JP2013150094A - Task scheduling device, task scheduling method, and task scheduling program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a task scheduling device which schedules tasks having more overlapping memory access regions than other tasks in such a way that they can be executed in succession.SOLUTION: Task information including task specific information for identifying a task is acquired in plurality, and access information indicating the arrangement of memory regions to be accessed in the task is acquired, which information is stored in access region storage means after being correlated to the task specific information. From the access information on two of the tasks, task correlation indicating the overlapping amount of memory regions to be accessed in two of the tasks is calculated. To ensure that, among combinations of the task at the top or the end of a task queue which is the queue of the tasks in which order they are executed with any of the tasks not included in the task queue, the tasks of combinations where the task correlation is highest are contiguous, the tasks not included in the task queue are added to the task queue and information indicating the task queue is output.

Description

  The present invention relates to a task scheduling apparatus, a task scheduling method, and a task scheduling program, and more particularly to a task scheduling apparatus, a scheduling method, and a scheduling program for an arithmetic unit having a cache memory.

  In recent years, with the demand for increasing the image size of moving images, the encoding techniques are MPEG-2 (Moving Picture Experts Group-2), MPEG-4, H.264 / MPEG-4 AVC (Advanced Video Coding, H.264). ) And have progressed to improve compression efficiency. These encoding methods achieve high encoding efficiency by compressing information using motion compensation prediction between frames. FIG. 20 is a diagram illustrating an example of a video encoding device based on the H.264 standard. Non-Patent Document 1 describes the contents of processing based on this standard. Further, Non-Patent Document 2 describes details of implementation of a moving image encoding device based on this standard.

  Motion compensation prediction is a technique for generating an image (predicted image) obtained by correcting an image of a reference frame using motion information between an encoding target frame and a reference frame. The moving image encoding apparatus compresses moving image information by encoding only a difference between a prediction image generated by motion compensated prediction and an encoding target image and a motion vector representing a motion between frames. In motion compensated prediction, an image to be encoded is predicted from an encoded image. A frame having one reference image is a P (Predictive) frame, and a frame having two reference images is a B (Bi-predictive) frame. The video encoding apparatus can select a reference frame for each block from a plurality of frames.

  A process for calculating motion information between the reference frame and the encoding target frame is motion estimation. The video encoding apparatus performs motion estimation by searching for a block having high similarity to the encoding target block from the reference frame. The position difference between the block with the highest similarity of the reference frame and the encoding target block is a motion vector. When there are a plurality of referable frames that are reference frame candidates, the video encoding apparatus performs a motion vector search in each referable frame. The image encoding apparatus selects a referenceable frame including a block having the highest similarity as a result of the search, and sets it as a reference frame, and sets a position difference between the block having the highest similarity and the encoding target block as a motion vector. The process of generating a predicted image using a reference image and a motion vector is called motion compensation.

  H.264 video coding processing including motion compensation prediction is performed in units of 16 × 16 pixel macroblocks. Since moving picture coding processing by recent coding schemes such as H.264 has a large amount of computation, speeding up is often achieved by parallel processing. Since motion estimation and motion compensation are almost independent processes for each macroblock, these processes can be performed in parallel on a macroblock basis.

  In motion compensation prediction, it is known that adjacent motion vectors have high correlation. Therefore, the coding amount of the motion vector is encoded by encoding only the difference between the prediction vector and the motion vector of the encoding target block using the motion vector information of the neighboring blocks already searched as a prediction vector. Can be reduced. The prediction vector in H.264 is the median of the motion vectors of the left, top, and top right blocks after processing.

  In motion estimation, as described in Non-Patent Document 2, an adaptive search algorithm is used in which a motion vector is searched from a point close to the search start point, and if it can be determined that a sufficiently good motion vector has been obtained, the search is terminated. There are many cases. Therefore, when searching for a motion vector, if the distance between the motion vector as a search result and the search start point is short, the search amount is small, and if the distance is long, the search amount is large. Since the prediction vector is predicted to be close to the motion vector, the search amount can be reduced by using the prediction vector as a search start point. If the prediction vector is a search start point, the search amount changes depending on the accuracy of the prediction vector, that is, the distance between the prediction vector and the motion vector.

  Next, the technical background in the processor architecture will be described.

  A processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) reads out data stored in a main memory and performs arithmetic processing. In general, since main memory access is slow for arithmetic processing, many processors are provided with a cache memory for reducing the access load on the main memory. The cache memory is smaller in capacity than the main memory, but can be accessed at high speed. The cache memory holds the contents of the main memory accessed in the past by the processor. A processor having a cache memory can access data at high speed without accessing the main memory and accessing only the cache memory if the memory access is temporally and spatially local.

  In the above-described motion estimation and motion compensation, since the amount of memory access is large with respect to the amount of calculation, the cache hit rate that indicates the ratio that only requires access to the cache memory greatly affects the processing time of these processes. Therefore, when executing a task for performing motion estimation or motion compensation, task scheduling that can continuously execute a task using the same data is important. Furthermore, in a general multi-core / meni-core parallel processing environment, each core or each core group that combines multiple cores has a cache, and tasks that use the same data are as much as possible so that the cache can be used effectively. Scheduling such as placing in the core is also required. In particular, when the image size is large, all of the image data cannot be held in the cache memory, and it is important to reuse a large amount of data once stored in the cache.

  In parallel processing, it is important for speeding up to reduce the time in which each processor core is in an idle state as much as possible by evenly distributing the load of each processor core. In addition, many core processors such as GPUs have more cores than multi-core CPUs, and the number of tasks that can be executed in parallel is large. The impact will be greater.

  Examples of techniques for improving the cache hit rate of motion compensation processing are described in Patent Document 1, Patent Document 2, and Patent Document 3. These techniques are techniques that increase the cache hit rate by collecting tasks that have similar memory access areas and executing the processes collectively while data is in the cache.

  FIG. 21 is a diagram illustrating an example of a frame structure in the encoding process. In the example of FIG. 21, P1, P2, and P3 represent P frames, and B1, B2, B3, and B4 represent B frames. In encoding processing including bi-predictive motion compensation, two B frames are sandwiched between P frames as shown in FIG. 21, and each B frame often has a frame structure that refers to the same P frame before and after.

  In the following description, a task executing a task accesses, for example, at least a part of the data in the frame b while executing the task a, and the task a refers to the frame b. In this case, the frame b is expressed as a frame accessed by the task a or a frame accessed by the task a. Further, a task executing a task accesses a certain memory area while executing the task a, for example, and indicates that the task a accesses the memory area. In this case, this memory area is referred to as a memory area accessed by task a or a memory area accessed by task a.

  Patent Document 1 describes a moving image decompression method that effectively uses a cache by continuously processing the same-position macroblocks of two consecutive B pictures. The method described in Patent Document 1 improves the cache hit rate by continuously processing the same reference frame.

  Patent Document 2 describes a motion compensation device that collects and continuously executes tasks that refer to the same frame so that a cache hits even when a plurality of frames can be referred to and the reference frame is different for each block. ing. The above-described task represents processing performed with reference to one frame in any block. The technique described in Patent Document 2 increases the cache hit rate by changing the task processing order more flexibly than the technique described in Patent Document 1 and executing more tasks that refer to the same frame continuously. Can do.

  FIG. 22 is a diagram illustrating an example of assignment of tasks to processor cores by the blocks referred to by the tasks and the motion compensation apparatus disclosed in Patent Document 2. In FIG. 22, for example, a task 1-1 represents a task that refers to the reference frame 1 in a block in the encoding target frame, and tasks 2-1 to 2-3 are references in the block in the encoding target frame. Represents a task that references frame 2. These tasks are not necessarily processing in the same block. PE (Processor Element) 1 to PE4 each represent a processor core. In FIG. 22, for example, task 1-1 is performed by PE1, and tasks 2-1 to 2-3 are performed by PE2.

  When performing parallel processing, the motion compensation device described in Patent Document 2 can repeatedly use data in a cache for each core by performing processing referring to the same frame with the same processor core as shown in FIG. it can.

  Patent Document 3 describes a decoding device based on a technique for increasing a cache hit rate by continuously executing processing of blocks that refer to the same area of the same frame. The decoding device of Patent Document 2 calculates the access range of each task, and increases the cache hit rate by scheduling the task so that tasks in the same access range are continuously executed. The technique described in Patent Document 3 schedules tasks so that tasks of different frames that access the same memory area are continuously executed so that the cache hit rate is higher than the technique described in Patent Document 2. can do.

  FIG. 23 shows an example of the relationship between a block in a reference frame and a block in an encoding target frame in which a task that refers to the block is executed, and the processing order of those tasks determined by the decoding device of Patent Document 3. It is a figure showing the example of. In FIG. 23, for example, task 1-1 represents a task executed in block 1-1 of encoding target frame 1, and task 2-1 represents a task executed in block 2-1 of encoding target frame 2. . The result of scheduling the tasks 1-1 to 3-4 in FIG. 23 by the decoding apparatus of Patent Document 3 is the processing order in FIG.

  FIG. 25 is a diagram illustrating an example of a configuration of a task scheduling system that schedules tasks using the technique described in Patent Document 3. The task scheduling system shown in FIG. 25 includes a task acquisition unit that acquires a task, a task access area calculation unit that calculates an access area of the acquired task, and a memory area where the task access area is arbitrarily divided. It includes a task grouping unit that determines whether or not a task is included and groups tasks for each divided memory area, and a schedule unit that sorts the processing order of tasks for each group.

  FIG. 26 is a diagram illustrating an operation example of the task scheduling system in FIG. 25 according to the technique described in Patent Document 3. FIG. 3 is a diagram illustrating an example of a memory range accessed by the tasks A, B, and C. An example of scheduling tasks A, B, and C that access a memory in the range shown in FIG. 3 will be described. The memory is divided into areas represented by (1) and (2) in the figure, and some of the divided memory areas overlap. First, the task acquisition unit acquires a task, for example, task A. The task access area calculation unit calculates the access area of task A. Since the access area of task A is included in the memory area of (1), the task grouping unit places task A in group G1. This is repeated until all tasks are grouped. Task B is arranged in group G1 in the same manner as task A. Since task C is included in the memory area (2), task C is placed in group G2. When the grouping of all tasks is completed, the schedule unit sorts the tasks using the group number as a key. For example, the tasks are sorted in the order of tasks A, B, and C, and the tasks are processed in the sorted order. The processing order of tasks within a group and the processing order of groups depend on the implementation.

  Patent Document 4 describes a system analysis apparatus that divides processes and data items and displays an interface between the processes based on the type and number of access instructions for the data items examined at each step of the program.

  Patent Document 5 describes a cache control device that associates an area obtained by dividing a cache memory with each of a plurality of tasks to be processed in parallel. In the technique described in Patent Document 5, each task uses only an associated cache memory area. As a result, the data in the cache memory area used by each task is not overwritten by another task.

  Patent Document 6 describes a performance monitoring method in which related processes are grouped according to a file to be accessed by the process, and operation performance is monitored in units of the process group.

JP 2000-278893 A JP 2011-063120 A WO2008 / 038513 JP 2001-092651 A JP 2004-178571 A JP 2009-265707 A

ITU-T Recommendation H.264 `` Advanced video coding for generic audiovisual services, '' May 2003. K. P. Lim, G. Sullivan and T. Wiegand, "Text Description of Joint Model Reference Encoding Methods and Decoding Concealment Methods", Joint Video Team (JVT) of ISO / IEC MPEG and ITU-T VCEG, Busan, Korea, Apr. 2005

  The technique described in Patent Document 1 has a problem that when the reference frame is different for each block, a cache hit is difficult and the effect is small.

  The technique described in Patent Document 2 has a problem that when the image size is increased, even in the same frame, it is not always placed in the cache, and cache misses may occur frequently. The technique described in Patent Document 2 further has a problem that it is difficult to equalize the processing amount per processor core in the case of parallel processing because the number of tasks per reference frame is not constant.

  The task scheduling method described in Patent Document 3 groups tasks that access a predetermined memory area so that they are continuously executed. For example, when scheduling a task that accesses a range as shown in FIG. 3, in order to improve the cache hit rate, it is necessary to continuously process tasks A and C having overlapping memory access areas. However, when the tasks A, B, and C are scheduled by the task scheduling method described in Patent Document 3, the tasks A and B extracted as tasks that access the area of (1) are processed in succession. Thereafter, the task C extracted as a task accessing the area (2) is processed. At this time, the processes of tasks A and B and task C are not always performed continuously. Since the memory access areas of task A and task B do not overlap, the cache memory cannot be used. On the other hand, since task A and task C are not processed consecutively even though part of the memory access area overlaps, the data in the cache cannot be reused.

  As described above, the technique of Patent Document 3 groups only whether or not the memory area accessed by the task is in the same memory area. The technique of Patent Document 3 has a problem that a plurality of tasks cannot be scheduled so that tasks having overlapping memory access areas are continuously executed.

  The techniques disclosed in Patent Documents 4, 5, and 6 have a problem that a plurality of tasks cannot be scheduled so that tasks having overlapping memory areas to be accessed are continuously executed.

  It is an object of the present invention to provide a task scheduling apparatus that schedules a task in which more memory access areas overlap so that they can be continuously executed.

