JP4743581B2 - Data processing system and control method thereof - Google Patents

Description

  The present invention relates to a data processing system capable of generating data paths having different configurations by a plurality of processing elements whose functions and connections between them can be changed, and a control method therefor.

Japanese Laid-Open Patent Publication No. 8-101761 discloses a reconfigurable architecture that can be quickly switched between a plurality of configurations. Among them, a core cell including a logic cell having a selectable function therein is disclosed. The configuration is described. The core cell further includes a configuration cache and an instruction cache, and the configuration cache further includes four bit data storage units. One of the four data storage units of the configuration cache is selected according to the data of the instruction cache.
JP-A-8-101761

  For example, various data flows (data) can be dynamically changed by switching the connection of a plurality of processing elements (elements) having an arithmetic function, a comparison function or a logic function of an appropriate scale such as ALU (Arithmetic Logic Unit). Technology to reconstruct the path) has been developed. Elements can be arranged like a matrix or tree and connected by a group of wires that can change the connection between those elements to create a data path, or elements can be connected to connect adjacent elements It is possible to generate a data path.

  In order to diversify the data paths that can be generated, it is desirable that the processing function of each element can be changed, and any appropriate arithmetic operation such as addition, subtraction, multiplication, division, etc., or appropriate logic such as logical sum, logical product, etc. It is desirable that one of operations can be selected. In order to change the data path configuration, it is necessary not only to change the connection of elements but also to change the processing function of each element. There are several ways to provide configuration information to change processing capabilities for an element. One is a method of supplying from the outside via a wiring group connecting elements. Although this method requires less hardware resources, it takes an enormous amount of time to supply configuration information by designating individual elements. Therefore, it is impossible to change the data path dynamically. If each element has functions and resources similar to those of a microcomputer, it may be possible to download all information for executing an application in advance via a data path wiring group. However, it is not flexible to dynamic application changes and cannot handle various exceptions. In order to cope with such a situation, communication wiring between elements is required, and hardware resources increase.

  One of the other methods is to provide a bus for supplying configuration information to each element separately from the wiring group that configures the data path. The system can supply configuration information to individual elements at the appropriate time under the control of a suitable central control system. Therefore, the flexibility to change the application is high, and if the central control system can detect the exception, it can cope with the exception. However, a central control system that manages the functions of tens or hundreds of elements in parallel and independently is required to have the ability to change the functions of all elements in one clock or several clocks. Processing capacity and high processing speed are required. Accordingly, the control system for managing the elements requires a processor that operates at a higher speed than the elements and has a much higher processing capacity, and therefore, space and cost of the LSI are consumed. For this reason, the merit of providing an LSI that generates a reconfigurable data path by a plurality of elements is diminished.

  The function of the element can be switched by a clock signal or a cyclic signal based on the clock signal, and management of the timing for switching the function of the element is not required, so the load on the system that manages the function of the element Will be greatly reduced and management hardware will be simplified. However, the functionality of the element only switches cyclically, so there is little configurable data path flexibility. For example, if two functions are switched based on a clock signal, it is only possible to implement a data path that is twice as many as the same number of elements as compared to hardware that cannot dynamically change the data path. On the other hand, the processing speed generally decreases, and various problems such as an increase in hardware such as FFs occur in order to ensure the synchronous design.

  There is a method in which a configuration memory having an appropriate capacity and a small-scale processor are mounted on each element, and the processor selects an appropriate function from the configuration memory according to a small-scale program and changes the function of the element. This method may be advantageous in that the management of elements can be distributed. However, in addition to a bus for supplying configuration information to the individual elements, a bus for supplying a program for controlling the individual elements to the individual elements is required, and the hardware increases. Furthermore, in order to synchronize the timing of changing functions between elements and to handle exceptions, communication hardware is required.

  A data path having a configuration suitable for executing a desired process by connecting a plurality of elements not only processes one packetized data but also the same process for a large number of data packets. Used to perform Therefore, even if the data path is dynamically changed, there is a certain amount of time after a data path is generated until a different data path is generated using the elements constituting the data path. For this reason, if it is not necessary to supply information for configuring the next data path to the element in units of clocks, hardware for providing configuration information to the element can be simplified. Then, after the configuration information for configuring the next data path is distributed to all the elements and the data packet to be processed in the previous data path has passed, all the elements that require the function change are all instructions that change the function. By broadcasting to this element, it is possible to change the function of the appropriate element and set the next data path at an appropriate timing in units of clocks. In this system, the hardware for managing element functions can be simplified, the flexibility to change the application is high, and the communication function for changing the functions at the same timing between elements is unnecessary. .

