JP5697102B2 - Configuration apparatus and configuration method of image processing system - Google Patents

Description

  The present invention relates to a technique for reconstructing an image processing system by disassembling an existing image processing system into a plurality of processing components (IP) and changing connections between the processing components.

  When there is an image processing system designed to perform specific image processing, if you want to increase flexibility, such as changing the internal processing order, disassemble the image processing system into multiple processing components. The image processing system is reconfigured by changing the connection between the processing components.

  Here, as a general method for changing the connection between the processing components constituting the image processing system, there is a method in which the image processing system has a NoC (Network-on-Chip) configuration.

  NoC is a technique for flexibly changing the connection between processing components when a plurality of processing components are mounted on an integrated circuit. This technology is characterized in that a “router” is added to each of a plurality of processing components. The routers are interconnected with neighboring routers by wiring for transmitting and receiving data. That is, the router has a plurality of transmission ports. The data output from the processing component is accompanied by a destination, and the router transmits the data from an appropriate transmission port according to the destination. The router can use the received data as an input to the processing component, transmit the output of the processing component, and transmit (transfer) the received data as it is. In this specification, a system formed by wiring between routers and a router is expressed as NoC in a narrow sense.

  Although the desired flexibility can be obtained by the function of the router mounted on the NoC, the NoC has a problem that the area required for adding the router increases.

  Regarding this problem, Non-Patent Document 1 describes that if the image processing system has a BMNoC (Bus-Mesh-NoC) configuration, the required area can be made smaller than that of the NoC configuration.

  BMNoC is one of the improved technologies of NoC and is characterized in that each processing component is mounted on a bus. There is a router on the bus, and the portion formed by the router and the wiring between the routers corresponds to a conventional NoC. There may be a configuration in which switches are placed on the bus and a plurality of switches are accommodated in a router.

  Here, NoC and BMNoC are compared.

  In NoC, in order to realize sufficient transfer capability between processing components, it is necessary to configure the wiring in a mesh or a complete mesh. However, this configuration increases the wiring between routers. In NoC, an area for performing processing associated with a data destination is required, and this area increases in proportion to the number of routers. On the integrated circuit, these areas are subtracted from the area for realizing the processing component, and the area for realizing the processing component is lost.

  On the other hand, in BMNoC, the latter (area of the router) can be saved mainly by using a bus for data transfer between close processing components. In addition, data transfer between processing components mounted on different buses goes through a switch and a router, but transfer with a small amount of data can be assigned to such a route. This is realized by the BMNoC configuration algorithm, and this allows the performance to be maintained even if the number of wiring between routers is small, so that the former (wiring between routers) can be mainly saved.

SeungJu Lee, Masao Yanagisawa, Tatsuo Ohtsuki and Nozomu Togawa, “BusMesh NoC: A Novel NoC Architecture Comprised of Bus-based Connection and Global Mesh Routers”, Circuits and Systems (APCCAS), 2010 IEEE Asia Pacific Conference on, December 2010 , p. 712-715

  As described above, when the image processing system has the BMNoC configuration, there is an advantage that the required area is smaller than that of the NoC configuration.

  However, in the BMNoC configuration algorithm, when a certain processing component is mounted on which bus, as an influence, the data transfer amount of the transfer path on the network connecting the buses between the buses is calculated. The arrangement of processing components is optimized with the data transfer capability inherent to each of the bus and the transfer path as the upper limit. Therefore, it is necessary to obtain the output data amount of each task executed by the processing component in advance while giving the specific data transfer capability of each bus and transfer path separately.

  However, since the amount of output data of a task is generally measured by a real processor or a dedicated simulator for simulating the processor, there is a problem that labor and time are consumed.

  An object of the present invention is to provide a technique capable of calculating the amount of output data of a task without using a real processor or a dedicated simulator.