  A task scheduling apparatus according to the present invention acquires task information that acquires a plurality of task information including task specifying information that specifies a task that is a unit of processing, and access information that indicates an arrangement of memory areas accessed by the task. From the task access acquisition means for storing in the access area storage means in association with the task identification information, the access area storage means for storing a set of the task identification information and the access information, and the access information of the two tasks, A task correlation calculating means for calculating a task correlation representing an overlapping amount of memory areas accessed by the two tasks, a first or last task in a task sequence which is a sequence of execution order of the tasks, and a task sequence Before the combination having the highest task correlation among combinations with any of the tasks not included As the task is continuous, it comprises a scheduling means for adding said tasks that are not included in the task sequence in the task sequence, and output means for outputting information representing the task sequence.

  The task scheduling method of the present invention acquires a plurality of task information including task specifying information for specifying a task that is a unit of processing, acquires access information indicating an arrangement of a memory area accessed by the task, and specifies the task Memory associated with information and stored in access area storage means, a set of the task specifying information and the access information is stored in access area storage means, and the memory accessed by the two tasks from the access information of the two tasks Calculating a task correlation representing an overlap amount of the area, and combining the first or last task in the task sequence, which is a sequence of the execution order of the tasks, with any of the tasks not included in the task sequence, The tasks that are not included in the task sequence are assigned to the tasks so that the tasks of the combination having the highest correlation are consecutive. In addition to the click string and outputs the information representing the task sequence.

  The task scheduling program of the present invention includes a task acquisition unit that acquires a plurality of task information including task specifying information for specifying a task that is a unit of processing, and access information that indicates an arrangement of memory areas accessed by the task A task access acquisition means for acquiring and storing in the access area storage means in association with the task specification information, an access area storage means for storing a set of the task specification information and the access information, and the access of the two tasks From the information, a task correlation calculating means for calculating a task correlation representing an overlap amount of memory areas accessed by the two tasks, the first or last task in the task sequence which is a sequence of the execution order of the tasks, In combination with any of the tasks not included in the task column, the task correlation is Even as the task of high the combination is continuous, and schedule means for adding the tasks that are not included in the task sequence in the task sequence, it is operated as an output means for outputting information representing the task sequence.

  According to the present invention, when a plurality of tasks including tasks having overlapping memory areas to be accessed are executed, the cache hit rate is improved and the task execution speed is increased.

It is a block diagram showing the structure of the task scheduling apparatus 1 of 1st, 2nd embodiment. It is a flowchart showing operation | movement of the task scheduling apparatus 1 of 1st Embodiment. It is a figure showing the example of the access area of a task. It is a figure showing the 2nd example of the access area of a task. It is a flowchart showing the example of operation | movement of the task scheduling apparatus 1 of 2nd Embodiment. It is a figure showing the example of the prediction vector of a some task. It is a figure showing the structure of the task scheduling apparatus 3 of 3rd Embodiment. It is a flowchart showing operation | movement of the task scheduling apparatus 3 of 3rd Embodiment. It is a figure showing the example of the access area of the some task in which parallel processing is possible. FIG. 10 is a diagram illustrating an access area accessed by the task in the example of FIG. It is a figure showing the correlation between tasks of the task of the example of FIG. FIG. 10 is a diagram illustrating a result of grouping the tasks of FIG. 9 by the task scheduling device 4. It is a figure showing the structure of the task scheduling apparatus 4 of 4th Embodiment. It is a flowchart showing the example of operation | movement of the task scheduling apparatus 4 of 4th Embodiment. It is a figure showing the example of assignment of the task to a plurality of processor cores. It is a figure showing the structure of the task scheduling apparatus 5 of 5th Embodiment. It is a flowchart showing the example of operation | movement of the task scheduling apparatus 5 of 5th Embodiment. It is a figure showing the structure of the task scheduling apparatus 6 of 6th Embodiment. It is a flowchart showing the example of operation | movement of the task scheduling apparatus 6 of 6th Embodiment. 1 is a diagram illustrating an example of a video encoding device based on the H.264 standard. It is a figure showing the example of the frame structure in an encoding process. FIG. 10 is a diagram illustrating an example of task reference frames and assignment of tasks to processor cores by the motion compensation device of Patent Literature 2. An example of a relationship between a block in a reference frame and a block in an encoding target frame in which a task that refers to the block is executed, and an example of processing order of those tasks determined by the decoding device of Patent Document 3 FIG. It is a figure showing the example of the result of having performed the grouping of the task by the technique of patent document 3, and the allocation of the task group to the processor core. FIG. 10 is a diagram illustrating an example of a configuration of a task scheduling system that schedules tasks by the technique described in Patent Literature 3. 10 is a diagram illustrating an example of an operation of a task scheduling system according to a technique described in Patent Literature 3. FIG. It is a figure showing the structure of the task scheduling apparatus 7 of 7th Embodiment. It is a flowchart showing the example of operation | movement of the task scheduling apparatus 7 of 7th Embodiment. FIG. 3 is a block diagram illustrating an example of a configuration of an information processing device 1000.

  Next, embodiments of the present invention will be described in detail with reference to the drawings.

(First embodiment)
FIG. 29 is a block diagram illustrating an example of the configuration of the information processing apparatus 1000 including the task scheduling apparatus 1 according to the first embodiment of this invention.

  Referring to FIG. 29, the information processing apparatus 1000 includes a task generation unit 1001, a task scheduling apparatus 1, a task execution unit 1002, a processor 1003, an execution result output unit 1004, a data input unit 1006, and a data storage unit. 1007 is included. The information processing apparatus 1000 may include a motion vector storage unit 1005.

  The data input unit 1006 receives input data and stores it in the data storage unit 1007. The input data is, for example, moving image data including a plurality of time-series frames.

  The data storage unit 1007 stores input data that has been input.

  The task generation unit 1001 generates a plurality of sets of tasks to be executed by the processor 1003 and task information representing the tasks from input data stored in the data storage unit 1007. A task is a group of processes for input data stored in the data storage unit 1007. The task is, for example, motion estimation processing in a certain block (encoding target block) included in a certain frame (encoding target frame) in the moving image. The task may be, for example, a motion compensation process in a certain block included in a certain frame in the moving image.

  The task scheduling apparatus 1 receives task information of a plurality of tasks, determines the task execution order (task sequence) as described later, and outputs information representing the task execution order.

  The task execution unit 1002 receives a task from the task generation unit 1001 and receives information indicating the task execution order from the task scheduling apparatus 1. The task execution unit 1002 causes the processor 1003 to execute each task according to the execution order determined by the task scheduling apparatus 1.

  The processor 1003 executes a task in accordance with an instruction from the task execution unit 1002. The processor 1003 includes a cache memory. The processor 1003 sequentially transmits task execution result data to the execution result output unit 1004.

  The execution result output unit 1004 receives information indicating the result of the task executed by the processor 1003 from the processor 1003 and outputs the information. When the task type is motion estimation, the execution result output unit 1004 stores the calculated motion vector of the encoding target block in association with the information for specifying the encoding target block in the motion vector storage unit 1005. May be.

  The motion vector storage unit 1005 stores a motion vector associated with information for specifying an encoding target block.

  Note that the information processing apparatus 1000 may not include the task scheduling apparatus 1. In this case, the task scheduling apparatus 1 and the information processing apparatus 1000 communicate with each other so that the task scheduling apparatus 1 can receive task information from the information processing apparatus 1000 and transmit information indicating the execution order to the information processing apparatus 1000. It is only necessary to be connected. In other embodiments described later, the information processing apparatus 1000 has a configuration in which the task scheduling apparatus 1 is replaced with the task scheduling apparatus of each embodiment.

  Hereinafter, an example in which the task processing is motion estimation processing will be described in detail. However, the task processing may be processing other than motion estimation, such as motion compensation or other processing capable of estimating an access area that is a memory area accessed by the task, or processing in which the access area is known. The task scheduling apparatus 1 can perform task scheduling in a similar manner if the task access area is known at the time of scheduling. The task access area may be a predicted value. In the example described in this embodiment, one task is processing in one macroblock. The task scheduling apparatus 1 performs scheduling with processing in one macroblock as one task.

  In the present invention, the amount of duplication of memory access areas between tasks is used as the correlation between tasks, and the cache is effectively used by continuously executing tasks with high correlation.

  FIG. 1 is a diagram illustrating a configuration of a task scheduling apparatus 1 according to the first embodiment of this invention.

  Referring to FIG. 1, a task scheduling apparatus 1 according to the present embodiment includes a task acquisition unit 101, a task access acquisition unit 102, an access information storage unit 105, a task correlation calculation unit 103, a schedule unit 104, and an output unit. 106. The task scheduling apparatus 1 may include a task access estimation unit 107.

  The task acquisition unit 101 inputs task information of a plurality of tasks to be scheduled.

  The task information that represents the task only needs to include task identification information that identifies the task. The task information may include information indicating the type of task such as motion estimation or motion compensation task. The task information may include information indicating the access area of the task. The information representing the access area may be access information representing (identifying) the arrangement of the access area. The information indicating the access area may be information for calculating the arrangement of the access area.

  The task information may include information for calculating the arrangement of the access areas, and the task scheduling apparatus 1 may be configured to include the task access estimation unit 107. The task access estimation unit 107 calculates the access area arrangement from the information for calculating the access area arrangement included in the task information. The task access estimation unit 107 transmits information identifying the task and access information indicating the calculated access area arrangement to the task access acquisition unit 102.

  Further, when the task information includes access information for specifying the arrangement of the access area, the task acquisition unit 101 may transmit the task specification information and the access information to the task access acquisition unit 102. If the task information does not include information indicating the access area of the task, the task acquisition unit 101 may transmit the task identification information to the task access acquisition unit 102.

  The task access acquisition unit 102 acquires access information of the task for which the task acquisition unit 101 has acquired task information. The task access acquisition unit 102 stores the task identification information and the access access information in the access area storage unit 105 in association with each other.

  The access area storage unit 105 stores a set of task specifying information and access information.

  The task correlation calculation unit 103 calculates a correlation between two tasks (task correlation) from the access information stored in the access area storage unit 105. The task correlation calculation unit 103 may determine the correlation between two tasks based on, for example, the size of the overlapping portion of the access areas of those tasks.

  The schedule unit 104 determines the task execution order so that two tasks having a high correlation between tasks are continuously executed.

  The output unit 106 outputs information representing the determined execution order.

  Next, the operation of the present embodiment will be described in detail with reference to the drawings.

  FIG. 2 is a flowchart showing the operation of the task scheduling apparatus 1 of this embodiment.

  FIG. 3 is a diagram illustrating an example of a task access area. The example of the operation described below is an example of the operation when the task scheduling apparatus 1 schedules tasks A, B, and C that access a range as shown in FIG.

  First, the task acquisition unit 101 acquires task information of a task, for example, task A (step S101).

  As described above, the task information may include access information indicating the arrangement of access areas. In this case, the task acquisition unit 101 may transmit access information to the task access acquisition unit 102. Further, the task information may include information for calculating the arrangement of access areas. When the task information includes information for calculating the access area arrangement, after the operation of step S101, the task access estimation unit 107 determines the access area information from the information for calculating the access area arrangement included in the task information. Calculate the placement. Then, the task access estimation unit 107 may transmit access information indicating the access area arrangement to the task access acquisition unit 102.

  For example, if the access area is arranged on the memory and the data accessed and used by the task is two-dimensional data such as an image and the access area is represented by a rectangle, the access area is arranged on the memory. The coordinates may be the upper left and lower right coordinates of the access area. If the data that the task accesses and uses in the memory is one-dimensional data, the access area can be arranged at the start and end addresses of the data, or the start and end addresses of the data. It only needs to be long.

  If the input data is a moving image data in which a plurality of frames that are two-dimensional images are arranged in time series, and the type of task is motion estimation, the task is performed in the encoding target block on the encoding target frame. This process is performed by referring to data on the reference frame. In this case, the information indicating the location of the access area may be information for specifying the reference frame and, for example, the upper left and lower right coordinates of the access area in the reference frame. In this case, each frame included in one moving image has the same size, and the coordinates in each frame are coordinates in a coordinate system in which the overlapping point is the origin when the frames are overlapped and the coordinate axes are in the same direction. If it is. For example, in all frames, the coordinates of the points in the frame may be represented by a coordinate system in which the upper left corner point is the origin, the right direction is the x-axis direction, and the lower direction is the y-axis direction. The information indicating the access area arrangement may be information indicating the address of the memory in which the access area in the reference frame is stored.

  Further, when the task information includes information for calculating the arrangement of the access area, the information for calculating the arrangement of the access area includes, for example, coordinates indicating the position of the encoding target block on the encoding target frame, Any information that identifies the reference frame may be used. Information for calculating the arrangement of access areas may include a prediction vector. Alternatively, the task access estimation unit 107 may use any existing method such as a method in which the median of the motion vector obtained as a result of the search in the left, upper, and upper right blocks of the encoding target block is used as a prediction vector, A prediction vector may be calculated. The coordinates representing the position of the encoding target block may be the coordinates of the upper left point in the encoding target frame of the encoding target block, for example. The task searches a search range centered on a point moved by the predicted vector from the same coordinates that represent the position of the encoding target block on the reference frame. In other words, the task accesses a memory area (access area) corresponding to the search range. The size of the search range may be a predetermined size.

  In this case, the task access estimation unit 107 may calculate the access area arrangement from the information specifying the reference frame, the coordinates representing the position of the encoding target block, and the prediction vector.

  Next, the task access acquisition unit 102 acquires the access information of the task that acquired the task information (for example, the above-described task A) (step S102), associates the task specifying information of the task with the acquired access information, Stored in the access area storage unit 105.

  The task access acquisition unit 102 acquires access information from the task acquisition unit 101 and the task access estimation unit 107. For example, the task access acquisition unit 102 calculates the access information stored in the storage unit for storing the access information of each task from the storage unit, which is calculated by the task generation unit 1001 and is not illustrated. You may read.