  This system can reconfigure the data path in units of clocks, but in order to broadcast a command to change the function for all elements, all the functions that are required to be changed to configure the next data path The process needs to be completed in the element. That is, the data path to be reconfigured needs to be processed for all of its elements. Therefore, focusing on the data path, it is reconfigured with one clock or several clocks. However, when viewed in units of elements, if a data packet is flowing through the data path in the pipeline method, the element on the input side of the data path The function is changed with the delay of the data path latency after the processing of the final data packet is completed, and the data path latency is delayed by the data path latency after the function is changed at the data path output element. Process is started. In a data processing system such as an LSI capable of reconfiguring a data path, it is important to further reduce the processing gap between elements when further increasing the processing speed, and is an object of the present invention.

One embodiment of the present invention includes a plurality of processing elements and a wiring group that connects the plurality of processing elements, and each processing element has at least one of a processing function and a connection between the processing elements by the wiring group. A control method for a data processing system including a control unit for changing an individual function including at least one of connection functions for changing a section, and comprising at least one data path through which a data packet flows in a pipeline manner by a plurality of processing elements Each of the processing elements includes a background operation unit that stores configuration information that defines a future individual function provided by the supply mechanism among configuration information that defines an individual function, and a future individual of the background operation unit. Configuration information that defines the function And a foreground operation unit for storing the configuration information defining the selected current individual function from within the data processing system further configuration information defining the future individual function, each of the individual through the feed mechanism A plurality of processing elements having a processor, a data input buffer , and a processor that can change the configuration information that defines the current individual function based on the configuration information that defines the future individual function . A first data path is configured by connecting at least a part of the data input buffer and at least a part of the wiring group, and the data input buffer is configured to process the first data path in the first data path. supplies data packet including the first data and a token of interest, the token Of the first data, the first shows the unwanted main-processing in the processing element included in the data path, further, the data input buffer, as a token, the final data or final data through the individual processing Including a function of including an end token indicating an opportunity to change the individual function of an element, and the control method includes a step of supplying configuration information defining a future individual function to an individual processing element via a supply mechanism. A data input buffer supplying a data packet including an end token to a processing element constituting the first data path via the first data path; and a processing element constituting the first data path control unit is, the advent of the end token, Masaru And a step of changing the configuration information for determining the current individual function based on the configuration information for determining the individual function .

  In this control method, the second data is supplied in the order of the first data to be processed to the individual processing elements constituting the data path along the data path. The second data is information that is supplied in a state of being distinguished from the first data and indicates an opportunity to change the individual function of each processing element. Therefore, if each processing element is decoded independently of the processing of the first data, and the processing element is provided with configuration information for changing the individual function, the second data is used as the second data. Individual functions of processing elements constituting one data path are changed. If the configuration information given to each processing element is configuration information that changes each processing function, for example, a parameter, without changing the connection, the second data passes through the first data path. Accordingly, the individual functions of the individual processing elements are sequentially changed, and a second data path having a configuration different from that of the first data path is generated. Therefore, even if not all of the second data path is reconfigured, a processing element whose individual function has been changed to configure a data packet to be processed by the second data path to form the second data path. It becomes possible to supply to. That is, the pipeline processing by the second data path can be started without waiting for the second data path to be completely configured. Therefore, in each element, it is possible to minimize the time gap from the end of the processing of the first data path to the start of the processing of the second data path. In a configurable data processing system such as an LSI, the processing speed can be further increased.

  If the reconstructed second data path also requires processing elements other than the processing elements included in the first data path, for example, included in the third data path, they are included in the third data path. If the individual function of the processing element has already been changed, the second data is supplied to the first data path, so that the second data path is reconfigured in the same manner as described above. Can start pipeline processing of data paths. On the other hand, if the individual function of the processing element included in the third data path has not been changed, the processing included in the third data path when the second data is supplied to the third data path. The pipeline processing of the second data path can be started in conjunction with the change of the individual function of the element. Therefore, in any case, the pipeline processing can be started before the reconfiguration of the second data path is completed.