The apparatus of the image processing system of the present invention is
A configuration device for configuring an image processing system that executes each of a plurality of tasks by a processing component,
A first storage unit that stores a task input data amount table describing a list of input data amounts of each input of the task for each task;
A second storage unit that stores, for each processing component, a component-specific output data amount table that describes a list of output data amounts of each output of the processing component;
A third storage unit for storing a task graph connection table describing a processing component for executing the task and a list of inputs of the next task to which the output data of the task is input for each task;
A first process for extracting a task whose input data amount is all determined from the task input data amount table as a target task, and a processing component for executing the target task and the next to the target task from the task graph connection table. A second process for extracting a list of task inputs, a third process for extracting a list of output data amounts of each output of the processing component executing the target task from the component-specific output data amount table, and the task A fourth process for extracting a list of input data amounts for each input of the target task from an input data amount table; and an output of each input of the target task and each output of a processing component that executes the target task based on the data amount, describe and calculate the output data of the output of the target task to calculate the list of output data of the output in the task output data amount table And a sixth process that describes output data of the output corresponding to the target task as the input data amount of the input of the next task of the target task in the task input data amount table. For all of the tasks, a data transfer amount calculation unit that executes sequentially from the task in which the amount of input data is all determined ,
A fourth storage for storing a task output data amount table describing a list of output data amounts of each output calculated by the data transfer amount calculation unit for each task, which is an output data amount of each output of the task Part.

The configuration method of the image processing system of the present invention includes:
A configuration method by a configuration device for configuring an image processing system that executes each of a plurality of tasks by a processing component,
For each task, a first step of describing a list of input data amounts of each input of the task in a task input data amount table;
A second step of describing, for each processing component, a list of output data amounts for each output of the processing component in a component-specific output data amount table;
A third step for describing, in the task graph connection table, a processing component for executing the task and a list of inputs of the next task to which output data of the task is input for each task;
Based on the task input data amount table, the component-specific output data amount table, and the task graph connection table, for each task, an output data amount of each output of the task is calculated, and the task input data amount table From the task graph connection table, a processing component for executing the target task and a list of inputs of the next task after the target task, From the second process for retrieving the output data, the third process for retrieving the output data amount list of each output of the processing component that executes the target task from the component-specific output data amount table, and the task input data amount table, A fourth process for retrieving a list of input data amounts of each input of the target task, and a process of executing the input data amount of each input of the target task and the target task The output data amount of each output of the target task is calculated based on the output data amount of each output of the product, and a list of calculated output data amounts of each output is described in the task output data amount table. And a sixth process for describing output data of output corresponding to the target task as an input data amount of an input of the next task of the target task in the task input data amount table. For all the tasks, the input data amount is sequentially executed from the task in which all the input data amounts are determined, and for each task, the output data amount of each output possessed by the task and the calculated output data amount list of each output are displayed as task output data. And a fourth step described in the quantity table.

  According to the present invention, an output data amount specific to a processing component is prepared in advance as a component-specific output data table, and the task output data amount is calculated using the component-specific output data table.

  As a result, there is an effect that a real processor and a dedicated simulator for simulating the processor are not required.

It is a block diagram of the component apparatus of one Embodiment of this invention. It is a figure which shows the example of the existing image processing system. It is a figure which shows the example which made NoC structure the image processing system. It is a figure which shows the example of the task graph required when making an image processing system BMNoC structure. It is a figure which shows the example which made the image processing system the BMNoC structure. It is a figure which shows the example of a task input data amount table | surface. It is a figure which shows the example of the output data amount table according to components. It is a figure which shows the example of a task graph connection table. It is a figure which shows the example of a task output data amount table | surface. It is a flowchart explaining the algorithm for producing a task output data amount table | surface. It is a flowchart explaining a BMNoC structure algorithm.

  EMBODIMENT OF THE INVENTION Below, the form for implementing this invention is demonstrated with reference to drawings.

  FIG. 1 shows a block diagram of a component device of the present embodiment.

  As shown in FIG. 1, the configuration apparatus according to the present embodiment includes a task input data amount table storage unit 11, an output data amount table storage unit 12 for each component, a task graph connection table storage unit 13, and a data transfer amount calculation unit. 14, a task output data amount table storage unit 15, a BMNoC configuration determination unit 16, and a BMNoC configuration data storage unit 17.

  The configuration apparatus according to the present embodiment is an apparatus for configuring an image processing system that executes each of a plurality of tasks by using processing components, and setting the image processing system to a BMNoC configuration.

  When the image processing system is configured to have a BMNoC configuration, a task graph that shows a configuration diagram representing task connection relationships and an output data amount of each output of each task is required.

  In the present embodiment, the task graph is divided into a task graph connection table representing the connection relation between tasks and a task output data amount table representing the output data amount of each output of each task, and the latter task output data amount A data transfer amount calculation unit 14 is provided as means for calculating the output data amount of a task necessary for creating a table.