  The task scheduling apparatus 1 repeats the operations in steps S101 to S102 until the task access acquisition unit 102 acquires access information for all tasks (No in step S103).

  When the task access acquisition unit 102 completes acquisition of access information for all tasks (step S103, Yes), the task correlation calculation unit 103 calculates a correlation between tasks (step S104).

  First, the task correlation calculation unit 103 may calculate the correlation of all the combinations of two tasks whose access information is stored in the access area storage unit 105 from the access information read from the access area storage unit 105.

  If the access information is information indicating the position and range of the access area, such as the coordinates of the upper left and lower right points of the access area, for example, the task correlation calculation unit 103 determines the access from the access information of the two tasks. What is necessary is just to calculate the overlap amount of an area | region. The overlap amount is, for example, the size of the memory of the overlap portion. The task correlation calculation unit 103 may set the calculated overlap amount as the correlation between two corresponding tasks. The access information may be information indicating the position of the access area (for example, coordinates indicating the center of the access area). And the task correlation calculation part 103 is good also as the correlation of a task by calculating the value showing the grade of the duplication amount of two tasks from the access information like 2nd Embodiment mentioned later.

  When the task correlation is an overlapping amount of access areas, in the example of the task in which the access area is the area of FIG. 3, the correlation between task A and task B and task B and task C is zero. The correlation between task A and task C is 6.

  Next, the scheduling unit 104 determines the task execution order so that a set of tasks having a high correlation continues (step S105).

  First, for example, the schedule unit 104 selects a set of tasks having the highest correlation and determines the execution order of these two tasks. The schedule unit 104 may determine the execution order of the two tasks having the highest correlation. The schedule unit 104 sets the task whose execution order is determined as scheduled.

  When the scheduling of all tasks has not been completed (No at Step S106), the scheduling unit 104 selects a task having the highest correlation with a task that has already been scheduled (Step S105). The schedule unit 104 determines the execution order of tasks so that tasks with high correlation are executed continuously.

  When there is a task that has been selected and scheduled, the scheduling unit 104 may select a task having the highest correlation with the last selected and scheduled task. When selecting the next task of the two tasks having the highest correlation, the scheduling unit 104 may select the task having the highest correlation with one of the two tasks. Then, the scheduling unit 104 may determine the execution order so that the task having the higher correlation with the next selected task among the already selected tasks is executed later.

  In the example of FIG. 3, the correlation between task A and task C is the highest, and the correlation between task A and task B and task B and task C is zero. In this case, the execution order determined by the schedule unit 104 is the order of A, C, B or the order of C, A, B.

  FIG. 4 is a diagram illustrating a second example of task access areas. In this example, the access areas of tasks A, B, and C are the same as in the example of FIG. In the example of FIG. 4, there is a task D whose correlation with the task C is 1. The correlation between task A and task D and the correlation between task B and task D are 0 respectively.

  In this example, the correlation between task A and task C is the highest, and then the correlation between task C and task D is high. Therefore, the schedule unit 104 first selects task A and task C, and assumes that scheduling has been completed. Next, the scheduling unit 104 selects the task D having the highest correlation with the task A or the correlation with the task C among the tasks not yet selected. Since the correlation between task C and task D is higher than the correlation between task A and task D, in addition to task A and task C, scheduling unit 104 executes task C and task D continuously. The execution order of task A, task C, and task D is determined. The scheduling unit 104 may determine the execution order so that task A, task C, and task D are in order. Next, the scheduling unit 104 may select a task B whose correlation with the task A or the correlation with the task C is lower than the task D, and determine the execution order so that the task B is executed last.

  The scheduling unit 104 may select a task using not only the correlation with the task selected last and scheduled, but also the correlation with other scheduled tasks. For example, the schedule unit 104 may calculate a weighting factor that has a greater influence as the correlation with the newly scheduled task is increased according to the distance (time) between tasks in the task processing order. The scheduling unit 104 may select a task based on the sum of values obtained by multiplying the calculated coefficient by the correlation with each scheduled task.

  In addition, the scheduling unit 104 may select a task having the highest correlation with the first task or the last task in the execution order among the tasks whose execution order has already been determined. Then, when the correlation between the selected task and the first task in the execution order is the highest, the scheduling unit 104 sets the execution order so that the execution order executes the task selected before the first task. You may decide. In addition, when the correlation between the selected task and the last task in the execution order is the highest, the scheduling unit 104 determines the execution order so that the selected task is executed after the last task. May be.

  When the scheduling unit 104 selects all tasks and the determination of the execution order of all tasks is completed (Yes in step S106), the output unit 106 outputs information indicating the execution order (step S107).

  The present embodiment described above has an effect of improving the cache hit rate and increasing the task execution speed when executing a plurality of tasks including tasks having overlapping memory areas to be accessed.

  The reason is that the task correlation calculation unit 103 calculates the overlap of the memory areas accessed by the task, and the task set having a high correlation between tasks is preferentially executed continuously. This is because the unit 104 determines the task execution order. When the processor executes tasks according to the execution order determined by the schedule unit 104, the cache hit rate is improved and the cache memory can be used effectively. As a result, the processing speed of the task is increased.

  For example, when a plurality of tasks in the example of FIG. 3 is scheduled with the technique of Patent Document 3, the execution order cannot be determined so that the tasks A and C having overlapping memory areas to be accessed are continuously executed. . On the other hand, when the task scheduling apparatus of the present embodiment similarly performs the task scheduling in the example of FIG. 3, the task execution order is such that task A and task C having overlapping memory areas to be accessed are executed in succession. In order. As a result, the cache data can be used effectively and the cache hit rate is improved.

(Second Embodiment)
Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

  Hereinafter, the case where the task type is motion estimation will be described. When the task type is motion estimation, as will be described later, the task correlation calculation unit 103 according to the present embodiment uses the prediction vector information of each task, which is one piece of information used for motion estimation, between tasks. Is calculated. The type of task may be motion compensation. When the task type is motion compensation processing, the prediction vector in the following description may be read as the motion vector.

  The configuration of the task scheduling apparatus 1 of the present embodiment is the same as the configuration of the task scheduling apparatus 1 of the first embodiment of FIG.

  However, in the first embodiment, the information for calculating the arrangement of the access area includes coordinates indicating the position of the encoding target block (for example, the coordinates of the upper left corner point of the encoding target block) and the task type. In the case of motion estimation, it is a prediction vector. The access information indicating the arrangement of the access area is the coordinates of the prediction vector finger point, which is the point moved by the prediction vector from the coordinates indicating the position of the encoding target block. When the type of task is motion compensation, the prediction vector in the present embodiment, excluding the description related to calculation of the prediction vector, may be read as the motion vector.

  Below, it demonstrates centering around difference with 1st Embodiment.

  The access region estimation unit 107 acquires the prediction vector, and calculates the coordinates of the prediction vector finger point from the prediction vector and the coordinates representing the position of the encoding target block. The access area estimation unit 107 uses the coordinates of the predicted vector finger point as access information. When the task information includes a prediction vector, the access area estimation unit 107 acquires the prediction vector from the task information. When the prediction vector is stored in a storage unit (not shown), the access area estimation unit 107 acquires the prediction vector from the storage unit. When the task is motion estimation, the access region estimation unit 107 may calculate the prediction vector without acquiring the prediction vector from the task information or the like. The access region estimation unit 107 may acquire a motion vector that has already been calculated around the encoding target block and is necessary for calculating the prediction vector of the encoding target block, and calculate the prediction vector. The motion vector that has already been calculated may be stored in the motion vector storage unit 1005 for storing the motion vector that has already been calculated, which is included in the information processing apparatus 1000 described above, for example. The access area estimation unit 107 may read a necessary motion vector from the motion vector storage unit 1005.

  The task access acquisition unit 102 acquires access information that is the coordinates of the predicted vector finger point.

  The access area storage unit 105 stores task identification information and access information that is the coordinates of the predicted vector finger point in association with each other.

  The task correlation calculation unit 103 calculates a distance between predicted vector finger points of two tasks and sets a value representing a correlation between the two tasks.

  The larger the overlapping amount of the access areas of two tasks, the closer the distance between the predicted vector finger points of those two tasks should be. Therefore, in the present embodiment, the higher the correlation between tasks, the smaller the value representing the correlation between tasks. Thus, the relationship between the high correlation between tasks and the magnitude of the value representing the correlation between tasks does not necessarily have to be such a relationship that the value representing the correlation increases as the correlation between tasks increases. The relationship between the height of the correlation between tasks and the magnitude of the value representing the correlation between tasks may be determined in advance according to the calculation method of the value representing the correlation.

  The schedule unit 104 according to the present embodiment determines the task execution order so that two tasks having a small correlation value and a high correlation value are successively executed.

  Next, the operation of the present embodiment will be described in detail with reference to the drawings.

  FIG. 5 is a flowchart showing an example of the operation of the task scheduling apparatus 1 of this embodiment. Compared with FIG. 2 representing the operation of the first embodiment, the operation of FIG. 5 is an operation in which step S102 of the operation of FIG. 2 is replaced with step S202, step S104 is replaced with step S204, and step S105 is replaced with step S205. is there. The operation of the step assigned with the same reference numeral as in FIG. 2 is the same as the operation of the step assigned with the reference numeral in the first embodiment, and thus the description thereof is omitted.

  FIG. 6 is a diagram illustrating an example of prediction vectors of a plurality of tasks. The end point of each arrow is a prediction vector finger point. Arrows to which A, B, and C are attached represent prediction vectors of task A, task B, and task C, respectively.

  In the following description, an operation example when the task scheduling apparatus 1 determines the execution order of each task whose prediction vector is shown in FIG. 6 will also be described.

  After step S101, when the access region estimation unit 107 calculates the access region arrangement, the access region estimation unit 107 calculates the coordinates of the predicted vector finger point.

  The task access acquisition unit 102 acquires access information that is the coordinates of the predicted vector finger point of the task for which task information has been acquired in step S101 (step S202).

  The task access acquisition unit 102 associates the task specifying information and the access information (in this embodiment, the coordinates of the predicted vector finger point) of the task for which the task information has been acquired in step S101, and stores them in the access area storage unit 105.

  When the storage of the combination of task identification information and access information of the previous task in the access area storage unit 105 is completed (step S103, Yes), the task correlation calculation unit 103 calculates the prediction vector finger point between the two task sets. The distance is calculated (step S204). The task correlation calculation unit 103 sets the calculated predicted vector finger point distance as a correlation between two corresponding tasks (a value indicating the correlation). As described above, in this embodiment, the higher the correlation between tasks, the smaller the value representing the correlation.

  First, the task correlation calculation unit 103 may calculate the predicted vector finger point distances for all the two task sets selected from the tasks for which the task acquisition unit 101 has acquired task information.

  In the case of the task example of FIG. 6, the distance between the predicted vector finger points of task A and task B is 5, the distance between the predicted vector finger points of task B and task C is √34 (about 5.8), The distance between the predicted vector finger points is √5 (about 2.2).

  Next, the scheduling unit 104 determines the execution order of these tasks so that the two tasks having the highest correlation are executed successively (S205). As described above, the smaller the distance between predicted vector finger points, which is a value representing the correlation, the higher the correlation between the two corresponding tasks. Therefore, the schedule unit 104 selects the task set having the smallest distance between the predicted vector finger points. To do.

  In the case of the example of the task in FIG. 6, since the distance between the predicted vector finger points of task A and task C is the smallest, the schedule unit 104 selects task A and task C and assumes that they have been scheduled.

  When the scheduling unit 104 selects all tasks and does not schedule them (step S106, No), the scheduling unit 104 selects a task having the highest correlation with the already scheduled tasks (step S205). The schedule unit 104 makes the selected task scheduled. The scheduling unit 104 continuously selects a task that has a higher correlation with the last selected task among the tasks that are executed first and the task that is executed first among the already selected tasks. Determine the task execution order to be executed. If the correlation between the task executed first among the tasks already selected is higher than the correlation between the task executed last and the task selected last, the scheduling unit 104 Make the execution order first. In the opposite case, the schedule unit 104 puts the last selected task in the execution order.

  In the example of the task in FIG. 6, the predicted vector finger point distance between task B and task C is smaller than the predicted vector finger point distance between task A and task B. Therefore, if the execution order of the task A and task C that have already been selected is the order of task A and task C, the schedule unit 104 sets the execution order of task B before (first) task A, and task A and task B. Is executed continuously.

  The operation of the schedule unit 104 in step S205 is the same as the operation in step S105 of the first embodiment, except that the level of correlation is determined by the small distance between predicted vector finger points.

  In step S205, the schedule unit 104 may select a task having the smallest distance between the predicted vector finger point and the center of gravity of the predicted vector finger point of the already selected task. In this case, if not all tasks have been selected (No at Step S106), the task scheduling apparatus 1 may return to the operation at Step S204.

  In step S204, the task correlation calculation unit 103 calculates the centroid of the predicted vector finger point of the already selected task, and calculates the distance between the calculated centroid and the predicted vector finger point of the task not yet selected. .

  When the schedule unit 104 determines the execution order of all tasks (step S106, Yes), the output unit 106 outputs information indicating the execution order (step S107).

  In addition to the effect of the first embodiment, the present embodiment described above has an effect that the amount of calculation for determining the task execution order can be reduced.