  The second data may be generated by a processing element constituting the first data path and supplied to the first data path together with the first data. For example, when an exception occurs in a processing element in the middle of the first data path, the downstream processing element can be changed to an individual function that performs processing for exception for subsequent data packets. Opportunities to change the configuration of the data path often occur due to the end of a group of data packets to be processed by the data path. Accordingly, the control method of the present invention further includes supplying the packetized first data to the first data path, and supplying the second data to the first data path together with or after the last packet. It is desirable to have

One of the different aspects of the present invention includes a plurality of processing elements and a wiring group connecting the plurality of processing elements, and each processing element has a respective processing function and a connection between the processing elements by the wiring group. A data processing unit comprising a control unit for changing an individual function including at least one of connection functions for changing at least a part of the data processing unit, and constituting at least one data path through which a data packet flows in a pipeline manner by a plurality of processing elements Is a data processing system. Each processing element is configured to store a background operation unit that stores configuration information that defines a future individual function provided by the supply mechanism among configuration information that defines an individual function, and a configuration that defines a future individual function of the background operation unit. And a foreground operating unit that stores configuration information that defines the current individual function selected from the information. In addition, the data processing system supplies configuration information that defines future individual functions to each individual processing element via a supply mechanism, and each processing element provides current information based on configuration information that defines future individual functions . The processor includes a processor that can change configuration information that defines an individual function, and a data input buffer, and at least a part of the plurality of processing elements and the data input buffer are connected by at least a part of the wiring group. The data input buffer supplies the first data to be processed in the first data path and the data packet including the token, and the token is supplied to the first data path of the first data. indicates No Yes-processing in the processing element contained, further, de Data input buffer, as a token, passing the final data or final data includes ability to include end token indicating the opportunity of changing the individual function of each processing element, the control unit of each processing element, the end This is a data processing system that changes the configuration information that defines the current individual function based on the configuration information that defines the future individual function when a token arrives. In this data processing system, the second data is supplied in units of data paths rather than individually supplying signals for instructing individual functions to be changed to individual processing elements. There is no significant complexity. Also, the control unit of each processing element may include a function for decoding the second data and changing the individual function, so that the function as a general-purpose processor that fetches an instruction using the program counter is also included. do not need.

  The control method and data processing system of the present invention do not exclude having a step and means for broadcasting an instruction or signal indicating the timing of reconfiguring the data path to all processing elements. When it is necessary to change the configuration of a wide range of processing elements including processing elements that do not constitute a data path, a method of broadcasting information indicating timing is effective. However, wiring for broadcasting is necessary.

  In the control method of the present invention, the second data also indicates validity / invalidity of the first data, and a processing element that constitutes the first data path processes the first data based on the second data. It is effective to have further. In the data processing system of the present invention, the data supply means adds information indicating validity / invalidity of the first data to the second data, and the control unit of each processing element adds to the second data. Based on this, it is effective to process the first data. In a data processing system that uses a token as data indicating validity / invalidity of the first data, it is possible to add information indicating an opportunity to change the individual function of each processing element to the token.

  As described above, in the data processing system and the control method thereof according to the present invention, the individual processing elements are individually and independently in the order in which the necessity of the current individual functions realized in the individual processing elements is completed. , Autonomously reconfigured. That is, the reconfiguration is the smallest unit that constitutes the data path, that is, the element, the previous process, that is, the data to be processed by the current individual function passes, the process ends, or the data to be processed By disappearing, the current individual function is performed at a timing that is no longer necessary.

  Therefore, according to the control method of the data processing system of the present invention, the process of supplying the configuration information defining the future individual functions to the individual processing elements, and the individual processing elements when the current individual functions are no longer necessary, And a step of switching to an individual function. In addition, the data processing system of the present invention has configuration information supply means for supplying configuration information defining future individual functions to each processing element, and each processing element does not require the current individual function. In other words, a control unit for switching to the next individual function is provided. For this reason, in the data processing system of the present invention, it is not necessary to wait for input of data in the next data path until the entire processing of the data processing system or each data path is completed. Since the processing element after processing is reconfigured for the next processing, data can be continuously input. Therefore, it is possible to increase the processing speed by eliminating the waiting time.

  FIG. 1 shows an example of a data processing system. This data processing system 1 is made into a chip, a reconfigurable data processing area (data processing unit) 19, a function for reconfiguring the reconfigurable data processing area 19, a function for controlling input / output, etc. And a peripheral region 18 for supporting the above.

  The reconfigurable circuit area 19 is divided into a plurality of segments 10 to 15 in FIG. 1, but a plurality of elements are two-dimensionally arranged in an array or matrix as shown in FIG. This is a configuration called a matrix. The matrix 19 of this example includes a plurality of processing elements (PE) 21 arranged two-dimensionally in the vertical and horizontal directions, wirings 22 arranged in a lattice pattern between them, and vertical and horizontal wirings 22 at connection points of the wirings 22. And a switching unit 23 that can freely switch the connection. Each PE 21 has a configuration in which the processing function and the connection with the wiring group 22 can be changed. To change the processing function, it is possible to adopt a configuration in which the function can be freely set by a lookup table or the like. In this example, there are some functional groups, such as elements for arithmetic and logic operations, elements for delay, elements for memory, elements that generate addresses for inputting or outputting data, and elements for inputting or outputting data. The space efficiency of the matrix 19 is improved by arranging elements having internal configurations suitable for the respective processes. In addition, since the redundancy is reduced by arranging the elements divided into a certain number of function groups, there is an advantage that AC characteristics and processing speed can be improved.