  Here, each component of the component apparatus of this embodiment is demonstrated in detail.

  The task input data amount table storage unit 11 is a first storage unit that stores, for each task, a task input data amount table that describes a list that lists the input data amount of each input of the task. An example of the task input data amount table is shown later in FIG. 6A.

  The component-specific output data amount table storage unit 12 is a second storage unit that stores, for each processing component, a component-specific output data amount table that describes a list that lists the output data amount of each output of the processing component. . An example of the output data amount table for each part is shown later in FIG. 6B.

  The task graph connection table storage unit 13 stores, for each task, a task graph connection table that describes a processing component that executes the task and a list that enumerates the input of the next task to which the output data of the task is input. It is the 3rd storage part. An example of the task graph connection table is shown later in FIG. 6C.

  Based on the task input data amount table, the component-specific output data amount table, and the task graph connection table, the data transfer amount calculation unit 14 calculates the output data amount of each output of the task for each task. An example of the algorithm of the data transfer amount calculation unit 14 is shown later in FIG. 6E.

  The task output data amount table storage unit 15 includes, for each task, a task output data amount describing a list that lists the output data amount of each output of the task and the output data amount calculated by the data transfer amount calculation unit 14 It is the 4th storage part which stores a table. An example of the task output data amount table is shown later in FIG. 6D.

  Based on the task graph connection table and the task output data amount table, the BMNoC configuration determination unit 16 determines the arrangement of processing components when the image processing system is configured to have a BMNoC configuration using the BMNoC algorithm. An example of the BMNoC algorithm is shown later in FIG.

  The BMNoC configuration data storage unit 17 stores data representing the BMNoC configuration of the image processing system for which the BMNoC configuration determination unit 16 has determined the arrangement of processing components.

  FIG. 2 shows an example of an existing image processing system to be configured by the configuration apparatus of the present embodiment.

  As shown in FIG. 2, the image processing system includes an input IF (input interface) 21, a processing component 22 that executes a plurality of tasks, and an output IF (output interface) 23.

  In FIG. 2, characters in parentheses () in the rounded rectangle representing the processing component 22 indicate the component type, and the character above it indicates the processing content. However, the processing component 22 shown in FIG. 2 is an example, and what processing component 22 constitutes the image processing system is not an essential part of the present invention. Therefore, the detailed description about the processing content of the processing component 22 is abbreviate | omitted.

  The image signal input from the input IF 21 is processed by each of the processing components 22. The output of the processing component 22 is passed to the next processing component 22 connected to the processing component 22. The image signal processed by each of the processing components 22 is output from the output IF 23.

  In order to reconfigure the image processing system thus configured, the processing component 22 constituting the image processing system is disconnected and the processing component 22 is reconnected.

  Therefore, in the present embodiment, the BMNoC configuration, which is one of the improved NoC technologies, is adopted, and the image processing system is configured as the BMNoC configuration. The reason will be described below.

  FIG. 3 shows an example in which the image processing system has a NoC configuration.

  As shown in FIG. 3, when the image processing system has a NoC configuration, a router 24 is added to each processing component 22.

  The image signal input from the input IF 21 first enters the router 24. The router 24 has several outputs, and it is possible to set in advance how the given input is identified and from which output is sent. For example, when an image signal is given as an input, it is sent to the processing component 22 for processing, or is it passed to the next router 24 or the like.

  It is also possible to give a predetermined identifier to the image signal in advance and change the setting of the router 24 so that the image signal is sent to either one according to the value of the identifier.

  When the image processing system is configured as NoC, it is not necessary to disconnect the connection when changing the combination and order of the processing components 22 when reconfiguring the image processing system, and only the setting given to the router 24 needs to be changed. .

  However, when the image processing system has a NoC configuration, there is a problem that a required area increases to add the router 24.

  For the above reason, in this embodiment, as described in the background art, the BMNoC configuration that can reduce the required area as compared with the NoC configuration is adopted.

  In order to adopt the BMNoC configuration for the image processing system, a task graph that shows the configuration diagram representing the connection relation between tasks and the output data amount of each output of each task is required.

  FIG. 4 shows an example of a task graph required when the image processing system has a BMNoC configuration.