  The reason is that the scheduling unit 104 determines the level of correlation based on the predicted vector finger point distance. Therefore, there is no need to calculate the overlapping amount of access areas. In addition, the correlation calculation does not require information indicating the arrangement of access areas, which are memory areas accessed by tasks. If information indicating the access area arrangement is necessary, even if the task scheduling apparatus 1 does not calculate the access area arrangement, for example, another processing unit such as the task generation unit 1001 of the information processing apparatus 1000 accesses the access area. It is necessary to calculate the arrangement of the areas. Therefore, by configuring the task scheduling apparatus as in the present embodiment, it is possible to simplify the process for determining the task execution order, and the amount of computation is reduced.

(Third embodiment)
Next, a third embodiment of the present invention will be described in detail with reference to the drawings.

  FIG. 7 is a diagram illustrating the configuration of the task scheduling apparatus 3 of the present embodiment.

  Referring to FIG. 7, the task scheduling apparatus 3 of the present embodiment includes a prediction vector accuracy estimation unit 303 and an access area correction unit 304 in addition to the configuration of the task scheduling apparatus 1 of the first embodiment of FIG. .

  In motion estimation, a motion vector search is performed centering on a prediction vector. Although the upper limit of the motion vector search range is determined by the setting, when an adaptive search algorithm is used, the actual search range is not constant. If it is known that the prediction vector is close to the motion vector, that is, the accuracy of the prediction vector is high, the search amount is small and the search range, that is, the memory access range may be narrow. On the other hand, if the accuracy of the prediction vector is known to be low, the memory access range needs to be wide. Therefore, in the present embodiment, the memory access range size is changed according to the accuracy of the prediction vector.

  The prediction vector accuracy can be estimated by comparing the surrounding motion vectors used for calculating the prediction vector. If the direction and size of the motion vectors of the surrounding blocks are similar, it is expected that the motion vectors of the encoding target blocks are also similar. Therefore, in this case, the accuracy of the prediction vector is estimated to be high.

  The prediction vector accuracy estimation unit 303 acquires information on motion vectors already calculated for the blocks around the encoding target block. The prediction vector accuracy estimation unit 303 estimates the accuracy of the prediction vector of the encoding target block from the acquired motion vector information.

  The access area correction unit 304 corrects the arrangement of the access areas stored in the access area storage unit 105 according to the accuracy of the prediction vector estimated by the prediction vector accuracy estimation unit 303. When the access area is an area having a predetermined size centered on the prediction vector finger point, the access area correction unit 304 may correct the size of the area according to the accuracy of the prediction vector.

  The other components of the present embodiment are the same as the components of the first embodiment given the same reference numerals, and thus description thereof is omitted.

  The access area correction unit 304 does not correct the arrangement of the access areas stored in the access area storage unit 105, but the access area estimation unit 107 determines the size of the access area according to the accuracy of the prediction vector. And the location of the access area may be estimated.

  Next, the operation of the task scheduling apparatus 3 of the present invention will be described in detail with reference to the drawings.

  FIG. 8 is a flowchart showing the operation of the task scheduling apparatus 3 of this embodiment. 8 is compared with the flowchart representing the operation of the task scheduling apparatus 1 of the first embodiment of FIG. 1, the difference is that steps S304 to S306 exist between step S102 and step S103.

  Since the operations in steps S101 and S102 are the same as those in the steps having the same reference numerals in the first embodiment, description thereof is omitted.

  The predicted vector accuracy measurement unit 303 receives from the task acquisition unit 101 information indicating task identification information and the position of the encoding target block of the task for which the task acquisition unit 101 has acquired task information in step S101. The prediction vector accuracy measurement unit 303 acquires information on motion vectors of blocks around the encoding target block (step S304).

  The prediction vector accuracy estimation unit 303 may acquire motion vector information from the motion vector storage unit 1005 described above. The motion vector from which the prediction vector accuracy estimation unit 303 obtains information may be a motion vector used for calculation of the prediction vector of the encoding target block. As described above, for example, H.264. In H.264, the motion vectors of the encoding target block are calculated using the motion vectors of the left, top, and upper right blocks of the encoding target block. In this case, the prediction vector accuracy estimation unit 303 may acquire the motion vectors of the left, upper, and upper right blocks of the encoding target block. The prediction vector accuracy estimation unit 303 may acquire motion vectors that have already been calculated for surrounding blocks other than the motion vector used for calculation of the prediction vector.

  The prediction vector accuracy estimation unit 303 estimates the accuracy of the prediction vector of the encoding target block from the acquired motion vector (step S305).

  The prediction vector accuracy estimation unit 303 may perform estimation of the accuracy of the prediction vector based on the magnitude of the obtained motion vector variation. For example, the prediction vector accuracy estimation unit 303 may calculate the variances of the acquired motion vector in the x direction and the y direction, respectively. The smaller the variance in the x and y directions of surrounding vectors, the higher the accuracy of the prediction vector. The predicted vector accuracy estimation unit 303 may further calculate a normalized value by taking, for example, the logarithm of the calculated variance value in the x direction / y direction. The predicted vector accuracy estimation unit 303 may use the calculated value as a coefficient for correcting the task access area (correction coefficient). The prediction vector accuracy estimation unit 303 transmits the task identification information and the two correction coefficients in the x direction and the y direction to the access area correction unit 304.

  The detailed calculation procedure of the correction coefficient is, for example, as follows. The prediction vector accuracy estimation unit 303 first calculates the variance of each motion vector in the x direction and the y direction. Next, the predicted vector accuracy estimation unit 303 calculates the logarithm of the calculated variance in each of the x direction and the y direction. The base of the logarithm may be determined appropriately. Further, if there is a possibility that the variance is less than 1, the prediction vector accuracy estimation unit 303 may calculate the logarithm after adding 1 to the variance value. The prediction vector accuracy estimation unit 303 sets the correction coefficient to 1 if the logarithm of the calculated variance in each of the x direction and y direction is equal to or greater than a predetermined value, and calculates the variance in the calculated x direction and y direction if it is less than the predetermined value. A value obtained by dividing the logarithm of x by a predetermined value may be used as the correction coefficient in the x direction and the y direction.

  In the process of each task, when changing the search range according to the accuracy of the prediction vector of the encoding target block of the task, if the algorithm for the change is known, the prediction vector accuracy estimation unit 303 The correction coefficient may be calculated as follows. That is, the prediction vector accuracy estimation unit 303 calculates the ratio of the search range in the x direction and the ratio of the search range in the y direction before and after the change of the search range, and sets the correction coefficient in the x direction and the correction coefficient in the y direction, respectively. do it.

  The access area correction unit 304 corrects the access information corresponding to the received task identification information stored in the access area storage unit 105 (step S306).

  The access area represented by the access information is an area having a predetermined size centered on the predicted vector finger point when the task access acquisition unit 102 stores the access area in the access area storage unit 105. The access area correction unit 304 may multiply the received x-direction and y-direction correction coefficients by the sizes of the x-direction and y-direction areas, respectively. When the correction coefficient calculated as described above is multiplied by the size of the region in the x direction and the y direction, the smaller the variance of the motion vectors of the surrounding blocks, the narrower the access region. The correction coefficient does not have to be a value calculated as described above as long as the variance of the motion vectors of the surrounding blocks is smaller and the access area is narrower. The predicted vector accuracy correction unit 303 sets an upper limit on the size of the correction coefficient so that the size of the access area does not exceed a predetermined size regardless of the degree of accuracy. Good. For example, as in the above-described example, the prediction vector accuracy correcting unit 303 sets the correction coefficient when the accuracy evaluation value (dispersion or logarithm of dispersion in the above example) exceeds a predetermined range to the worse accuracy. It may be set to 1.

  Since the operations in steps S103 to S107 are the same as those in the steps having the same reference numerals in the first embodiment, description thereof will be omitted.

  Further, the task scheduling apparatus 3 may include an access region estimation unit 107 that calculates the arrangement of access regions, and the prediction vector accuracy estimation unit 303 may transmit a correction coefficient to the access region estimation unit 107. The access area estimation unit 107 may multiply the received predetermined correction coefficient by a predetermined size to obtain the access area size. In this case, the access area correction unit 304 does not correct the access information. Furthermore, the access area correction unit 304 may not exist.

  Next, the effect of this embodiment will be described.

  In the present embodiment, when the tasks are executed in the execution order determined by the task scheduling device 3, in addition to the effects of the first embodiment, the cache hit rate when executing the task is further improved, There is an effect that the execution speed is further improved.

  The reason is that the prediction vector accuracy estimation unit 303 calculates a value representing the accuracy of the prediction vector, and the access region correction unit 304 increases the size of the access region so that the size of the access region decreases as the prediction vector accuracy increases. It is because it corrects.

  In motion estimation, a task does not always access the entire set search range. For example, if a block with a similarity greater than or equal to a predetermined criterion is found before the task completes a search for blocks similar to the encoding target block for the entire set search range, the block May be detected as the finger point of the motion vector, and the search may be terminated. In such a case, the range that the task actually accesses is only the range that is accessed through the search path from the search start point that is the predicted vector finger point to the detected motion vector finger point, in the entire search range. Therefore, when the accuracy of the prediction vector is high and the prediction vector finger point is close to the motion vector finger point, the range that the task accesses becomes narrow. On the other hand, when the accuracy of the prediction vector is low and the prediction vector finger point is far from the motion vector finger point, the range that the task accesses becomes wide.

  The access area correction unit 304 according to the present embodiment is not an access area when the entire search area set in advance is searched, but an area that is actually accessed by the task estimated from the accuracy of the prediction vector. Modify the access area so that Thereby, the task correlation calculation unit 103 can use a range close to the actual memory access range for calculation of the task correlation. Therefore, the accuracy of task correlation, which is an overlapping portion of access areas, calculated by the access correlation calculation unit 103 is improved. Therefore, the cache hit rate when the tasks are executed in the execution order determined by the task scheduling device 3 is further improved, the cache data can be used more efficiently, and the task execution speed is further improved.

(Fourth embodiment)
Next, a fourth embodiment will be described in detail with reference to the drawings.

  FIG. 13 is a diagram showing the configuration of the task scheduling apparatus 4 of this embodiment.

  The task scheduling apparatus 4 according to the present embodiment performs scheduling when tasks are processed in parallel by a multicore processor or a menicore processor. In order to effectively use the cache in a parallel processor having a cache for each core, it is important to continuously execute a task referring to the same memory area on the same processor core. Furthermore, in order to perform parallel processing efficiently, it is necessary to equalize the processing amount per processor core.

  When scheduling is performed using the technique of Patent Document 2, scheduling is performed without considering the processing amount, and thus the processing amount for each processor core is biased. Furthermore, when the technique of Patent Document 3 is assigned to a processor core for each group divided for each memory access area, the processing amount for each group is not equal.

  FIG. 24 is a diagram illustrating a result of grouping tasks with the technique of Patent Document 3 and an example of assignment of task groups to processor cores (PE1, PE2).

  In the example of the task accessing the range shown in FIG. 24, the area (1) includes task A, and the area (2) includes tasks B, C, and D. When each group is assigned to another processor core by the technique of Patent Document 3, the number of tasks is biased as 1 and 3.

  In this embodiment, the processor 1003 in FIG. 29 executes a task with a plurality of processor cores. The task scheduling apparatus 4 according to the present embodiment groups a plurality of tasks executed by the processor 1003 into a plurality of task groups assigned to each processor core. The task scheduling apparatus 4 receives the number of processor cores as the number of task group groups, and performs the operations described below so that the number of tasks to be executed is equalized to the number of processor core task groups. assign. If the processing amount per task is equal, the total processing amount of tasks included in each task group is equal. Then, when executing a task, for example, the task execution unit 1002 in FIG. 29 assigns each group to a processor core.

  Referring to FIG. 13, the task scheduling apparatus 4 includes a task acquisition unit 101, a task access acquisition unit 102, an access area storage unit 105, a task correlation calculation unit 103, a schedule unit 104, an output unit 106, a group An access calculation unit 406 and a group storage unit 407 are included. The task scheduling device 4 may include an access area estimation unit 107.

  In the present embodiment, the task for which the task acquisition unit 101 acquires task information is a task that can be processed in parallel.

  The task correlation calculation unit 103 calculates a correlation between a task group including a plurality of tasks and a task in addition to calculating a correlation between two tasks.

  The schedule unit 104 registers highly correlated tasks in the same task group. The schedule unit 104 also determines the task execution order within the task group.

  The group access calculation unit 406 calculates a memory area accessed by the task group.

  The group storage unit 407 stores information for specifying a task group and information for specifying a task included in the task group in association with each other.

  The other components of the task scheduling device 4 are the same as the components of the task scheduling device 1 of the first embodiment given the same reference numerals, and thus the description thereof is omitted.

  Next, the operation of the task scheduling apparatus 4 of this embodiment will be described in detail with reference to the drawings.

  FIG. 9 is a diagram illustrating an example of a memory range (access area) accessed by a plurality of tasks (task A, task B, task C, and task D) that can be processed in parallel. In the description of the operation of the present embodiment, an example of an operation in which the task scheduling apparatus 4 assigns these tasks to two processor cores and performs scheduling, that is, an operation in which these tasks are divided into two groups and scheduled is also described. To do.

  FIG. 14 is a flowchart showing an example of the operation of the task scheduling device 4.

  For example, the schedule unit 104 of the present embodiment only needs to hold the number of task groups in advance. Alternatively, in step S101, the task acquisition unit 101 may also acquire information about the number of task groups in addition to the task information. In this case, the task acquisition unit 101 may transmit information on the acquired number of task groups to the schedule unit 104.