  3 and 4 are examples of PE21. The PE 21 includes an internal data path area 29 whose function can be changed, and a control unit 60 that sets the function of the internal data path area 29. The internal data path area 29a of the PE 21a shown in FIG. 3 includes an address generation circuit 28 composed of a counter and the like, and a selector SEL, and an address generated under the conditions set by the control unit 60 is used as the output signal do. Output to the wiring 22. The output signal do is fed back to the PE 21a via the row wiring and the column wiring as it is or after being processed by another PE 21 as the input signal dix or diy in the data packet 90. The address selected by the selector SEL under the conditions set by the control unit 60 is output from the matrix 19 as an address for data input or output. The PE 21 a also includes a selector (not shown) for selecting input data from any of the wirings 22 and outputting output data, and the setting is also performed by the control unit 60.

  The PE 21b shown in FIG. 4 has a configuration suitable for arithmetic operations and logical operations. The internal data path unit 29b includes a shift circuit SHIFT, a mask circuit MASK, and a logical operation unit ALU. Similarly to the PE 21a, the setting unit 60 sets the states of the shift circuit SHIFT, the mask circuit MASK, and the logical operation unit ALU. Therefore, the input data dix and diy supplied in the data packet 90 can be added or subtracted, compared, logical sum or logical product can be calculated, and the result is output as the output signal do in the form of the data packet 90. It is output to the wiring (bus) 22.

  As shown in FIG. 2, the matrix 19 of PU1 includes 368 PEs 21. In order to separately form transfer paths for supplying configuration information (configuration data) to them, the matrix 19 of PU1 is shown in FIG. Thus, it is divided into six segments 10 to 15 in the form. However, the wiring group 22 for connecting a plurality of PEs 21 is transparently arranged with respect to these segments 10 to 15, and a plurality of PEs 21 for the segments 10 to 15 are connected to form various data paths. Can do. Therefore, the switching unit (SU) 23 accepts configuration information as one type of processing element and changes the connection between PEs. Therefore, similarly to PE 21, a control unit 60 that sets a function of switching the connection of the wirings of the wiring group 22 is provided.

  Data is input / output to / from a data path constituted by a plurality of PEs 21 of the matrix 19 using the input buffer 33 and the output buffer 34 in the peripheral area 18. The input buffer 33 includes four input elements LDB, and the configuration and control of the buffer 33 can be set by configuration information. Similarly, the output buffer 34 includes four output elements STB, and the configuration and control can be set by configuration information. In addition, a direct input 31 and a direct output 32 are prepared, and the processing of an application in which circuit elements are insufficient in one chip (PU) 1 can be handled by connecting a plurality of chips 1. Yes.

  In the data processing system 1 of this example, at least one of the processing function and the connection function for changing at least part of the connection between the PEs 21 and SU 23 and the input / output buffers 33 and 34 of the matrix 19 is provided. Configuration information for changing the individual functions to be included is supplied by a shift register type supply mechanism 50. The main supply source of the configuration information is a processor 35 such as a RISC arranged in the peripheral area 18, and the configuration of the matrix 19 can be controlled by software executed by the processor 35. It is also possible to supply configuration information from another PU 1 or the like. The configuration information supply mechanism 50 includes a transfer control unit (TCU) 59, and transfer paths 51 to 58 that extend around each of the segments 10 to 15, the input buffer 33, and the output buffer 34. The TCU 59 is connected to a bus switching unit (bus interface, BSU) 36, and the processor 35 supplies configuration information to the TCU 59 via the BSU 36.

  FIG. 5 shows a more detailed configuration of the configuration information supply mechanism 50. The supply mechanism 50 includes transfer paths 51 to 56 that connect the PEs 21 included in the segments 10 to 15. The transfer paths 51 to 56 connect transfer registers (FF) of the control unit 60 of each PE 21 in series. For example, in the transfer path 51, 1-word data transmitted to the transfer FF of one PE 21 is transmitted to the transfer FF of the next PE 21 with a delay of one clock or one cycle.