  As shown in FIG. 4, in the task graph, the task 30 is represented by a circle. A unique number is assigned to the task 30 on the task graph. This is a task ID (Identification). The upper character in the circle representing the task 30 represents the component type of the processing component that executes the task 30, and the lower character represents the task ID of the task 30. The task 30 on the task graph corresponds to a processing component in the configuration diagram of the image processing system. When the output of task A becomes the input of task B, they are connected by an arrow line from task A to task B. This line corresponds to the connection between the processing components in the configuration diagram of the image processing system. This is done for all tasks 30. The numerical value shown near the line between the tasks 30 represents the amount of data transferred between the tasks 30. In the present embodiment, the unit of data amount is arbitrary. Therefore, the data amount does not need to be a real number and is a non-negative integer.

  In the BMNoC configuration algorithm, the image processing system can be configured as a BMNoC configuration based on this task graph.

  FIG. 5 shows an example in which the image processing system has a BMNoC configuration.

  In FIG. 5, MR represents a router, ES represents a switch, NI represents a network interface, IP represents a processing component, Bus represents a bus, and CN represents a node.

  FIG. 5 shows an example in which the image processing system shown in FIG. 4 is arranged in a BMNoC configuration under the condition of five processing components on a bus, three buses under a switch, and two switches under a router.

  However, in order to make the image processing system BMNoC configuration as shown in FIG. 5, a task graph as shown in FIG. 4 is necessary, and in order to obtain this task graph, it is necessary to determine the output data amount of each task. There is.

  However, since the amount of output data of a task is generally measured by a real processor or a dedicated simulator that simulates the processor, it takes time and effort.

  On the other hand, in this embodiment, the data transfer amount calculation unit 14 calculates the output data amount of the task using the output data amount specific to the processing component that executes the task.

  Here, the calculation method of the output data amount of a task is demonstrated in detail.

  The data transfer amount calculation unit 14 calculates the output data amount of each output of the task for each task based on the task input data amount table, the output data amount table for each part, and the task graph connection table. Write a list that lists the output data amount in the task output data amount table.

  Therefore, in the initial state before calculating the output data amount of the task, it is necessary to describe in advance in each column of the task input data amount table, the component-specific output data amount table, and the task graph connection table. This description may be performed by the data transfer amount calculation unit 14 or may be performed by other means.

  However, for the output data amount table for each part and the task graph connection table, it is necessary to complete all the numerical values, but for the task input data amount table, if the numerical value is not fixed, it will be treated as a special value. Write nil and leave it unfinished.

  FIG. 6A shows an example of a task input data amount table.

  As shown in FIG. 6A, the task input data amount table is a table describing a list of input data amounts of each input of the task with respect to the task ID of the task.

  Since each task has a unique number of inputs, the inputs are arranged in advance in the order of the linear list, and the input data amount of each input is described as an element of each linear list. This is called an input data amount list.

  Of the task inputs, some inputs are directly connected to external inputs of the image processing system. Therefore, for these inputs, non-negative integer values are described as elements of the input data amount list.

  The numerical value indicating the amount of data transferred between tasks is treated as a relative value in the BMNoC configuration algorithm, and its unit may be arbitrary. Therefore, the input data amount does not need to be a real number, and is a non-negative integer here.

  For example, a task with a task ID of 1 has two inputs, and the input data amounts of these inputs are 33 and 17, respectively.

  The unit of the input data amount is arbitrary as described above. For example, the number of bits per time and the number of packets per cycle of the operation of the image processing system can be considered.

  On the other hand, some of the task inputs are not directly connected to the external input of the image processing system, so the amount of input data cannot be determined immediately.

  Therefore, in the initial state, these inputs may be prepared by describing nil as a special value as an element of the input data amount list.

  As will be described later, this nil is eventually replaced with a real number.

  FIG. 6B shows an example of the output data amount table for each component.

  As shown in FIG. 6B, the output data amount table for each component is a table describing a list of output data amounts of each output of the processing component with respect to the component type of the processing component.

  Since each processing component has a unique number of outputs, the outputs are arranged in advance in the order of the linear list, and the output data amount of each output is described as an element of each linear list. This is called an output data amount list.

  Since the same processing component may be used in two or more tasks on the task graph, the output data amount list can be managed for each component type, not for the task ID.

  An element of the output data amount list is an output data amount specific to the processing component.

  In a processing component, the amount of output data may depend on the amount of input data, so it is expressed by a function with the amount of input data as a variable.