  The operations of step S101 to step S103 of the task scheduling apparatus 4 are the same as the operations of steps S101 to S103 of the task scheduling apparatus 1 of the first embodiment, except that the number of groups is acquired and transmitted to the scheduling unit 104. The same. Therefore, the description of the operation in steps S101 to S103 is omitted.

  The schedule unit 104 calculates the number of tasks (number of group tasks) per task group from the received number of groups and the number of tasks whose access information is stored in the access information storage unit 105 (step S401). The schedule unit 104 may be obtained by counting the number of tasks stored in the access information storage unit 105. The schedule unit 104 may receive the number of tasks to be scheduled from the task acquisition unit 101.

  In the example of FIG. 9, the number of tasks per group is two.

  Next, the task correlation calculation unit 103 calculates the correlation between the tasks for all the combinations of the two tasks selected from the task that received the task information, similarly to step S104 in FIG. 2 (step S104).

  FIG. 10 is a diagram showing an access area accessed by the task in the example of FIG. 9 as a table.

  FIG. 11 is a diagram showing the correlation between tasks calculated from the information indicating the access area of FIG. The overlap size in FIG. 11 corresponds to the correlation between tasks.

  Next, the scheduling unit 104 selects a task pair having the highest correlation among the tasks that have not been grouped to assign the task to any task group, and registers (groups) the task pair in the new task group i (grouping). Step S405). i is a task group number. The initial value of the number i may be 0, for example. The schedule unit 104 associates information for identifying the task group i with information for identifying each task constituting the selected pair, and stores the information in the group storage unit 407. Note that in the operation of step S405, the scheduling unit 104 first registers one of the task pairs having the highest correlation in the task group, and further registers the task having the highest correlation with the task group in the task group. It can also be regarded as an action to perform.

  In the example of FIG. 9, if there is no grouped task, the task pair with the highest correlation is task C and task D. I is an initial value of 0. The schedule unit 104 registers task C and task D in group 0. The schedule unit 104 associates the information specifying the task group 0 with the information specifying the task C and the information specifying the task D, and stores them in the group storage unit 407.

  When the number of tasks registered in the task group i is smaller than the number of group tasks (step S408, Yes), the group access calculation unit 406 calculates the access area of the task group (task group i) in which the task is newly registered. . The group access calculation unit 406 may set the sum of memory areas accessed by the tasks included in the task group i as the access area of the task group i. The group access calculation unit 406 transmits information indicating the calculated access area of the task group i to the task correlation calculation unit 103.

  The task correlation calculation unit 103 calculates a correlation (group correlation) between each task that is not grouped and the task group i. Group correlation is the correlation between tasks and groups. The schedule unit 104 selects a task having the highest correlation with the task group i from tasks that are not grouped. The schedule unit 104 registers the selected task in the task group i (step S407).

  Similar to the correlation between tasks, the scheduling unit 104 may set the size of the overlapping portion of the access area of the task group i and the access area of the task to the correlation between the task group i and the task.

  The schedule unit 104 associates information for specifying a task group with each task specifying information for each task included in the task group, and stores the information in the group storage unit 407.

  If the number of tasks registered in task group i is equal to or greater than the number of group tasks (step S408, No), if there is a task for which grouping has not ended (step S409, No), i is incremented and step Return to S405. In this case, if the grouping of all tasks has been completed (step S409, Yes), the task scheduling device 4 outputs the execution order for each task group (step S107) and ends the operation.

  When the number of tasks registered in the task group i is equal to or greater than the number of group tasks (step S408, No), the scheduling unit 104 performs steps S104 to S106 in FIG. 2 among the tasks included in the task group i. The execution order may be determined in the same manner as in FIG.

  In the example of the task of FIG. 9, when task C and task D are grouped into task group 0 and task A and task B are not grouped, the number of tasks in task group 0 is equal to the number of group tasks. Furthermore, there are tasks A and B that are not grouped. In this case, the operation returns to step S405, and task A and task B are registered in task group 1.

  As a result, the task scheduling apparatus 4 groups the tasks in FIG. 9 as in the example in FIG. FIG. 12 is a diagram illustrating a result of grouping the tasks of FIG. 9 by the task scheduling apparatus 4.

  In step S405, the schedule unit 104 may determine the execution order so that the pair of tasks is continuously executed when the pair of tasks is registered in the new task group i. Then, in step S407, the schedule unit 104 may determine the execution order of tasks newly registered in the task group i in the same manner as in step S105 in FIG. Furthermore, the schedule unit 104 may store in the group storage unit 407 a set of task specifying information and task specifying information arranged in the order of the determined task execution order (information representing a task sequence).

  The task correlation calculation unit 103 also correlates a task and a task group, a correlation between the task and the first task in the task group, and a task whose execution order is the last in the task group. Of these, the higher one may be used. In this case, the operation in step S406 is not necessary, and the group access calculation unit 406 does not have to exist.

  Moreover, the schedule part 104 may repeat as many times as the number of groups which received step S405 first. In step S407, the scheduling unit 104 repeats an operation of selecting a task and a group having the highest correlation among the ungrouped tasks and groups, and grouping the selected task into the selected group. Also good.

  In step S107, the output unit 106 outputs information representing the task execution order for each group.

  Each task group grouped as described above is assigned to a processor core included in the processor 1003, for example. Each processor core processes each task included in the assigned task group in parallel for each task group.

  FIG. 15 is a diagram illustrating an example of task assignment to a plurality of processor cores. When the processor 1003 includes two processor cores (PE1, PE2), the number of task groups is two. If the number of task groups is 2, the task scheduling apparatus 4 groups the tasks shown in FIG. 9 (left side in FIG. 15) into two task groups as shown on the right side in FIG. For example, the task execution unit 1002 allocates two task groups to the two processor cores (PE1 and PE2) of the processor 1003.

  In addition to the effect of the first embodiment, the present embodiment described above has an effect that the processing amount of each processor core can be equalized when a plurality of processor cores execute tasks. is there.

  The reason is that the scheduling unit 104 sequentially assigns the task having the highest correlation with each task group to the task group so that the number of tasks included in each task group corresponding to the number of processor cores is equal. It is. In this case, if the processing amount of each task is equal, the total processing amount of tasks included in each task group is equal.

  When scheduling is performed as in Patent Documents 2 and 3 and tasks are assigned to processor cores for each group, the number of tasks in the group cannot be considered during grouping. Therefore, in the task example of FIG. 24, the number of tasks per processor is biased. The task scheduling apparatus 4 according to the present embodiment can equalize the number of tasks included in each task group during task grouping. Therefore, the task scheduling apparatus 4 according to the present embodiment can equalize the processing amount for each processor core when the processing amount of each task is equal.

(Fifth embodiment)
Next, a fifth embodiment of the present invention will be described in detail with reference to the drawings.

  If the amount of processing per task varies from task to task, schedule not only the number of tasks per task group but also the amount of processing of tasks, so that the amount of processing per group, and thus per processor core, is equalized. Can be

  For example, the amount of processing per task in motion estimation can be estimated from the number of referable frames. It can be estimated that the processing amount in the case where the number of referenceable frames is two is double the processing amount in the case of one frame. In the present embodiment, even when the processing amount per task is not equal, the load per processor core is equalized. The processing amount is a numerical value indicating the degree of calculation amount. The unit of calculation amount may be arbitrary.

  FIG. 16 is a diagram illustrating the configuration of the task scheduling apparatus 5 of the present embodiment.

  Referring to FIG. 16, the task scheduling apparatus 5 of the present embodiment includes a task processing amount estimation unit 508, a task processing amount storage, in addition to the components of the task scheduling apparatus 4 of the fourth embodiment shown in FIG. Part 509.

  The task processing amount estimation unit 508 estimates the processing amount of the task from the task information acquired by the task acquisition unit 101. For example, the amount of processing per task in motion estimation correlates with the number of frames that can be referred to and the number of search block types described briefly below. That is, the calculation amount of the task of motion estimation is roughly proportional to the number of referable frames and the number of search block types. Therefore, the task processing amount estimation unit 508 may estimate the processing amount per task in motion estimation from these pieces of information.

  The number of search block types is a block size type for searching for a motion vector in one task. For example, the task is H.264. In the case of a task for performing motion search according to the H.264 standard, the task is to search for a motion vector in a smaller block obtained by dividing a macroblock in addition to searching for a motion vector in units of macroblocks (16 × 16 pixels). May do. In this case, the task selects the best motion vector from the results of a search performed with a plurality of blocks of a plurality of sizes. In such a case, for example, when a block for performing a search is a 16 × 16 pixel macroblock and a block of 8 × 8 pixel and 4 × 4 pixel macroblocks, the number of search block types is 3.

  For example, the task processing amount estimation unit 508 may set the assumed maximum number of comparisons performed between the encoding target block and the block in the search frame as the processing amount in motion vector search. . If the algorithm for comparison performed between the encoding target block and the block in the search frame in the motion vector search is the same for each task, the task throughput estimation unit 508, for example, searches for task motion estimation. The value of the total area of the range may be used as the processing amount. The task information acquired by the task acquisition unit 101 only needs to include the number of referable frames, the number of search block types, and other information for estimating the processing amount of the task in addition to the type of task.

  The schedule unit 104 assigns tasks to each task group so that the processing amount of each task group approaches the guideline of the processing amount per task group calculated from the total processing amount of all tasks and the number of groups. The scheduling unit 104 further performs task scheduling for each task group.

  The task processing amount estimation unit 508 stores the task specifying information and the processing amount of the task specified by the task specifying information in the task processing amount storage unit 509 in association with each other.

  The task processing amount storage unit 509 stores task specifying information and processing amount in association with each other.

  The other components of the task scheduling device 5 are the same as the components of the same number of the task scheduling device 4 of the fourth embodiment, and thus description thereof is omitted.

  FIG. 17 is a flowchart showing an example of the operation of the task scheduling apparatus 5 of this embodiment. Comparing FIG. 17 and FIG. 14, the task scheduling apparatus 5 has the point that the operation of step S501 is performed after step S102 and the operation of step S502 is performed before step S104. Different from the device 4. Furthermore, the task scheduling apparatus 5 is different from the task scheduling apparatus 4 in that it performs the operation of step S508 instead of step S408. Since the operations of the other steps of the task scheduling device 5 are the same as the operations of the steps having the same number in the task scheduling device 4 of the fourth embodiment, the description thereof is omitted.

  After step S102, the task processing amount estimation unit 508 estimates the processing amount of the corresponding task from the task information acquired by the task acquisition unit 101 (step S501). The task processing amount estimation unit 508 stores the estimated processing amount in the task processing amount storage unit 509 in association with the task specifying information of the corresponding task.

  When the task acquisition unit 101 completes the acquisition of the access areas of all the tasks for which task information has been acquired and the estimation of the task processing amount (step S103, Yes), the task processing amount estimation unit 508 determines the processing amount per task group. Is calculated (step S502).

  The task processing amount estimation unit 508 may receive the number of task groups from the task acquisition unit 101 in addition to the task information. The task processing amount estimation unit 508 calculates the total processing amount of each estimated task, and calculates the processing amount guide (processing amount guide) per task group by dividing the calculated total by the number of received groups. do it.

  In step S509, the schedule unit 104 compares the total processing amount of tasks assigned to the task group i with the processing amount standard estimated by the task processing amount estimation unit 508. When the total processing amount of tasks assigned to task group i is less than the processing amount guide (step S509, Yes), the schedule unit 104 continues to assign tasks to task group i (steps S406 to S407). To do. When the total processing amount of tasks assigned to the task group i is equal to or larger than the processing amount guide (No in step S509), the scheduling unit 104 ends the assignment of tasks to the task group i.

  In the present embodiment described above, in addition to the effects of the fourth embodiment, even when the processing amount for each task is different, the processing amount per processor core can be equalized, and the processing of all tasks There is an effect that it is possible to shorten the processing time until the process is completed.

  The reason is that the task processing amount estimation unit 508 estimates the processing amount for each task. Using the estimated processing amount for each task, the scheduling unit 104 assigns tasks to each task group so that the total processing amount for each task group, which is a set of tasks to be executed by individual processor cores, becomes equal. . If the processing amount of the task executed by each processor core is equal, the processing time until the processing of all tasks is completed is shorter than when the processing amount of the task executed by each processor core is not equal.

(Sixth embodiment)
Next, a sixth embodiment of the present invention will be described in detail with reference to the drawings.

  The amount of processing per task in motion estimation is determined by the amount of motion vector search. The search amount corresponds to a range in which the search is performed from when the search is started until the motion vector finger point is detected. The search amount per task is roughly determined by the number of frames that can be referred to by the task. However, the search may be aborted when it is determined that a sufficiently good motion vector has been found, so the search amount per task is not always proportional to the number of referable frames. For example, the search amount per frame / block differs depending on the distance between the search start point and the motion vector finger point. The smaller the distance between the search start point and the motion vector finger point, the greater the distance.

  Since the motion vector has a high correlation with the motion vectors of the surrounding blocks, a good motion vector can be determined early by performing a motion vector search centering on the prediction vector calculated from the surrounding motion vectors. If the direction and size of the prediction vector and the motion vectors of the surrounding blocks are similar, the prediction vector is close to the motion vector and the search amount is expected to be small. On the other hand, if the directions and sizes of the motion vectors of the surrounding blocks are various, it is considered that the accuracy of the prediction vector is insufficient and the search amount increases. Therefore, the task scheduling apparatus 6 of this embodiment estimates the processing amount of each task using the accuracy of the prediction vector calculated as in the third embodiment. This increases the accuracy of estimating the processing amount of each task. Therefore, the task scheduling apparatus 6 of this embodiment can group tasks so that the processing amount per task group is equal.