  The TCU 59 includes a data area management unit 91 and a delivery unit 92 that collects and distributes the configuration information set by the data area management unit 91 to the transfer paths 51 to 58. The data area management unit 91 includes a bus control unit 95 that manages data exchange with the BSU 36, a transmission unit 93, and a reception unit 94. All the transfer paths 51 to 58 are closed, and the data transmitted from the TCU 59 is returned to the TCU 59, and the configuration information is obtained by the desired PE 21 in the middle of the transfer paths 51 to 58, and is necessary. For example, the configuration information is returned from the PE 21 to the TCU 59.

  FIG. 6 shows the configuration of the control unit 60 of the PE 21 connected to the transfer path 51. The control unit 60 includes a transfer FF 61 that constitutes the transfer path 51, a decoder 62 that decodes the data 75 stored in the transfer FF 61, and a background operation unit 63 that operates in the background to store the data 75. And a foreground operation unit 64 in which configuration information for setting the internal data path area 29 is stored. The background operation unit 63 is a mechanism capable of storing future configuration information. The background memory 65 of the three banks and the data 75 stored in the transfer FF 61 are directly connected to each bank of the background memory 65. A selector 66 that distributes the lines to be output and a selector 67 that can select a bank of the memory 65 and output data are provided. Therefore, configuration information for determining at least three future individual functions can be provided to each PE 21 at an appropriate timing by the transfer path 51. Since the configuration information operating in the foreground can be used as future configuration information, even in a memory of about 3 banks, various data paths can be dynamically reconfigured by switching individual functions. .

  Therefore, the configuration information supplied as data 75 to each PE 21 by the supply mechanism 50 is information for setting individual functions of each PE 21, such as connection of the internal data path area 29, functions or configurations of individual logic gates, and the like. Information for determining the connection between the internal data path area 29 and the wiring group 22, initial conditions or initial conditions when processing is started in the internal data path area 29, or processing as a data path is started Includes data. Furthermore, it is possible to include information for selecting configuration information to be loaded next, and it is also possible to include information related to control of the control unit 60. The configuration information supplied to the switching unit 23 mainly includes information related to the connection of the wiring group 22.

  The foreground operation unit 64 is a mechanism for storing configuration information for realizing the current individual function. The foreground operation unit 64 stores the setting data (configuration information) supplied to the internal data path region 29 to store the internal data path region 29. A foreground memory 68 that maintains the current configuration, and a selector 69 that selects and supplies data from the background memory 65 or the transfer FF 61 to the foreground memory 68. Selectors 67 and 69 that select configuration information to be loaded into the foreground memory 68 are controlled by the decoder 62. It is also possible to control from the data set in the transfer FF 61 and the configuration information set in the foreground memory 68 by switching the second selectors 72 and 70.

  Further, the control unit 60 can select the data 75 to be supplied to the transfer FF 61 of the downstream PE 21 from any of the data of the FF 61 of its own PE 21, the data of the bank of the background memory 65, and the data of the foreground memory 68. An output selector 71 is provided. The decoder 62 analyzes the data 75 transferred to its own transfer FF 61 to switch the output selector 71 and select the data 75 transferred to the transfer FF 61 of the downstream PE 21.

  The control unit 60 further includes a clock signal control unit 77 that can operate and stop the internal data path area 29 by turning on and off the clock signal 79 supplied to the internal data path area 29. The clock signal control unit 77 receives input data (first data) dix and diy to be processed in the internal data path area 29 and tokens (second data) dtx and dty supplied from the wiring group 22. ing. In this example, input data dix and token dtx constitute one data packet 90 and are supplied to each PE 21 via a data path constituted by a plurality of PEs 21 and a wiring group 22 including a switching unit 23. . The same applies to input data diy and token dty. The tokens dtx and dty indicate validity / invalidity of the respective input data dix and diy. In the clock signal control unit 77, if the tokens dtx and dty are valid, the clock signal 79 is supplied to the internal data path area 29, and the internal data set so as to perform a desired processing function according to the configuration information of the foreground memory 68. The input data dix and diy are processed by the path area 29, and the output data do is output from the internal data path area 29. Therefore, the internal data path area 29 does not operate when valid input data is not supplied, so that power consumption can be saved in units of PE 21.

  The internal data path area 29 not only outputs the output data do as the processing result of the input data dix and diy, which is the processing target data, but also includes a token to indicating validity / invalidity of the output data do, The packet 90 is output. When the state of data processing in the internal data path area 29 does not require the processing of the PE 21 having the structure constituting the data path to be interrupted, the data packet 90 includes a token to indicating the validity of the data. When it is preferable to interrupt the processing of the subsequent PE 21, it is possible to generate a token to indicating invalidity of the data and include it in the data packet 90, and communicate with other PEs 21 via the data path by the token. be able to.

