
The object code for sequentially switching the context for each operation state is separated from the data input netlist by separating a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit. Object generation means for generating data so as to satisfy the constraint conditions;
Object output means for outputting the data of the generated object code;
A data processing apparatus comprising: The constraint conditions include the number of the plurality of processing circuits, the number of input ports of the processing circuit, the presence or absence of usable circuit resources of the processing circuit, the processing capability of the processing circuit or the wiring circuit, and the number of wiring circuits 5. The data processing apparatus according to claim 1, comprising at least one of a delay time, the number of bits of the circuit resource, and a clock interval for switching the operation state. 6. Condition storage unit, a data processing method source input means, the object generation means, and the data processing apparatus comprising an object output means for generating an object code,
The condition storage means includes a plurality of processing circuits that individually execute data processing in response to first operation instructions set individually, and a plurality of processing circuits that correspond to the first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control connection relations are arranged in a matrix; and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to a second operation command; The constraint condition corresponding to at least the physical structure and the physical characteristics of the parallel processing device separately provided is registered as data,
The source input means, the steps of operation of the parallel arithmetic unit for data entry a set of source code written in a high level language,
The object generation means, wherein the data input the source code consists of the first operation command for each sequential cutting switched operating state of the plurality of the processing circuit and a plurality of the wiring circuit context and the second A series of the object code including at least an operation instruction and sequentially switching the context for each operation state, separating a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit, Generating data so as to satisfy the constraints;
The object output means, the steps of data output the object code the is data generated,
A data processing method comprising: Condition storage unit, a data processing method source input means, RTL generation means, and the data processing apparatus provided with RTL output means generates the RTL description,
The condition storage means includes a plurality of processing circuits that individually execute data processing in response to first operation instructions set individually, and a plurality of processing circuits that correspond to the first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control connection relations are arranged in a matrix; and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to a second operation command; The constraint condition corresponding to at least the physical structure and the physical characteristics of the parallel processing device separately provided is registered as data,
The source input means, the steps of operation of the parallel arithmetic unit for data entry a set of source code written in a high level language,
The RTL generation means, wherein the data input the source code consists of the first operation command for each sequential cutting switched operating state of the plurality of the processing circuit and a plurality of the wiring circuit context and the second A series of RTL descriptions that include at least an operation instruction and sequentially switch the context for each operation state, separate a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit, Generating data so as to satisfy the constraints;
The RTL output means for outputting the data-generated RTL description;
A data processing method comprising: Condition storage unit, RTL input means, the net list generating means, and a data processing method in which the data processing apparatus comprising a netlist output means generates a netlist,
The condition storage means includes a plurality of processing circuits that individually execute data processing in response to first operation instructions set individually, and a plurality of processing circuits that correspond to the first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control connection relations are arranged in a matrix; and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to a second operation command; The constraint condition corresponding to at least the physical structure and the physical characteristics of the parallel processing device separately provided is registered as data,
The RTL input means, the steps of operation of the parallel arithmetic unit for data entry a set of RTL descriptions written in high-level language,
The net list generating means, wherein the data input the RTL description, sequentially switching the context for each successively cut switched operating state of the plurality of the processing circuit and a plurality of said wiring circuit, the parallel operation device comprises circuitry Separating the data netlist of resources from a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit, and generating data so as to satisfy the constraints;
The netlist output means, the steps of data output the netlist said the data generated,
A data processing method comprising: Condition storage unit, a data processing method netlist input means, the object generation means, and the data processing apparatus comprising an object output means for generating an object code,
The condition storage means includes a plurality of processing circuits that individually execute data processing in response to first operation instructions set individually, and a plurality of processing circuits that correspond to the first operation instructions. A matrix circuit unit in which a plurality of wiring circuits that individually switch and control connection relations are arranged in a matrix; and a state management unit that sequentially switches the context of the matrix circuit unit for each operation cycle in response to a second operation command; The constraint condition corresponding to at least the physical structure and the physical characteristics of the parallel processing device separately provided is registered as data,
The netlist input means, the steps of data input a series of netlist operation of the parallel arithmetic unit written in a high-level language,
The object generation means, wherein the data input the netlist, composed of the first operation command for each sequential cutting switched operating state of the plurality of the processing circuit and a plurality of the wiring circuit context and the second A series of the object code including at least an operation instruction and sequentially switching the context for each operation state, separating a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit, Generating data so as to satisfy the constraints,
The object output means, the steps of data output the object code the is data generated,