  FIG. 18 is a diagram illustrating the configuration of the task scheduling apparatus 6 according to the present embodiment.

  Referring to FIG. 18, the task scheduling device 6 includes a prediction vector accuracy estimation unit 303 and a task processing amount correction unit 614 in addition to the configuration of the task scheduling device 5 in FIG. 16.

  The prediction vector accuracy estimation unit 303 is the same as the prediction vector accuracy estimation unit 303 of the third embodiment.

  The task processing amount correction unit 614 estimates the processing amount of each task estimated by the task processing amount estimation unit 508 and stored in the task processing amount storage unit 509 according to the accuracy of the prediction vector calculated by the prediction vector accuracy estimation unit 303. To correct.

  The other components of the task scheduling device 6 are the same as the components with the same number of the task scheduling device 5 of the fifth embodiment, and thus description thereof is omitted.

  Next, the operation of the task scheduling apparatus 6 of this embodiment will be described in detail with reference to the drawings.

  FIG. 19 is a flowchart showing an example of the operation of the task scheduling apparatus 6. 19 and FIG. 17 representing the operation of the fifth embodiment are the same except that steps S304, S305, and S606 are provided between steps S501 and S103. Hereinafter, differences from the operation of the fifth embodiment in FIG. 17 will be mainly described.

  After the task processing amount estimation unit 508 estimates the processing amount of the task from the task information acquired by the task acquisition unit 101 (step S501), the prediction vector accuracy estimation unit 303 moves the blocks around the encoding target block. A vector is acquired (step S304).

  The prediction vector accuracy estimation unit 303 calculates the accuracy of the prediction vector of the encoding target block from the acquired motion vector (step S305).

  Since step S304 and step S305 are the same as step S304 and step S305 of the third embodiment in FIG. 8, respectively, detailed description thereof is omitted.

  The task processing amount correction unit 614 uses the correction coefficient that represents the accuracy estimated by the prediction vector accuracy estimation unit 303 to estimate the task processing amount estimated by the task processing amount estimation unit 508 and stored in the task processing amount storage unit 509. The amount is corrected (step S606).

  The task processing amount correction unit 614, for example, with respect to the processing amount of the task processing amount storage unit 509, the correction coefficient in the x direction and the y direction corresponding to the prediction vector of the encoding target block targeted by the task. What is necessary is just to apply a correction coefficient.

  Since the operation after step S103 is the same as that of the fifth embodiment, the description thereof is omitted.

  Note that the predicted vector accuracy estimation unit 303 may transmit the calculated correction coefficient to the task processing amount estimation unit 508. Then, the task processing amount estimation unit 508 may store the value obtained by multiplying the calculated processing amount by the received correction coefficient in the task processing amount storage unit 509 in association with the task specifying information. In this case, the task processing amount correction unit 614 may not exist.

  In the present embodiment described above, in addition to the effects of the fifth embodiment, when a task is executed by a plurality of processor cores, the processing amount per processor core can be further equalized and the processing speed can be further increased. There is an effect that parallel processing becomes possible.

  The reason is that the task processing amount correction unit 614 corrects the task processing amount estimated by the task processing amount estimation unit 508 using the accuracy of the prediction vector corresponding to the task estimated by the prediction vector accuracy estimation unit 303. Because it does. Therefore, the estimation accuracy of the processing amount of each task is improved. The scheduling unit 104 uses the corrected processing amount for each task, and assigns tasks to each task group so that the total processing amount for each task group, which is a set of tasks to be executed by individual processor cores, becomes equal. assign. For this reason, the processing amount of each task group is further equalized. Accordingly, the time until all tasks are completed is further shortened. That is, higher-speed parallel processing becomes possible.

(Seventh embodiment)
Next, a seventh embodiment of the present invention will be described in detail with reference to the drawings.

  Each of the embodiments described so far is a task scheduling apparatus that calculates all the correlations between a task group that has already been scheduled and other tasks and selects a task to be assigned to the task group. The first to third embodiments correspond to the case where the task group is 1. On the other hand, in this embodiment, tasks are assigned to task groups by selecting tasks one by one and selecting a task group to which the selected task is assigned according to the correlation between the selected task and each task group. A method for performing scheduling will be described.

  FIG. 27 is a block diagram showing the configuration of the task scheduling apparatus 7 of this embodiment.

  27 and FIG. 13 is different in that the task scheduling apparatus 7 includes a group access storage unit 704 in addition to the configuration of the task scheduling apparatus 4 of the fourth embodiment. In the following, the difference between the task scheduling device 7 and the task scheduling device 4 will be mainly described.

  The group access calculation unit 406 calculates the access area of the updated task group, and stores it in the group access storage unit 704 in association with information for specifying the task group.

  The group access storage unit 704 stores information indicating the access area for each task group calculated by the group access calculation unit 406 in association with information for specifying the task group.

  The task correlation calculation unit 103 calculates a correlation from the task access information read from the access area storage unit 105 and the task group access information read from the group access storage unit 704.

  Next, the operation of the present embodiment will be described in detail with reference to the drawings.

  FIG. 28 is a flowchart showing the operation of the task scheduling apparatus 7 of this embodiment.

  Referring to FIG. 28, first, the task scheduling device 7 acquires task information of tasks of the number of groups, and registers each task for which task information has been acquired in different task groups (step S701).

  Specifically, first, the task acquisition unit 101 receives the number of groups that is the number of task groups. The task acquisition unit 101 may receive the number of tasks to be scheduled in addition to the number of groups. The task acquisition unit 101 acquires task information for the number of received groups. The task access acquisition unit 102 acquires the access information of the task represented by the task information acquired by the task acquisition unit 101 and stores it in the access area storage unit 105 in association with the task identification information. The schedule unit 104 registers each acquired task in a different task group. The schedule unit 104 associates information specifying each task with information specifying different task groups, and stores the information in the group storage unit 407. At this time, each task group includes only one task, and the access area of each task group is equal to the access area of the task included in each task. The group access calculation unit 406 reads the access information of the tasks included in each task group from the access area storage unit 105 and stores the access information in the group access storage unit 704 in association with the information specifying the task group.

  Upon receiving task information in step S701, for example, the task acquisition unit 101 divides the encoding target frame into regions of the number of groups, selects one encoding target block from each region, and selects the selected encoding target block. It is sufficient to acquire one task at a time for processing. When the task acquisition unit 101 acquires the task in this way, the task processing position (position of the encoding target block) for each task group is dispersed within the encoding target frame. Alternatively, the task acquisition unit 101 may select and acquire a task using each frame as a reference frame for each frame. When the task acquisition unit 101 acquires a task in this way, the access area for each task group is distributed.

  Next, the task acquisition unit 101 acquires task information of a task that has not yet been scheduled (step S101). The task acquisition unit 101 may select an arbitrary task from tasks that have not yet been scheduled, and acquire task information of the selected task.

  The access area estimation unit 107 may estimate the access area from the task information acquired from the task acquisition unit 101.

  Next, the task access acquisition unit 102 acquires the access information of the task whose task information is selected from, for example, the access region estimation unit 107 that estimates an access region (step S102). The task access acquisition unit 102 stores the acquired access information in the access area storage unit 105 in association with the task identification information of the corresponding task.

  Next, the task correlation calculation unit 103 reads the access information of the task for which the task acquisition unit 101 has acquired the task information from the access area storage unit 105. The task correlation calculation unit 103 calculates the correlation between the task and the task group from which the task acquisition unit 101 has acquired the task information from the read access information and the access information of each task group stored in the group access storage unit 704. , Respectively (step S704).

  The task correlation calculation unit 103 may calculate the correlation between the task and the task group in the same manner as the calculation of the correlation between tasks. For example, when the access information is information indicating the range of the access area as in the first embodiment, the task correlation calculation unit 103 calculates the size of the overlapping area between the task access area and the task group access area. do it. As in the fourth embodiment, the task group access area may be, for example, the sum of the task access areas included in the task group. For example, as in the second embodiment, when the access information is the coordinates of the predicted vector finger point, the task correlation calculation unit 103 calculates the distance between the predicted vector finger point of the task and the predicted vector finger point of the task group, What is necessary is just to make it a correlation. In this case, the coordinates of the predicted vector finger point of the task group may be, for example, the center of gravity of the predicted vector finger point of the task included in the task group.

  Next, the schedule unit 104 registers the task for which the task information has been acquired in a task group that has a high correlation with the task and that includes a small number of tasks (step S705).

  For example, the scheduling unit 104 has the highest evaluation value calculated with an evaluation formula using the number of tasks and the correlation as a variable such that the evaluation value is higher as the number of tasks included is smaller and the evaluation value is higher as the correlation is higher. Select a high task group and register the task. The evaluation formula may be an evaluation formula that becomes an evaluation value at which no task is registered in the task group when the number of tasks included exceeds the number of tasks per task group. For example, the scheduling unit 104 selects and selects a task group having the highest correlation with the task for which task information is acquired from among task groups in which the number of tasks included is less than the number of tasks per task group The task may be registered in the task group. The operation of the schedule unit 104 in this case outputs an evaluation value proportional to the correlation when the number of tasks is less than the number of tasks per task group, and is 0 as the evaluation value when the number of tasks is equal to or greater than the number of tasks per task group. This corresponds to the above-described operation based on the evaluation expression that outputs The scheduling unit 104 may calculate the number of tasks per task group by dividing the number of tasks to be scheduled acquired by the task acquisition unit 101 by the number of groups.

  Next, the group access calculation unit 406 calculates an access area of the updated task group, and stores it in the group access storage unit 704 in association with information specifying the task group (step S706). The group access calculation unit 406 may receive information indicating the updated task group from the schedule unit 104. The group access calculation unit 406 may detect an update of information representing a task included in each task group stored in the group storage unit 407.

  If scheduling of all tasks has not been completed (No at Step S106), Steps S101 to S706 are repeated. The group access calculation unit 406 performs the operation of updating the access information of the task group (step S706) when the scheduling of all tasks has not been completed after the completion of step S705 (No in step S106). Also good. In this case, the group access calculation unit 406 does not have to perform the operation in step S706 immediately after step S705.

  When scheduling of all tasks has been completed (step S106, Yes), the output unit 106 outputs information indicating the task execution order for each task group (step S107).

  Before outputting information indicating the execution order of tasks, the scheduling unit 104 determines the execution order of each task included in the task group in the same manner as in any of the first, second, and third embodiments, for example. It may be determined for each group. In this case, the output unit 106 may output information indicating the execution order for each task group determined by the schedule unit 104. The execution order for each task group may be, for example, the order in which tasks are registered in each task group. In this case, the schedule unit 104 may store the task specifying information for each task group in the group storage unit 407 so that the task specifying information of the tasks included in each task group is in the order registered in the task group.

  In addition, the group storage unit 407 may store the execution order of tasks assigned to each task group in association with information specifying the task group. In step S706, the group access calculation unit 406 may set a set of access information of the first task and access information of the last task in the task group as the access information of the task group. Then, the group access calculation unit 406 stores the combination of the access information of the first task and the access information of the last task in the task group in the group access storage unit 704 in association with the information specifying the task group. do it.

  In this case, in step S704, the task correlation calculation unit 103 determines whether the task correlation calculation target task has a task correlation with the task whose execution order is the first in the task group and the execution order in the task group. It is only necessary to calculate both of the task correlation with the last task. The task correlation calculation unit 103 may set the larger of the two calculated task correlations as the correlation between the task for which the task correlation is calculated and the task group.

  In step S705, the schedule unit 104 may select a task group to which a task is assigned in the same manner as described above. The scheduling unit 104 determines the execution order so that the task with the larger task correlation between the first task and the last task in the execution order of the selected task group and the task to be assigned is continuous. That's fine. The schedule unit 104 may store the task identification information of tasks included in each task group in the group storage unit 407 in the order of the determined execution order.

  When the group storage unit 407 stores a set of task identification information in the order of execution associated with information identifying each task group, in step S106, the output unit 106 The stored information may be output as it is.

  Further, the task scheduling device 7 may include a task processing amount estimation unit 508 and a task processing amount storage unit 509, similarly to the task scheduling device 5 of the fifth embodiment. In that case, the schedule unit 104 may use the total value of the processing amounts calculated by the task processing amount estimation unit 508 instead of the number of tasks.

  Further, the task scheduling device 7, similarly to the task scheduling device 6 of the sixth embodiment, in addition to the task processing amount estimation unit 508 and the task processing amount storage unit 509, the predicted vector accuracy estimation unit 303, the task processing amount A correction unit 614 may be included. In this case, the task processing amount correction unit 614 may correct the task processing amount in the same manner as in the sixth embodiment. Furthermore, the schedule unit 104 may use the total value of the processing amounts calculated by the task processing amount estimation unit 508 and corrected by the task processing amount correction unit 614 instead of the number of tasks.

  In addition to the effect of the fourth embodiment, the present embodiment described above has an effect that the amount of processing in task scheduling can be reduced.

  This is because the task correlation calculation unit 103 does not necessarily calculate the task correlation for all combinations of two tasks selected from the tasks to be scheduled.

  Next, modified examples of the first to seventh embodiments will be described.

  When the processor 1003 is a parallel processor that processes a task in parallel with a plurality of processor cores, the configuration of the processor 1003 is a configuration in which a core group is formed by one cache shared with a plurality of processor cores, and there are a plurality of core groups. But it ’s okay. In this case, the task execution unit 1002 may assign a task group for each core group sharing the cache. In this case, each processor core can process tasks in parallel even in the core group.