  The tokens dtx and dty are also input to the decoder 62. The tokens dtx and dty include not only the validity / invalidity of input data but also information (end token) indicating an opportunity for changing the individual function of the PE 21. Therefore, when one of the tokens dtx and dty requests to change the function of the internal data area 29, the decoder 62 at least one of the selectors 66, 67, and 69 of the background operation unit 63 and the foreground operation unit 64 To select one of the configuration information stored in the background memory 65 or the transfer FF 61 and set it in the foreground memory 68. Thereby, the individual function of the internal data area 29 is changed, and the connection with the wiring group 22 is also changed as necessary. The end token is included in the data packet 90 by being reflected in the token to of the output data do via the internal data path area 29 as in the case of data validity / invalidity. Therefore, the end token is inherited from the input data packet to the output data packet, and is supplied to the subsequent PE 21 via the data path constituted by the plurality of PEs 21. As will be described later, the end token is normally generated by the data input buffer 33 and incorporated in the data packet 90. Further, by changing the configuration of the internal data path area 29, an end token can be generated or extinguished in the PE 21, and the change of the individual function of the subsequent PE 21 can be controlled by the PE 21. is there.

  The configuration of the internal data path unit 29 is switched after the maximum data stay time has elapsed. When the clock signal control unit 77 detects the end token, the clock signal control unit 77 stops data input to the internal data path unit 29, and the decoder 62 switches the configuration by switching the data in the foreground memory 68. Then, after a certain clock wait, the data input is permitted and when the tokens dtx and dty indicating that the data is valid are obtained, the clock signal control unit 77 activates the internal data path area 29. As a result, each PE 21 switches to the next individual function constituting the next data path and starts processing.

  FIG. 7 shows control related to configuration information in the control unit 60 of each PE 21. In step 81, if the data 75 supplied to the transfer FF 61 by the configuration information supply mechanism 50 is the configuration information of its own PE 21, it is stored in the background memory 65 in step 82. In step 83, if either of the tokens dtx or dty supplied from the wiring group 22 constituting the data path is an end token, one of the configuration information of the background memory 65 is stored in the foreground memory 68 in step 84. Thus, the configuration of the internal data path unit 29 is switched. Next, if it is determined in step 85 that the data supplied to the PE 21 by the tokens dtx and dty supplied from the wiring group 22 constituting the data path is valid, the internal data path unit 29 is set in step 86. Processing is performed and data is sent to the PE 21 next to the data path connected by the wiring group 22.

  FIG. 8A shows an outline of the data packet 90 supplied to the internal data path unit 29 of each PE 21 by the wiring group 22. The data packet 90 includes a 2-bit token area 92 and a 32-bit data area 91. FIG. 8B shows the value of the token area 92 and its meaning. When the value of the token area 92 is 2 or 3, it is an end token, and if there is the next configuration information in the background memory 65, it is moved to the foreground memory 68 and the function of the PE 21 is changed. If there is no next configuration information in the background memory 65, the configuration of the internal data path area 29 is not changed and the same function is maintained. Alternatively, when the next configuration information stored in the background memory 65 is the same as that in the foreground memory 68, the configuration of the internal data path area 29 is not changed and the same function is maintained. The end token can be supplied independently of the validity / invalidity of the data, and the end token can be supplied together with the final data processed by the individual function or after the final data.

  The token (second data) 92 is included in the data packet 90 together with the data to be processed (first data) 91, but is stored in different predetermined bit positions. Are distinct. Therefore, the control unit 60 of each PE 21 can decode the token 92 independently of the first data 91 or by ignoring the first data 91. For this reason, as described above, each PE 21 can understand the end token regardless of the validity / invalidity of the first data 91, and the process of changing the individual function can proceed. Further, regarding the change of the individual function, the control unit 60 can perform processing in parallel independently of the internal data path area 29.

  Instead of providing an end token, the internal data path area 29 may interpret the data to be processed, determine the end of the data to be processed, and determine that the current individual function is no longer necessary. In this case, the control unit 60 waits for the determination of the internal data path area 29 and performs the process of moving the next configuration information to the foreground memory. Even in this control method, each PE 21 can autonomously determine that its current individual function is no longer necessary, and can switch the configuration information so as to shift to the next individual function. . However, since it is necessary to add a function for determining that the data is the final data to the internal data path area 29, a series of processes for changing the configuration after the determination in the internal data path area 29 is started. In this change, a delay of a cycle in which data resides in the PE 21 occurs. Furthermore, since the end token is not different from the data to be processed, there is no choice but to change the configuration of the PE 21 in synchronization with the data to be processed. -This means loss, and the flexibility of communication between PEs 21 related to the change of configuration information is limited compared to the case using end tokens.