  Since each processing component has a unique number of inputs, the inputs are arranged in advance in the order of the linear list, and the respective data amounts are set as variables x1, x2, x3,. For these variables, the output data amount is expressed as a function.

  For example, a processing component with a component type of PRG has two outputs, one of which is expressed as x1 * 50 by the input data amount x1, that is, outputs 50 times the amount of input data. To do.

  Since the output data amount is information specific to the processing component, it can be determined regardless of how the image processing system is configured.

  FIG. 6C shows an example of a task graph connection table.

  As shown in FIG. 6C, the task graph connection table includes, for the task ID of the task, a component type of a processing component that executes the task, an input list of the next task to which output data of the task is input, Is a table describing

  For example, the task with task ID 3 is executed by a processing component whose component type is PRG. In addition, there are two outputs of the task with the task ID 3, and one of the outputs is data to an input whose task ID 7 is the unique number of the task. The next task list is an enumerated list that includes all such elements for a task with a certain task ID.

  If the image processing system can be represented as a configuration diagram of a task graph, all elements of the task graph connection table can be defined.

  FIG. 6D shows an example of a task output data amount table.

  As shown in FIG. 6D, the task output data amount table is a table describing the output data amount of each output of the task with respect to the task ID of the task.

  Since each task has a unique number of outputs, outputs are arranged in advance in the order of a linear list, and the output data amount of each output is described as an element of each linear list. This is called an output data amount list.

  For example, a task with task ID 4 has two outputs, and the output data amounts of these outputs are 100 and 250, respectively.

  The data transfer amount calculation unit 14 is intended to complete the task output data amount table, and inputs for the task output data amount table (initial state), component-specific output data table, and There are three task graph connection tables.

  FIG. 6E shows a flowchart for explaining an algorithm for creating a task output data amount table.

  FIG. 6E is prepared in advance with a task input data amount table (in an initial state), an output data table for each part, and a task graph connection table shown in FIG. 6A, FIG. 6B, and FIG. 6C, respectively. This is an example of creating the task output data amount table shown in FIG. 6D.

  As shown in FIG. 6E, first, the data transfer amount calculation unit 14 prepares a list L. Then, in the task input data amount table, all task IDs having all input data amounts are extracted, and the extracted task IDs are set as initial values of the list L (step A1). The list L is a table having, as elements, task IDs in which the input data amount is all determined in the task input data amount table.

  Next, the data transfer amount calculation unit 14 initializes the task output data amount table (step A2). This initialization state is a state in which the task ID existing in the task graph connection table is a value in the task ID column and each output data amount list is blank.

  Next, the data transfer amount calculation unit 14 extracts one task ID from the list L (step A3). Hereinafter, a task having this task ID is referred to as a target task.

  Next, the data transfer amount calculation unit 14 uses the task ID extracted in step A3 to look up the task graph connection table as a dictionary, and obtains the component type of the processing component that executes the target task and the next task list of the target task. Take out (step A4).

  Next, the data transfer amount calculation unit 14 uses the component type extracted in step A4 to look up a dictionary for the output data amount table for each component, and extracts an output data amount list for the component type of the processing component that executes the target task. (Step A5).

  Next, the data transfer amount calculation unit 14 searches the task input data amount table using the task ID extracted in step A3, and extracts the input data amount list of the target task (step A6).

  Next, the data transfer amount calculation unit 14 extracts in step A6 each function (a function for obtaining the output data amount of each output of component type) that is an element of the output data amount list of the component type extracted in step A5. Calculate the output data amount of each output of the target task by applying the input data amount of each input in the input data amount list, and display the output data amount list that lists the calculated output data amount of each output. It is described in a table (step A7).

  Next, the data transfer amount calculation unit 14 extracts one element (input of the next task) from the next task list extracted in step A4. Here, in the task input data amount table, if the input data amount of the input of the next task is blank (nil), the output data amount of the output corresponding to the input of the next task is output here. Therefore, the output data amount calculated in step A7 is described (step A8). This is performed for all elements of the next task list.

  When the above is completed, the data transfer amount calculation unit 14 checks the list L and determines whether or not there is an unprocessed task ID in L (step A9).

  If there is an unprocessed task ID in the list L in step A9, the data transfer amount calculation unit 14 returns to the process in step A3, extracts one unprocessed task ID from L, and repeats the subsequent processes. .

