JP3364274B2 - Task assignment device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to assignment of important tasks in a parallel processing computer, and in particular, by assigning tasks and taking parallelism out at the same time, it is possible to achieve optimal optimization among limited hardware resources. The present invention relates to a task assigning device capable of taking out parallelism and assigning tasks.

[0002]

2. Description of the Related Art In recent years, computers using parallel processing have been frequently used. Among them, a processing device that executes a task such as an arithmetic unit and a task allocation device that determines a time step are important. Here, the task means a basic unit of parallel processing such as an expression, a control statement, a loop statement, a function and an object in a program. As a task allocation device, for example, Srinivas D
evadas and A. Richard Newton, "Algorithms for Hardwa
re Allocation in Data Path Synthesis "(IEEE Transac
tions on Computer-Aided Design, Vol.8, No.7, July 198
9, pp.768-781) is known.

A conventional task allocation device will be described below with reference to FIG. In FIG. 11, reference numeral 1 denotes a binary operation expanding means for expanding a task into a binary operation to extract parallelism, 2 denotes initial task allocation data, initial task allocation means for obtaining an evaluation value thereof, and 3 a task Task moving means for moving one or more of the above to different arithmetic units and time steps, 4 is a task allocation data evaluating means for obtaining an evaluation value of task allocation data that has been updated with the movement of the task, and 5 is current task allocation data. And task allocation data selecting means for deciding the next task allocation data from the evaluation value of the new task allocation data, and 6 is an end judging means for judging the end of the processing.

In the above-mentioned structure, first, the binomial operation expanding means 1 expands a given program into a binary operation. The expansion to this binary operation determines which task and which task can be executed in parallel. The method of extracting the parallelism at the time of the binary operation expansion is described in detail in the massively parallel processing compiler (Yoichi Muraoka, Corona Co.).

Next, from the initial task assignment means 2 to the end determination means 6, science (S. Jarkpatric, CDGela
tt and MPVecchi, '' Optimization by Simulated Anne
aling '', Science, vol. 220, No. 4598, pp. 671-680, 1983), the optimal task assignment is obtained by the configuration using the simulated annealing method. The operation will be described below.

The initial task allocating means 2 appropriately allocates which arithmetic unit and at which time step the task expanded into the binomial operation is executed, and sets it as initial task allocation data. This may be assigned randomly, or the input data may be used as it is. The task moving means 3 randomly selects one task and moves it to a different computing unit and time step. The task allocation data evaluation means 4 obtains an evaluation value of the new task allocation data as the task moves. The evaluation value is calculated from the required number of arithmetic units and the number of time steps.
The task allocation data selecting means 5 uses the current task allocation data as the next task allocation data from the evaluation value of the current task allocation data, the evaluation value of the new task allocation data, and a parameter called temperature indicating the progress of the process. Or whether to use new task assignment data. Here, the evaluation value difference ΔC
When (evaluation value of new task allocation data-evaluation value of current task allocation data) is Δc <0, the new task allocation data is set as the next task allocation data, and Δc> 0.
In the case of, the new task allocation data is used as the next task allocation data with a probability of exp (-Δc / T). In other cases, the current task assignment data is used as the next task assignment data. Where T is the temperature parameter,
The value decreases as the process progresses. By probabilistically adopting new task allocation data even when the evaluation value deteriorates, and by decreasing the probability of adopting new task allocation data when the evaluation value deteriorates as the processing progresses, It is possible to search task assignment data having an optimum evaluation value. The end judging means 6 judges whether or not the evaluation value is not improved even if the processing is repeated. If the evaluation value does not decrease, the processing is ended, and if it decreases, the processing from the task moving means 3 is repeated.

[0007]

However, in the above-mentioned conventional configuration, since the expansion of the program to the binary operation and the assignment of the task are executed separately, when the parallelism is maximized, the storage capacity and the like are reduced. It requires a large amount of hardware resources, and what part of parallelism can be extracted and
There is a problem that it is not possible to find out whether the optimum parallelism can be taken out from the limited hardware resources.

The present invention solves such a conventional problem. By performing task transformation and task assignment at the same time, optimum parallelism can be taken out in limited hardware and the execution speed can be increased. It is an object of the present invention to provide a task assignment device that can be realized.

[0009]

In order to achieve the above object, the present invention provides a task allocation for executing task allocation processing.
Generates initial task allocation data as a allocating device
Initial task allocation means and generated initial task allocation
Intermediate conversion of tasks that include multiple operations based on assigned data
Transformed into multiple tasks consisting of a single operation for finding numbers
The task transformation means to perform and the plurality of decomposed tasks
Task moving means to move to multiple computing means,
Task assignment data updated by movement or transformation
The evaluation value based on the time step to execute
Task assignment data evaluation means to be evaluated by
And new task assignment data and current task assignment
Out of the data, the task assignment data is selected stochastically.
Task allocation data selection means to be used and the evaluation value
End determination means for determining the end of the process when no more
And the task transforming means formulates the plurality of tasks.
Task fusion method that fuses to one task by using substitution of
And separate tasks with multiple operations into several tasks
It is composed of a task separation means for performing.

[0010]

According to the present invention, by the above configuration, the allocation is started from the task allocation data given by the initial task allocation means, and the movement and modification of the task are repeated,
Optimal parallelism can be taken out and the execution speed can be increased with limited hardware resources.

[0011]

【Example】

(Embodiment 1) Hereinafter, a first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of a task assignment device according to the first exemplary embodiment of the present invention. In FIG. 1, 11 is an initial task allocation means for generating initial task allocation data, 12 is a task deforming means for deforming a task based on the generated initial task allocation data, and 13 is a task moving means for moving a task. , 14 is a task assignment data evaluation means for evaluating the task assignment data that has been updated by moving or transforming the task, and 15 is the next task assignment data determined from the evaluation values of the new task assignment data and the current task assignment data. Task assignment data selection means, 16 is an end determination means for determining the end of processing.