  FIG. 9 shows a configuration of a part that generates a data packet including an end token 92 in the input buffer 33 of the data processing system 1 of the present example. The end token generation unit 33t sets the number of data to be read via the setting data supply mechanism 57, reads desired data from the external memory accessed via the BSU 36, packetizes it into 32 bits, and forms a matrix. Input to 19 data paths. Then, an end token 92 is attached to the final data packet 90.

  FIG. 10A shows a matrix 19 having a very simple configuration including 16 PEs 21 as a sample. As shown in this figure, the maximum data stay time (processing time) in one PE 21 in the data path 100 that outputs OUT-1 and OUT-2 with respect to the inputs IN-1, IN-2, and IN-3. ) Is 20 clocks, and the time for switching the configuration information of each PE 21 to change the individual function is 1 clock. After the processing of the data path 100 is completed, when reconfiguring the data path 100 by changing the individual functions of the PE 21 at once by the broadcast method, the longest data path 100 passes through the four PEs 21, so the data path The configuration change requires a configuration switching time of 81 clocks + α, and data input is interrupted during this time. That is, a blank time of 81 clocks + α occurs between the last data of the configuration before switching and the head data after switching.

  On the other hand, when the configuration is switched using the end token, when each PE 21 detects the end token, the configuration information of each PE 21 is switched to change the individual function. For this reason, the time required for switching the configuration is 21 clocks + α. Ideally, the data to be processed by the next data path can be input after a blank time of 1 clock. Therefore, the processing delay due to the switching of the configuration can be reduced to several tenths.

  Actually, since the processing times of the PEs 21 having different functions are different, as shown in FIG. 10B, the data D1 and D2 are sequentially input at the time T0, and the PE 21 indicated by “A” at the time T1. At time T2, data is processed at PE21 indicated by "B", data is output from OUT-1, and data D2 is output from OUT-1 at time T3. The data D1 is output from the PE-2 of OUT-2 indicated by “,” and the data D2 is output from OUT-2 at time T4. If the functions of the PEs 21 of the matrix 19 are to be changed all at once by the broadcast method, the functions are changed at time T5 after time T4 to reconfigure the data path.

  In the case of the end token method, the change of the individual functions of the PE 21 indicated by “A” and the upstream PE 21 at the time T2 is permitted, and the PE 21 indicated by “B” and the upstream PE 21 at the time T3 are permitted. The individual functions are allowed to be changed, and at time T4, the individual functions of all the PEs 21 are allowed to be changed. Thus, in the data processing system 1 of the present invention, it is not necessary to wait for the next data input until the entire processing of the data processing system or each data path 100 is completed. Since the PE 21 that has completed the processing of the current individual function is reconfigured to the individual function for the next processing, data can be continuously input. Therefore, it is possible to increase the processing speed by eliminating the waiting time.

  In the data processing system 1 of the present example, the configuration information is provided to the individual PEs 21 by the supply mechanism 50 of the shift register method. Configuration information may be provided by a method. In the data processing system 1 of this example, the wiring resources 22 and 23 are prepared separately from the individual PEs 21 so as to configure the data path. However, the PE 21 is connected to the adjacent PE as the wiring resource. The data path may be configured by changing the above. Furthermore, the present invention can also be applied to a reconfigurable data processing system in which all PEs 21 have a common configuration.

  In the present invention, in a data processing system having means for providing configuration information to individual PEs and means for connecting individual PEs via a network to form a data path, the configuration is made via the data path. The information that permits the change is provided, and the effect of reducing hardware resources and waiting time related to data path reconfiguration is obtained. In particular, the control method of the present invention is suitable for the operation of changing the individual function of the PE constituting the existing data path and reconfiguring the data path of a new function, and a new function is provided by different parameters. The data path to be provided can be reconfigured with one clock so that data can be processed continuously with little latency.

  The means for broadcasting permission of reconfiguration to all PEs constituting the data processing system by wire or wireless is suitable for reconfiguring data paths collectively including PEs not included in the data paths. . Therefore, it is effective to provide a broadcasting means for the present invention. In both the broadcast method and the method using the end token of the present invention, in order to give different settings to individual devices, that is, PEs, it is not necessary to access each device individually, Redundant hardware for that purpose is unnecessary, and time spent for that purpose is also unnecessary.