  The configuration of the processor 1003 may be a multiprocessor configuration in which the processors do not share memory. In this case, since the main memory is not shared between different processors, data localization is more important.

  The processor 1003 may have a scratch pad memory instead of the cache memory. Like the cache memory, the scratch pad memory is a small-capacity and high-speed memory, and the user can explicitly manage the data to be held, so that the data locality according to the present invention can be used more effectively.

  In each of the embodiments described above, the task indicates a process in which processes to be performed on one macro block are grouped regardless of whether different frames are referred to. However, for example, as in Patent Document 2 and Patent Document 3, even if processing is performed on the same macroblock, different tasks may be used as long as the reference frames are different.

  In general, the cache memory stores not only actually used data but also data existing at an address close to the used data. Therefore, even if the area where the access areas of the two tasks overlap is the same, the cache hit rate may differ depending on the arrangement of the access areas on the memory. For example, if the data of each pixel of each frame is stored in the memory so that the addresses are continuous in pixels adjacent in the horizontal direction, the data stored in the cache after the task is executed is the access area of the task. And the data of the area adjacent to the access area in the horizontal direction. Therefore, even if the area of the overlapping area of the access areas of the two tasks is the same, for example, the difference in the vertical direction is larger than the case where the vertical difference in the coordinates of the predicted vector finger points of the two tasks is larger than the horizontal difference. It is considered that the cache hit rate is higher when is smaller than the horizontal difference. Therefore, the task correlation calculation unit 103 can also calculate the correlation between tasks by adding the direction to the overlapping amount of access areas and the distance between predicted vector finger points. For example, the task correlation calculation unit 103 uses the task correlation of the two areas as the overlap amount of the areas obtained by adding the areas of a predetermined width that are adjacent to the access areas in the horizontal direction of the two tasks and adding the areas to the access areas. That's fine. Alternatively, the task correlation calculation unit 103 calculates the distance between the prediction vector finger points calculated by multiplying the horizontal component of the coordinates of the prediction vector finger points of the two tasks by, for example, a predetermined weight having a value of less than 1. What is necessary is just to make the task correlation of two tasks.

  The prediction vector accuracy estimation unit 303 according to the third embodiment or the fifth embodiment is not vector dispersion, such as a difference between an average value of prediction vectors of surrounding blocks and a prediction vector of an encoding target block. A value representing the accuracy of the prediction vector may be calculated from another value. Further, in the third embodiment and the fifth embodiment, the prediction vector accuracy estimation unit 303 calculates the prediction vector of the spatially neighboring blocks, that is, the prediction vectors of the blocks around the coding target block in the coding target frame. Then, a value representing the accuracy of the prediction vector of the encoding target block is calculated. However, the prediction vector accuracy estimation unit 303 uses the prediction vector of the temporally close block in addition to the prediction vector of the spatially close block to calculate a value indicating the accuracy of the prediction vector of the encoding target block. May be. A block that is close in time is a block at the same position as the encoding target block in a frame adjacent to the encoding target frame. The prediction vector accuracy estimation unit 303 may further use a prediction vector of a block existing within a predetermined range from the same position as the encoding target block in an adjacent frame of the encoding target frame.

  The task scheduling device 1, the task scheduling device 3, the task scheduling device 4, the task scheduling device 5, the task scheduling device 6, the task scheduling device 7, and the information processing device 1000 are respectively a computer, a program for controlling the computer, and dedicated hardware. Or a combination of a computer and a program for controlling the computer and dedicated hardware.

  Task acquisition unit 101, task access acquisition unit 102, task correlation calculation unit 103, schedule unit 104, output unit 106, access region estimation unit 107, prediction vector accuracy estimation unit 303, access region correction unit 304, group access calculation unit 406, The task processing amount estimation unit 508 and the task processing amount correction unit 614 include, for example, a dedicated program for realizing the function of each unit, which is read into a memory from a recording medium that stores the program, and a processor that executes the program. Can be realized. The access area storage unit 105, the group storage unit 407, the task processing amount storage unit 509, and the group access storage unit 704 can be realized by a memory or a hard disk device included in a computer. Alternatively, task acquisition unit 101, task access acquisition unit 102, task correlation calculation unit 103, schedule unit 104, access region storage unit 105, output unit 106, access region estimation unit 107, prediction vector accuracy estimation unit 303, access region correction unit 304, a group access calculation unit 406, a group storage unit 407, a task processing amount estimation unit 508, a task processing amount storage unit 509, a task processing amount correction unit 614, a part or all of the group access storage unit 704, and a function of each unit. It can also be realized by a dedicated circuit to be realized.

  The task generation unit 1001, the task execution unit 1002, and the execution result output unit 1004 execute, for example, a dedicated program for realizing the function of each unit read from the recording medium storing the program into the memory. It can be realized by a processor. The motion vector storage unit 1005 can be realized by a memory or a hard disk device included in the computer. Alternatively, part or all of the task generation unit 1001, the task execution unit 1002, the processor 1003, the execution result output unit 1004, and the motion vector storage unit 1005 can be realized by a dedicated circuit that implements the functions of the respective units.

  Moreover, although a part or all of said embodiment can be described also as the following additional remarks, it is not restricted to the following.

(Appendix 1)
Task acquisition means for acquiring a plurality of task information including task identification information for identifying a task that is a unit of processing;
Task access acquisition means for acquiring access information representing an arrangement of a memory area accessed by the task, and storing the access information in an access area storage means in association with the task identification information;
Access area storage means for storing a set of the task identification information and the access information;
Task correlation calculating means for calculating a task correlation representing an overlap amount of memory areas accessed by the two tasks from the access information of the two tasks;
Among the combinations of the first or last task in the task sequence, which is the sequence of execution of the tasks, and any of the tasks not included in the task sequence, the tasks of the combination having the highest task correlation are consecutive. Scheduling means for adding the task not included in the task sequence to the task sequence,
An output means for outputting information representing the task sequence.

(Appendix 2)
The task acquisition means for acquiring the task information further including task content information representing the content of processing of the task;
From the task information and the access information, including a task processing amount estimation means for estimating a processing amount representing the degree of calculation amount of the task represented by the task information,
The task correlation calculating means further includes: one or more of the tasks, wherein the total processing amount of each of the tasks included in each of the task groups configured by the task sequence is less than a predetermined value. Each of the task groups includes a group correlation that represents a size of an overlap between a memory area accessed by each task and a memory area accessed by the task included in the target task group that is the task group. Not calculated for each of the tasks
The scheduling means adds the task having the highest group correlation with the target task group among the tasks not included in any of the task groups to the task column of the target task group. ,
The task scheduling apparatus according to claim 1, wherein the output unit outputs information representing the task sequence for each task group.

(Appendix 3)
The task acquisition means for acquiring the task information further including task content information representing the content of processing of the task;
From the task information, including task processing amount estimation means for estimating a processing amount representing the degree of calculation amount of the task represented by the task information,
The task correlation calculation means is further included in any one of the plurality of task groups and a memory area accessed by each task included in each task group configured by the task sequence of one or more of the tasks. Calculating a group correlation for each of the plurality of task groups, which is selected from each of the non-tasks and represents a size of an overlap of memory areas accessed by the task;
The scheduling means sequentially selects a target task from the tasks not included in any of the plurality of task groups, and the group correlation with the selected target task for each of the plurality of task groups And calculating an evaluation value from the processing amount of the target task, adding the target task to the task column of the task group having the highest evaluation value among a plurality of task groups,
The task scheduling apparatus according to claim 1, wherein the output unit outputs information representing the task sequence for each task group.

(Appendix 4)
A target frame in which the task content information is a frame including task type information indicating the type of the task and a target block that is a target area for processing the task in a moving image including a plurality of time-series frames. Information for specifying the position of the target block, and information for specifying a reference frame in the moving image that is referred to in the processing for the target block of the task. The task acquisition means for acquiring the task information;
The task type information is obtained by estimating the position of the region corresponding to the target block in the reference frame, and the position in the reference frame corresponding to the position of the target block in the target frame. In the case of representing motion estimation, which is a process including calculation of a motion vector of the target block, which represents a distance and direction to the estimated position in a reference frame, a plurality of surroundings of the target block in the target frame Each of the motion vectors already calculated for the block of the current block, generate a prediction vector that is a predicted value of the motion vector of the target block from each of the acquired motion vectors, and generate the predicted vector and the task An area in the reference frame accessed by the task specified by the task specifying information from information And generating the access information representing an arrangement of the memory area storing the estimated area, and when the task type information represents motion compensation, obtaining the motion vector of the target block, The task scheduling apparatus according to any one of supplementary notes 1 to 3, further comprising: task access estimation means for generating the access information from a motion vector and the task information.

(Appendix 5)
The access area storage means for storing a set of the task identification information and access information including a size of the memory area accessed by the task identified by the task identification information;
When the task type information represents motion estimation, the motion vector of each of a plurality of surrounding blocks of the target block is acquired, and the accuracy of the prediction vector is estimated from each of the acquired motion vectors Means,
Using the estimated accuracy, the size of the memory region is such that the higher the accuracy, the smaller the size of the memory region accessed by the task that processes the corresponding target block. The task scheduling apparatus according to claim 4, further comprising: an access area correcting means for correcting

(Appendix 6)
The task scheduling apparatus according to appendix 4 or 5, further comprising: a task processing amount correcting unit that corrects the processing amount so that the processing amount decreases as the accuracy increases using the estimated accuracy.

(Appendix 7)
When the task type information represents motion estimation, the task access acquisition unit acquires the prediction vector, and as the access information, coordinates of a prediction vector finger point that is a point moved by the prediction vector from the position of the target block When the task type information represents motion compensation, the motion vector finger point is a point obtained by obtaining the motion vector of the target block and moving as the access information by the motion vector from the position of the target block. The coordinates of
The task correlation calculating means calculates a distance between the coordinates, which is the access information associated with the task specifying information of two tasks, and calculates the task correlation between the two tasks. Scheduling device according to any of (8)
A task scheduling apparatus according to any one of appendices 1 to 7,
A processor;
Task generation means for generating task information of a task to be executed by the processor and inputting the task information to the task scheduling device;
A task scheduling system comprising: task execution means for acquiring information representing the task sequence from the task scheduling device and causing the processor to execute the task according to the order of the acquired task sequence.

(Appendix 9)
Get multiple pieces of task information including task identification information that identifies the task that is a unit of processing,
Obtaining access information representing the location of the memory area accessed by the task, storing it in the access area storage means in association with the task identification information;
Storing a set of the task identification information and the access information in an access area storage means;
From the access information of the two tasks, calculate a task correlation that represents the overlap amount of the memory areas accessed by the two tasks,
Among the combinations of the first or last task in the task sequence, which is the sequence of execution of the tasks, and any of the tasks not included in the task sequence, the tasks of the combination having the highest task correlation are consecutive. So that the task not included in the task column is added to the task column,
A task scheduling method for outputting information representing the task sequence.

(Appendix 10)
Obtaining the task information further including task content information representing the processing content of the task;
From the task information and the access information, a processing amount representing the degree of calculation amount of the task represented by the task information is estimated,
Further, the task group is one task group in which the total processing amount of each of the included tasks is less than a predetermined value in each of the task groups configured by the task sequence of one or more of the tasks. A group correlation indicating a size of an overlap between a memory area accessed by each task and a memory area accessed by the task included in the target task group is set for each of the tasks not included in any of the task groups. Calculated,
In each of the tasks not included in any of the task groups, the task having the highest group correlation with the target task group is added to the task column of the target task group,
The task scheduling method according to claim 9, wherein information representing the task sequence for each task group is output.

(Appendix 11)
Obtaining the task information further including task content information representing the processing content of the task;
From the task information, a processing amount representing the degree of calculation amount of the task represented by the task information is estimated,
Further, one or more of the tasks are selected from a memory area accessed by each task included in the task group configured by the task column and each of the tasks not included in any of the plurality of task groups. In addition, a group correlation representing the size of the overlap of the memory areas accessed by the task is calculated for each of the plurality of task groups,
A target task is sequentially selected from the tasks not included in any of the plurality of task groups, and for each of the plurality of task groups, the group correlation with the selected target task, and the target task The evaluation value is calculated from the processing amount of the task, and the target task is added to the task column of the task group having the highest evaluation value among the plurality of task groups,
The task scheduling method according to claim 9, wherein information representing the task sequence for each task group is output.

(Appendix 12)
A target frame in which the task content information is a frame including task type information indicating the type of the task and a target block that is a target area for processing the task in a moving image including a plurality of time-series frames. Information for specifying the position of the target block, and information for specifying a reference frame in the moving image that is referred to in the processing for the target block of the task. , Get the task information,
The task type information is obtained by estimating the position of the region corresponding to the target block in the reference frame, and the position in the reference frame corresponding to the position of the target block in the target frame. In the case of representing motion estimation, which is a process including calculation of a motion vector of the target block, which represents a distance and direction to the estimated position in a reference frame, a plurality of surroundings of the target block in the target frame Each of the motion vectors already calculated for the block of the current block, generate a prediction vector that is a predicted value of the motion vector of the target block from each of the acquired motion vectors, and generate the predicted vector and the task An area in the reference frame accessed by the task specified by the task specifying information from information And generating the access information representing an arrangement of the memory area storing the estimated area, and when the task type information represents motion compensation, obtaining the motion vector of the target block, The task scheduling method according to any one of appendices 9 to 11, wherein the access information is generated from a motion vector and the task information.