A data processing method comprising: The constraint conditions include the number of the plurality of processing circuits, the number of input ports of the processing circuit, the presence or absence of usable circuit resources of the processing circuit, the processing capability of the processing circuit or the wiring circuit, and the number of wiring circuits The data processing method according to any one of claims 6 to 9, including at least one of a delay time, the number of bits of the circuit resource, and an interval of clocks for switching the operation state. Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device comprising: a matrix circuit unit in which a plurality of wiring circuits are arranged in a matrix; and a state management unit that sequentially switches a context of the matrix circuit unit in response to a second operation instruction for each operation cycle. Context including the first operation instruction for each operation state in which a plurality of processing circuits and a plurality of the wiring circuits are sequentially switched from a series of source codes whose operations are described in a high-level language, and the second operation instruction a set of object code that is at least contained to data generation, and the computer,
Condition storage means for registering data of constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device;
Source input means for inputting the source code as data;
The object code for sequentially switching the context for each operation state is separated from the source code inputted with the data by separating a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit. , An object generation means for generating data referring to the constraint condition,
An object output means for outputting the data of the generated object code;
A computer program for causing a data processing apparatus to function. Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device comprising: a matrix circuit unit in which a plurality of wiring circuits are arranged in a matrix; and a state management unit that sequentially switches a context of the matrix circuit unit in response to a second operation instruction for each operation cycle. operation from a series of source code written in a high level language, RTL description of the parallel processing circuit of each successively cut switched operating state of the plurality of the processing circuit and a plurality of the wiring circuit to data generation, computer The
Condition storage means for registering data of constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device;
Source input means for inputting the source code as data;
The RTL description for sequentially switching the context for each operation state is separated from the source code inputted with the data by separating the data path corresponding to the matrix circuit portion and the finite state machine corresponding to the state management portion. RTL generation means for generating data so as to satisfy the constraint conditions,
RTL output means for outputting the data-generated RTL description;
A computer program for causing a data processing apparatus to function. Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device comprising: a matrix circuit unit in which a plurality of wiring circuits are arranged in a matrix; and a state management unit that sequentially switches a context of the matrix circuit unit in response to a second operation instruction for each operation cycle. from RTL description, for each successively cut switched operating state of the plurality of the processing circuit and a plurality of said wiring circuit, a netlist of a circuit resource the parallel operation device comprises for data generation, the computer,
Condition storage means for registering data of constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device;
RTL input means for inputting data of the RTL description;
The netlist for sequentially switching the context for each operation state is separated from the RTL description inputted with the data, by separating the data path corresponding to the matrix circuit portion and the finite state machine corresponding to the state management portion. , Netlist generation means for generating data so as to satisfy the constraint conditions,
Netlist output means for outputting data of the generated netlist;
A computer program for causing a data processing apparatus to function. Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. A parallel processing device that separately includes a matrix circuit section in which a plurality of wiring circuits are arranged in a matrix and a state management section that sequentially switches the context of the matrix circuit section in response to a second operation instruction for each operation cycle A series including at least a context including the first operation instruction and the second operation instruction for each operation state in which the plurality of processing circuits and the plurality of wiring circuits are sequentially switched from a net list of circuit resources provided. to the object code data generation, the computer,
Condition storage means for registering data of constraint conditions corresponding to at least the physical structure and physical characteristics of the parallel computing device;
Net list input means for inputting data into the net list;
The object code for sequentially switching the context for each operation state is separated from the data input netlist by separating a data path corresponding to the matrix circuit unit and a finite state machine corresponding to the state management unit. Object generation means for generating data so as to satisfy the constraint conditions;
An object output means for outputting the data of the generated object code;
A computer program for causing a data processing apparatus to function. The constraint conditions include the number of the plurality of processing circuits, the number of input ports of the processing circuit, the presence or absence of usable circuit resources of the processing circuit, the processing capability of the processing circuit or the wiring circuit, and the number of wiring circuits The computer program according to claim 11, comprising at least one of a delay time, the number of bits of the circuit resource, and an interval of clocks for switching the operation state.   A computer-readable information storage medium in which the computer program according to any one of claims 11 to 15 is stored. A data processing device according to claim 1 or 4,
Switching control is individually performed for connection relations between a plurality of processing circuits that individually execute data processing corresponding to the first operation command set individually and a plurality of processing circuits corresponding to the first operation command. a matrix circuit in which a plurality of wiring circuits are arranged in a matrix of, a state management unit in response to the second operation command switch sequentially for each operation cycle a context of the matrix circuit, the data by the data processing device A parallel computing device including a storage unit in which the output object code is stored ;
A data processing system comprising: The constraint conditions include the number of the plurality of processing circuits, the number of input ports of the processing circuit, the presence or absence of usable circuit resources of the processing circuit, the processing capability of the processing circuit or the wiring circuit, and the number of wiring circuits The data processing system according to claim 17, comprising at least one of a delay time, the number of bits of the circuit resource, and a clock interval for switching the operation state.

Images