  On the other hand, if there is no unprocessed task ID in the list L in step A9, the data transfer amount calculation unit 14 determines whether or not there is a blank in the output data amount list column of the task output data amount table. (Step A10).

  In step A10, if there is still a blank in the output data amount list column of the task output data amount table, the output data amount list has not been obtained for all of the plurality of tasks. Therefore, the data transfer amount calculation unit 14 extracts a task ID for which all the input data amounts are newly determined from the task input data amount table, and sets the task ID in the list L (step A11). Then, the data transfer amount calculation unit 14 returns to the process of step A3, extracts one task ID from L, and repeats the subsequent processes.

  On the other hand, if no blank exists in the output data amount list column of the task output data amount table in step A10, the output data amount list is obtained for all of the plurality of tasks. Therefore, the data transfer amount calculation unit 14 ends the process.

  The task output data amount table is determined as described above.

  A task graph can be configured by the task graph connection table and the task output data amount table. Therefore, after that, the BMNoC configuration determination unit 16 determines the arrangement of processing components when the image processing system is set to the BMNoC configuration using the BMNoC algorithm.

  FIG. 7 is a flowchart illustrating an example of a BMNoC configuration algorithm for determining the arrangement of processing components when the image processing system is configured to have a BMNoC configuration.

  As shown in FIG. 7, first, the BMNoC configuration determination unit 16 generates one or more divisions that divide all the processing components into a number equal to or less than the number that can be mounted on the bus (step B1). Any combinations of combinations may be generated in advance, or may be generated sequentially. There are multiple buses, for example, all buses are 5 each, or there are 5 buses and half buses that can be mounted 7 and so on. Give it.

  Next, the BMNoC configuration determining unit 16 takes out one from the dividing methods generated in Step B1 (Step B2).

  Next, when all the processing components are mounted on the bus by the division method extracted in step B2, the BMNoC configuration determining unit 16 outputs the amount of output data of the task executed by the processing components mounted on the bus for all the buses. Is compared with the data transfer capability of the bus (step B3). Here, when calculating the total sum of the output data amounts, an output data amount of a task executed by each processing component mounted on the bus, that is, a task output data amount table is required.

  Next, the BMNoC configuration determining unit 16 determines whether there is a bus whose total output data amount exceeds the data transfer capability (step B4).

  In step B4, if there is a bus whose total output data amount exceeds the data transfer capability, the BMNoC configuration determining unit 16 returns to the process of step B2, extracts another division and repeats the subsequent processes.

  On the other hand, if there is no bus whose total output data amount exceeds the data transfer capability in step B4 and the total output data amount is within the data transfer capability for all buses, the BMNoC configuration determining unit 16 determines whether the buses are connected to each other. For all the transfer paths connected through the switches and routers (that is, transfer paths between the buses), the amount of transfer data transferred through the transfer path is calculated and compared with the data transfer capability of the transfer path ( Step B5). Here, when calculating the transfer data amount, an output data amount of a task executed by each processing component mounted on the bus, that is, a task output data amount table is required.

  Next, the BMNoC configuration determining unit 16 determines whether there is a transfer path in which the transfer data amount exceeds the data transfer capability (step B6).

  In step B6, if there is a transfer path in which the transfer data amount exceeds the data transfer capability, the BMNoC configuration determining unit 16 returns to the process of step B2, takes out another division method, and repeats the subsequent processes.

  On the other hand, if there is no transfer path whose transfer data amount exceeds the data transfer capability in step B6 and the transfer data amount is within the data transfer capability for all transfer paths, the BMNoC configuration determining unit 16 ends the process. Then, the BMNoC configuration determination unit 16 selects a division method that has reached the end of the process, and determines the BMNoC configuration so that all the processing components are mounted on the bus according to the selected division method. The data of the BMNoC configuration determined here is stored in the BMNoC configuration data storage unit 17 by the BMNoC configuration determination unit 16.

  As described above, in this embodiment, the output data amount specific to the processing component is prepared in advance as the component-specific output data table, and the task output data amount is calculated using the component-specific output data table.

  As a result, there is an effect that a real processor and a dedicated simulator for simulating the processor are not required.

  Here, the above effect will be described in detail.