(Appendix 13)
Storing a set of the task identification information and access information including the size of the memory area accessed by the task identified by the task identification information in the access area storage means;
When the task type information represents motion estimation, obtain each of the motion vectors of a plurality of surrounding blocks of the target block, estimate the accuracy of the prediction vector from each of the obtained motion vectors,
Using the estimated accuracy, the size of the memory region is such that the higher the accuracy, the smaller the size of the memory region accessed by the task that processes the corresponding target block. The task scheduling method according to attachment 12, wherein:

(Appendix 14)
The task scheduling method according to appendix 12 or 13, wherein the processing amount is corrected using the estimated accuracy so that the processing amount decreases as the accuracy increases.

(Appendix 15)
When the task type information represents motion estimation, the prediction vector is acquired, and as the access information, coordinates of a prediction vector finger point that is a point moved by the prediction vector from the position of the target block are calculated, and the task type When the information represents motion compensation, the motion vector of the target block is acquired, and as the access information, a coordinate of a motion vector finger point that is a point moved by the motion vector from the position of the target block is calculated,
The task according to any one of appendices 9 to 12, wherein a distance between the coordinates as the access information associated with the task specifying information of two tasks is calculated to obtain the task correlation between the two tasks. Scheduling method (Appendix 16)
Computer
Task acquisition means for acquiring a plurality of task information including task identification information for identifying a task that is a unit of processing;
Task access acquisition means for acquiring access information representing an arrangement of a memory area accessed by the task, and storing the access information in an access area storage means in association with the task identification information;
Access area storage means for storing a set of the task identification information and the access information;
Task correlation calculating means for calculating a task correlation representing an overlap amount of memory areas accessed by the two tasks from the access information of the two tasks;
Among the combinations of the first or last task in the task sequence, which is the sequence of execution of the tasks, and any of the tasks not included in the task sequence, the tasks of the combination having the highest task correlation are consecutive. Scheduling means for adding the task not included in the task sequence to the task sequence,
A task scheduling program that operates as output means for outputting information representing the task sequence.

(Appendix 17)
Computer
The task acquisition means for acquiring the task information further including task content information representing the content of processing of the task;
From the task information and the access information, task processing amount estimation means for estimating a processing amount representing the degree of calculation amount of the task represented by the task information;
Further, the task group is one task group in which the total processing amount of each of the included tasks is less than a predetermined value in each of the task groups configured by the task sequence of one or more of the tasks. A group correlation indicating a size of an overlap between a memory area accessed by each task and a memory area accessed by the task included in the target task group is set for each of the tasks not included in any of the task groups. The task correlation calculating means for calculating
Among the tasks not included in any of the task groups, the scheduling means for adding the task having the highest group correlation with the target task group to the task column of the target task group;
Item 17. The task scheduling program according to appendix 16, which is operated as the output unit that outputs information representing the task sequence for each task group.

(Appendix 18)
Computer
The task acquisition means for acquiring the task information further including task content information representing the content of processing of the task;
From the task information, task processing amount estimation means for estimating a processing amount representing the degree of calculation amount of the task represented by the task information;
Further, one or more of the tasks are selected from a memory area accessed by each task included in the task group configured by the task column and each of the tasks not included in any of the plurality of task groups. Further, the task correlation calculating means for calculating a group correlation representing the size of the overlap of memory areas accessed by the task for each of the plurality of task groups,
A target task is sequentially selected from the tasks not included in any of the plurality of task groups, and for each of the plurality of task groups, the group correlation with the selected target task, and the target task The scheduling means for calculating an evaluation value from the processing amount of and adding the target task to the task column of the task group having the highest evaluation value among a plurality of task groups;
Item 17. The task scheduling program according to appendix 16, which is operated as the output unit that outputs information representing the task sequence for each task group.

(Appendix 19)
Computer
A target frame in which the task content information is a frame including task type information indicating the type of the task and a target block that is a target area for processing the task in a moving image including a plurality of time-series frames. Information for specifying the position of the target block, and information for specifying a reference frame in the moving image that is referred to in the processing for the target block of the task. The task acquisition means for acquiring the task information;
The task type information is obtained by estimating the position of the region corresponding to the target block in the reference frame, and the position in the reference frame corresponding to the position of the target block in the target frame. In the case of representing motion estimation, which is a process including calculation of a motion vector of the target block, which represents a distance and direction to the estimated position in a reference frame, a plurality of surroundings of the target block in the target frame Each of the motion vectors already calculated for the block of the current block, generate a prediction vector that is a predicted value of the motion vector of the target block from each of the acquired motion vectors, and generate the predicted vector and the task An area in the reference frame accessed by the task specified by the task specifying information from information And generating the access information representing an arrangement of the memory area storing the estimated area, and when the task type information represents motion compensation, obtaining the motion vector of the target block, The task scheduling program according to any one of supplementary notes 16 to 18, which is operated as a task access estimation unit that generates the access information from a motion vector and the task information.

(Appendix 20)
Computer
The access area storage means for storing a set of the task identification information and access information including a size of the memory area accessed by the task identified by the task identification information;
When the task type information represents motion estimation, the motion vector of each of a plurality of surrounding blocks of the target block is acquired, and the accuracy of the prediction vector is estimated from each of the acquired motion vectors Means,
Using the estimated accuracy, the size of the memory region is such that the higher the accuracy, the smaller the size of the memory region accessed by the task that processes the corresponding target block. 20. The task scheduling program according to appendix 19, wherein the task scheduling program is operated as an access area correction unit that corrects an error.

(Appendix 21)
Computer
The task scheduling program according to appendix 19 or 20, wherein the task scheduling program is operated as task processing amount correcting means for correcting the processing amount so that the processing amount decreases as the accuracy increases, using the estimated accuracy.

(Appendix 22)
Computer
When the task type information represents motion estimation, the prediction vector is acquired, and as the access information, coordinates of a prediction vector finger point that is a point moved by the prediction vector from the position of the target block are calculated, and the task type When the information represents motion compensation, the task obtains the motion vector of the target block, and calculates, as the access information, the coordinates of a motion vector finger point that is a point moved by the motion vector from the position of the target block Access acquisition means;
Supplementary notes for calculating the distance between the coordinates as the access information associated with the task specifying information of two tasks and operating as the task correlation calculating means for obtaining the task correlation between the two tasks The task scheduling program according to any one of 16 to 19.

  The present invention has been described above with reference to the embodiments, but the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

  The present invention can be applied to a processing task scheduling apparatus in an apparatus equipped with a cache memory, and in particular, to an apparatus for compressing and encoding moving images, moving image compression and encoding software corresponding to multi-core and menicore, and the like.

1, 3, 4, 5, 6, 7 Task scheduling apparatus 101 Task acquisition unit 102 Task access acquisition unit 103 Task correlation calculation unit 104 Scheduling unit 105 Access region storage unit 106 Output unit 107 Access region estimation unit 303 Predictive vector accuracy estimation unit 304 Access area correction unit 406 Group access calculation unit 407 Group storage unit 508 Task processing amount estimation unit 509 Task processing amount storage unit 614 Task processing amount correction unit 704 Group access storage unit 1000 Information processing device 1001 Task generation unit 1002 Task execution unit 1003 Processor 1004 Execution result output unit 1005 Motion vector storage unit 1006 Data input unit 1007 Data storage unit

Claims (10)

Task acquisition means for acquiring a plurality of task information including task identification information for identifying a task that is a unit of processing;
Task access acquisition means for acquiring access information representing an arrangement of a memory area accessed by the task, and storing the access information in an access area storage means in association with the task identification information;
Access area storage means for storing a set of the task identification information and the access information;
Task correlation calculating means for calculating a task correlation representing an overlap amount of memory areas accessed by the two tasks from the access information of the two tasks;
Among the combinations of the first or last task in the task sequence which is the sequence of execution of the tasks and any of the tasks not included in the task sequence, the tasks of the combination having the highest task correlation are consecutive. Scheduling means for adding the task not included in the task sequence to the task sequence,
An output means for outputting information representing the task sequence. The task acquisition means for acquiring the task information further including task content information representing the content of processing of the task;
From the task information and the access information, including a task processing amount estimation means for estimating a processing amount representing the degree of calculation amount of the task represented by the task information,
The task correlation calculating means further includes: one or more of the tasks, wherein the total processing amount of each of the tasks included in each of the task groups configured by the task sequence is less than a predetermined value. Each of the task groups includes a group correlation that represents a size of an overlap between a memory area accessed by each task and a memory area accessed by the task included in the target task group that is the task group. Not calculated for each of the tasks
The scheduling means adds the task having the highest group correlation with the target task group among the tasks not included in any of the task groups to the task column of the target task group. ,
The task scheduling apparatus according to claim 1, wherein the output unit outputs information representing the task sequence for each task group. The task acquisition means for acquiring the task information further including task content information representing the content of processing of the task;
Task processing amount estimation means for estimating a processing amount representing the degree of calculation amount of the task represented by the task information from the task information,
The task correlation calculation means is further included in any one of the plurality of task groups and a memory area accessed by each task included in each task group configured by the task sequence of one or more of the tasks. Calculating a group correlation for each of the plurality of task groups, which is selected from each of the non-tasks and represents a size of an overlap of memory areas accessed by the task;
The scheduling means sequentially selects a target task from the tasks not included in any of the plurality of task groups, and the group correlation with the selected target task for each of the plurality of task groups And calculating an evaluation value from the processing amount of the target task, adding the target task to the task column of the task group having the highest evaluation value among a plurality of task groups,
The task scheduling apparatus according to claim 1, wherein the output unit outputs information representing the task sequence for each task group. A target frame in which the task content information is a frame including task type information indicating the type of the task and a target block that is a target area for processing the task in a moving image including a plurality of time-series frames. Information for specifying the position of the target block, and information for specifying a reference frame in the moving image that is referred to in the processing for the target block of the task. The task acquisition means for acquiring the task information;
The task type information is obtained by estimating the position of the region corresponding to the target block in the reference frame, and the position in the reference frame corresponding to the position of the target block in the target frame. In the case of representing motion estimation, which is a process including calculation of a motion vector of the target block, which represents a distance and direction to the estimated position in a reference frame, a plurality of surroundings of the target block in the target frame Each of the motion vectors already calculated for the block of the current block, generate a prediction vector that is a predicted value of the motion vector of the target block from each of the acquired motion vectors, and generate the predicted vector and the task An area in the reference frame accessed by the task specified by the task specifying information from information And generating the access information representing an arrangement of the memory area storing the estimated area, and when the task type information represents motion compensation, obtaining the motion vector of the target block, The task scheduling apparatus according to claim 1, further comprising: task access estimation means for generating the access information from a motion vector and the task information. The access area storage means for storing a set of the task identification information and access information including a size of the memory area accessed by the task identified by the task identification information;
When the task type information represents motion estimation, the motion vector of each of a plurality of surrounding blocks of the target block is acquired, and the accuracy of the prediction vector is estimated from each of the acquired motion vectors Means,
Using the estimated accuracy, the size of the memory region is such that the higher the accuracy, the smaller the size of the memory region accessed by the task that processes the corresponding target block. The task scheduling apparatus according to claim 4, further comprising: an access area correction unit that corrects The task scheduling apparatus according to claim 4, further comprising: a task processing amount correction unit that corrects the processing amount so that the processing amount decreases as the accuracy increases, using the estimated accuracy. When the task type information represents motion estimation, the task access acquisition unit acquires the prediction vector, and as the access information, coordinates of a prediction vector finger point that is a point moved by the prediction vector from the position of the target block When the task type information represents motion compensation, the motion vector finger point is a point obtained by obtaining the motion vector of the target block and moving as the access information by the motion vector from the position of the target block. The coordinates of
The task correlation calculating unit calculates a distance between the coordinates, which is the access information associated with the task specifying information of two tasks, and obtains the task correlation between the two tasks. 5. The task scheduling apparatus according to any one of 4 above. A task scheduling apparatus according to any one of claims 1 to 7;
A processor;
Task generation means for generating task information of a task to be executed by the processor and inputting the task information to the task scheduling device;
A task scheduling system comprising: task execution means for acquiring information representing the task sequence from the task scheduling device and causing the processor to execute the task according to the order of the acquired task sequence. Get multiple pieces of task information including task identification information that identifies the task that is a unit of processing,
Obtaining access information representing the location of the memory area accessed by the task, storing it in the access area storage means in association with the task identification information;
Storing a set of the task identification information and the access information in an access area storage means;
From the access information of the two tasks, calculate a task correlation that represents the overlap amount of the memory areas accessed by the two tasks,
Among the combinations of the first or last task in the task sequence, which is the sequence of execution of the tasks, and any of the tasks not included in the task sequence, the tasks of the combination having the highest task correlation are consecutive. So that the task not included in the task column is added to the task column,
A task scheduling method for outputting information representing the task sequence. Computer
Task acquisition means for acquiring a plurality of task information including task identification information for identifying a task that is a unit of processing;
Task access acquisition means for acquiring access information representing an arrangement of a memory area accessed by the task, and storing the access information in an access area storage means in association with the task identification information;
Access area storage means for storing a set of the task identification information and the access information;
Task correlation calculating means for calculating a task correlation representing an overlap amount of memory areas accessed by the two tasks from the access information of the two tasks;
Among the combinations of the first or last task in the task sequence, which is the sequence of execution of the tasks, and any of the tasks not included in the task sequence, the tasks of the combination having the highest task correlation are consecutive. Scheduling means for adding the task not included in the task sequence to the task sequence,
A task scheduling program that operates as output means for outputting information representing the task sequence.

Images