  The data processing system and the control method thereof according to the present invention can be provided as a system LSI or ASIC capable of executing various data processing. The data processing system of the present invention is not limited to an electronic circuit, and can be applied to an optical circuit or a data processing apparatus and system based on an optoelectronic circuit, and further to a distributed processing system connected to a network. It is. Since the data processing system of the present invention can execute data processing at high speed with reconfigurable hardware, it is suitable for data processing devices that require high-speed and real-time properties such as network processing and image processing. It is.

It is a figure which shows the outline | summary of a data processing system. It is a figure which shows the outline | summary of a matrix. It is a figure which shows an example of a processing element (PE). It is a figure which shows the other example of PE. It is a figure which shows the state which connected PE by the transfer path. It is a figure which shows the structure of the control unit of PE. It is a flowchart which shows control of the control unit of PE. FIG. 8A shows the structure of the data packet, and FIG. 8B shows the token value and description. It is a figure which shows the structure of an input buffer. FIG. 10A is a diagram for explaining the present invention with a simple matrix, and FIG. 10B is a time chart.

Explanation of symbols

1 data processing system, 10-15 segment 19 matrix, 21 processing element (PE)
29 Internal data path area 50 Data transmission mechanism 60 Control unit

Claims (4)

A plurality of processing elements and a wiring group connecting the plurality of processing elements, each processing element changing a processing function and at least a part of a connection between the processing elements by the wiring group A control method for a data processing system comprising a control unit for changing an individual function including at least one of functions, wherein the plurality of processing elements constitute at least one data path through which a data packet flows in a pipeline manner ,
The individual processing element includes a background operation unit that stores configuration information that defines a future individual function provided by a supply mechanism among configuration information that defines the individual function, and the future individual of the background operation unit. A foreground actuating unit that stores configuration information that defines the current individual function selected from the configuration information that defines the function,
Said data processing system further configuration information defining the future of the individual function is supplied to each of the individual processing elements via the feed mechanism, the individual processing elements defining the future of the individual functional configuration A processor capable of changing the configuration information defining the current individual function based on the information, and a data input buffer,
A first data path is configured by connecting at least a part of the plurality of processing elements and the data input buffer by at least a part of the wiring group,
The data input buffer, the first data path, and supplies the data packet including the first data and a token to be processed of said first data path, the token of the first data, The processing element included in the first data path indicates the necessity / unnecessity of processing. Further, the data input buffer passes the final data or the final data as the token , and the individual processing elements Including the ability to include end tokens that indicate opportunities for individual function changes,
The control method is
The processor supplying configuration information defining the future individual functions to the individual processing elements via the supply mechanism;
The data input buffer supplying the data packet including the end token to the processing elements constituting the first data path via the first data path;
The control unit of the processing element constituting the first data path changing the configuration information defining the current individual function based on the configuration information defining the future individual function when the end token arrives; Control method having. The data processing system according to claim 1, further comprising a broadcasting unit that instructs all of the plurality of processing elements to change the configuration information that defines the current individual function by wire or wireless.
The control method is
A control method comprising the step of instructing the plurality of processing elements to change the individual function via the means for broadcasting by the processor. A plurality of processing elements and a wiring group connecting the plurality of processing elements are provided, and each processing element is a connection that changes at least a part of the processing function and the connection between the processing elements by the wiring group. A data processing system including a control unit that changes an individual function including at least one of functions, and having a data processing unit that configures at least one data path through which a data packet flows in a pipeline manner by the plurality of processing elements. And
The individual processing element includes a background operation unit that stores configuration information that defines a future individual function provided by a supply mechanism among configuration information that defines the individual function, and the future individual of the background operation unit. A foreground actuating unit that stores configuration information that defines the current individual function selected from the configuration information that defines the function,
Furthermore, the data processing system
The configuration information defining the future of the individual function supplied through the feed mechanism for each of the individual processing elements, the current individual function based on the configuration information the individual processing elements defining the future of the individual function a processor that allows changing the configuration information defining a first by having a data input buffer, and at least a portion of said plurality of processing elements and the data input buffer is connected by at least a portion of the wiring group 1 data path is configured, and the data input buffer supplies the data packet including the first data and the token to be processed in the first data path, and the token is the first data, Processing in the processing element included in the first data path; - shows the unnecessary,
The data input buffer, as the token passes through the last data or final data includes ability to include end token indicating the opportunity of changing the individual function of the individual processing elements,
The data processing system, wherein the control unit of the individual processing element changes the configuration information defining the current individual function based on the configuration information defining the future individual function when the end token arrives. 4. The data processing system according to claim 3, further comprising a broadcasting means for instructing all of the plurality of processing elements to change the configuration information defining the current individual function by wired or wireless.

Images