  Conventionally, as a method for calculating the amount of output data of a task, a program having a processing function corresponding to the entire target task graph (or a part of the target graph extracted) (this program has write access to the memory). In this method, the number of write accesses to the memory measured during the processing is converted into the amount of output data of the task. When it is difficult to use a real processor, simulation is performed using a specially developed simulator (such as CoMET from Synopsys).

  On the other hand, in the present embodiment, as described above, since the output data amount of the task is calculated using the component-specific output data table prepared in advance, a real processor or a dedicated processor for simulating the processor is used. There is an effect that a simulator becomes unnecessary.

  In addition, this embodiment has an effect that labor can be reduced compared to the conventional method using a real processor. In the conventional method using a real processor, a task already held as a hardware component is temporarily written in software, the relationship between the output data amount of the task and the memory access is defined, and the memory access is further performed. A library that can measure data is combined with software corresponding to a task, and after preparing appropriate (very large) input data for measurement, measurement is performed and the measurement is also effective. Although there is a trouble that it is necessary to evaluate whether or not there is an operation result, this embodiment does not require this trouble.

  Moreover, this embodiment has the effect that time can be reduced with respect to the conventional method using a dedicated simulator. In the simulator, the measurement of the number of times of memory access is prepared as a library in advance, and the labor for rearranging the number of processors and the configuration between processors is reduced. However, although 100 to 1000 times as long as the method using a real processor is required, this embodiment does not require this time.

  In this embodiment, since the image processing system has a BMNoC configuration, an image processing system having a required area smaller than that of the NoC configuration and having performance equal to or higher than that of the NoC configuration can be obtained.

  Note that Non-Patent Document 1 describes that the BMNoC configuration has performance equal to or higher than that of the NoC configuration. According to Non-Patent Document 1, it is described that latency, which is a typical performance index, can be improved by 50% or more.

  In the present embodiment, the image processing system has a BMNoC configuration that is one of the improved NoC technologies. However, the present invention is not limited to this, and other NoCs, specifically, processing components are used. The present invention can also be applied to other NoCs that are mounted on a bus and connected to each other by a switch and a router.

DESCRIPTION OF SYMBOLS 11 Task input data amount table storage part 12 Output data amount table storage part according to part 13 Task graph connection table storage part 14 Data transfer amount calculation part 15 Task output data amount table storage part 16 BMNoC structure determination part 17 BMNoC structure data storage part 21 Input IF (input interface)
22 Processing parts 23 Output IF (Output interface)
24 routers 30 tasks

Claims (8)

A configuration device for configuring an image processing system that executes each of a plurality of tasks by a processing component,
A first storage unit that stores a task input data amount table describing a list of input data amounts of each input of the task for each task;
A second storage unit that stores, for each processing component, a component-specific output data amount table that describes a list of output data amounts of each output of the processing component;
A third storage unit for storing a task graph connection table describing a processing component for executing the task and a list of inputs of the next task to which the output data of the task is input for each task;
A first process for extracting a task whose input data amount is all determined from the task input data amount table as a target task, and a processing component for executing the target task and the next to the target task from the task graph connection table. A second process for extracting a list of task inputs, a third process for extracting a list of output data amounts of each output of the processing component executing the target task from the component-specific output data amount table, and the task A fourth process for extracting a list of input data amounts for each input of the target task from an input data amount table; and an output of each input of the target task and each output of a processing component that executes the target task based on the data amount, describe and calculate the output data of the output of the target task to calculate the list of output data of the output in the task output data amount table And a sixth process that describes output data of the output corresponding to the target task as the input data amount of the input of the next task of the target task in the task input data amount table. For all of the tasks, a data transfer amount calculation unit that executes sequentially from the task in which the amount of input data is all determined ,
A fourth storage for storing a task output data amount table describing a list of output data amounts of each output calculated by the data transfer amount calculation unit for each task, which is an output data amount of each output of the task A component device. A configuration device for configuring an image processing system that executes each of a plurality of tasks by a processing component,
A first storage unit that stores a task input data amount table describing a list of input data amounts of each input of the task for each task;
A second storage unit that stores, for each processing component, a component-specific output data amount table that describes a list of output data amounts of each output of the processing component;
A third storage unit for storing a task graph connection table describing a processing component for executing the task and a list of inputs of the next task to which the output data of the task is input for each task;
Based on the task input data amount table, the component-specific output data amount table, and the task graph connection table, for each task, a data transfer amount calculation unit that calculates the output data amount of each output of the task;
A fourth storage for storing a task output data amount table describing a list of output data amounts of each output calculated by the data transfer amount calculation unit for each task, which is an output data amount of each output of the task And
When the image processing system has a specific NoC configuration in which the processing components are mounted on a bus and the buses are connected by a switch and a router, based on the task graph connection table and the task output data amount table, A configuration determining unit that determines an arrangement of processing components . When the image processing system has a specific NoC configuration in which the processing components are mounted on a bus and the buses are connected by a switch and a router, based on the task graph connection table and the task output data amount table, The configuration apparatus according to claim 1, further comprising a configuration determination unit that determines an arrangement of processing parts. The configuration determining unit
Generate a way to divide all processing parts into less than the number that can be mounted on the bus,
When all the processing components are mounted on the bus in the division method, the total amount of output data of tasks executed by the processing components mounted on the bus is the data of the bus. For the first condition that the transfer capacity is within the transfer capacity and all the transfer paths that connect the buses and pass through the switch and the router, the transfer data amount transferred through the transfer path is the data transfer capacity of the transfer path Select a split method that satisfies the second condition of being within,
The component device according to claim 2 , wherein the arrangement of the processing components is determined so that all the processing components are mounted on the bus according to the selected division method. A configuration method by a configuration device for configuring an image processing system that executes each of a plurality of tasks by a processing component,
For each task, a first step of describing a list of input data amounts of each input of the task in a task input data amount table;
A second step of describing, for each processing component, a list of output data amounts for each output of the processing component in a component-specific output data amount table;
A third step for describing, in the task graph connection table, a processing component for executing the task and a list of inputs of the next task to which output data of the task is input for each task;
Based on the task input data amount table, the component-specific output data amount table, and the task graph connection table, for each task, an output data amount of each output of the task is calculated, and the task input data amount table From the task graph connection table, a processing component for executing the target task and a list of inputs of the next task after the target task, From the second process for retrieving the output data, the third process for retrieving the output data amount list of each output of the processing component that executes the target task from the component-specific output data amount table, and the task input data amount table, A fourth process for retrieving a list of input data amounts of each input of the target task, and a process of executing the input data amount of each input of the target task and the target task The output data amount of each output of the target task is calculated based on the output data amount of each output of the product, and a list of calculated output data amounts of each output is described in the task output data amount table. And a sixth process for describing output data of output corresponding to the target task as an input data amount of an input of the next task of the target task in the task input data amount table. For all the tasks, the input data amount is sequentially executed from the task in which all the input data amounts are determined, and for each task, the output data amount of each output possessed by the task and the calculated output data amount list of each output are displayed as task output data. And a fourth step described in the quantity table. A configuration method by a configuration device for configuring an image processing system that executes each of a plurality of tasks by a processing component,
For each task, a first step of describing a list of input data amounts of each input of the task in a task input data amount table;
A second step of describing, for each processing component, a list of output data amounts for each output of the processing component in a component-specific output data amount table;
A third step for describing, in the task graph connection table, a processing component for executing the task and a list of inputs of the next task to which output data of the task is input for each task;
Based on the task input data amount table, the component-specific output data amount table, and the task graph connection table, the output data amount of each output of the task is calculated for each task, and the calculated output of each output A fourth step of describing a list of data amounts in the task output data amount table;
When the image processing system has a specific NoC configuration in which the processing components are mounted on a bus and the buses are connected by a switch and a router, based on the task graph connection table and the task output data amount table, And a fifth step of determining an arrangement of the processing components . When the image processing system has a specific NoC configuration in which the processing components are mounted on a bus and the buses are connected by a switch and a router, based on the task graph connection table and the task output data amount table, The configuration method according to claim 5, further comprising a fifth step of determining an arrangement of the processing components. In the fifth step,
Generate a way to divide all processing parts into less than the number that can be mounted on the bus,
When all the processing components are mounted on the bus in the division method, the total amount of output data of tasks executed by the processing components mounted on the bus is the data of the bus. For the first condition that the transfer capacity is within the transfer capacity and all the transfer paths that connect the buses and pass through the switch and the router, the transfer data amount transferred through the transfer path is the data transfer capacity of the transfer path Select a split method that satisfies the second condition of being within,
The configuration method according to claim 6 or 7, wherein the arrangement of the processing components is determined so that all the processing components are mounted on the bus according to the selected division method.