FIG. 2 is a schedule table showing the arithmetic units that execute the tasks and the time steps to be executed in this embodiment.

The operation of the task allocating apparatus having the above configuration will be described below. As an example of task assignment, consider assigning an expression to a computing unit. As a condition, it is assumed that only one operation can be executed in the same arithmetic unit and time step, and the movement range of the expression is limited to the range that does not change the execution result of the program. Further, here, the description of the control statement is based on the description method of C language.

When the initial task allocation means 11 reads the input program and generates the initial task allocation data as shown in FIG. 2A, the generated initial task allocation data is stored in the task allocation data evaluation means 14. It is evaluated based on the assigned task assignment data.
Here, as the evaluation method, it is assumed that the evaluation value is better as the execution time required for execution, that is, the total number of time steps to be executed is smaller, and the number of necessary variables is smaller, and the better the evaluation is, the smaller the evaluation value is. . However, when there are a plurality of operations in the same arithmetic unit and time step, it is determined as inexecutable and a bad evaluation value is given. Figure 2
Since the task y = a + b + c + d + e + f in (a) has a plurality of operations, the evaluation value is relatively poor. Here, the evaluation value is 10.

Next, either or both of the task transforming means 12 and the task moving means 13 are executed. These choices are randomly determined. Now, it is assumed that both the task transforming means 12 and the task moving means 13 are executed as a result of the random selection. The task transforming means 12 is shown in FIG.
The task y = a + b + c + d + e + f in (a) is assigned to task # 1 = a + b, task
# 2 = c + d, task # 1 = # 1 + # 2, task # 2 = e + f, task y = # 1 + #
Disassemble into 2. Here, # 1 and # 2 are intermediate variables generated by the task transforming means. In this embodiment, up to two intermediate variables can be generated.

Next, the task moving means 13 moves the tasks decomposed by the task transforming means 12 as shown in FIG. 2 (b). Then, the task allocation data evaluation means 14 evaluates the task allocation data renewed by the task transformation means 12 and the task movement means 13. Here, there is no more than one calculation on the same computing unit and time step, and the required number of time steps is as small as 4, so 4 is given as the evaluation value.

Next, the task allocation data selecting means 15
Then, the evaluation value 4 of the new task allocation data and the evaluation value 10 of the current task allocation data are compared to determine which task allocation data is to be used as the next task allocation data. Here, task assignment data is selected based on the simulated annealing method. That is, when the difference ΔC (evaluation value of new task allocation data-evaluation value of current task allocation data) between evaluation values is Δc <0, the new task allocation data is set as the next task allocation data, and when Δc> 0. , Exp
The new task allocation data is used as the next task allocation data with a probability of (-Δc / T). In other cases, the current task assignment data is used as the next task assignment data. It should be noted that T is a temperature parameter, and its value is decreased as the processing progresses. Here, the new task assignment data has a better evaluation value, so this new task assignment data is adopted as the next task assignment data. The end determining unit 16 determines the end of the process, ends the process when the evaluation value does not improve even after repeating the conversion several times, and otherwise repeats the process from the task transforming unit 12. Here, since the evaluation value has improved, the process returns to the task transformation unit 12 and the processing is repeated.

When both the task transforming means 12 and the task moving means 13 are executed again, the task transforming means 12 will be executed as shown in FIG.
Task # 1 = a + b, task # 2 = c + d, task # 1 = # 1 + # in (b)
2, task # 2 = e + f, task y = # 1 + # 2 to task # 1 = a + b, task # 1 = # 1 + c, task # 2 = d + e, task # 2 = # 2 + f, task y = # 1
It is transformed into + # 2, and the task moving means 13 moves the tasks decomposed in this way as shown in FIG. 2 (c).

In the task assignment data evaluation means 14,
Evaluate the new task assignment data. In the new task assignment data, the required number of time steps has become 3, so that the evaluation value is further improved and the evaluation value of 3 is given. In the task assignment data selecting means 15, the evaluation value is improved, so that the new task is again provided. Adopt allocation data. The end determination means 16 repeats the processing after the task transformation means 12 until the evaluation value does not change. After several iterations, there is finally no evaluation value higher than that of FIG. 2C, so the task assignment data of FIG. 2C is output as the final result.

As described above, in the first embodiment, under the constraint that only two intermediate variables can be generated, optimum parallelism is extracted to obtain the task allocation with the fastest execution time. You can

(Second Embodiment) Next, a second embodiment of the present invention will be described with reference to the drawings. The task assigning device of this embodiment differs from that of the first embodiment shown in FIG. In FIG. 3, reference numeral 21 is a task fusing means for fusing a plurality of tasks into one task by substituting expressions, and 22 is a task fusing means for separating a task having a plurality of operations into several tasks.

FIG. 4 is a schedule table showing the arithmetic units that execute the tasks and the time steps to be executed in this embodiment.

The operation of the task assigning device having the above configuration will be described below. First, initial task allocation means 11 is used as initial task allocation data.
As a result, the task assignment shown in FIG. 2A is obtained as in the first embodiment, and the evaluation value thereof is 10. Further, as in the first embodiment, it is assumed that only two intermediate variables can be generated.

Next, either or both of the task fusion means 21 and the task separation means 22 in the task transformation means 12 are executed. These choices are randomly determined. Also, how tasks are separated and how they are merged is randomly determined. Here, it is assumed that only the task separation unit 22 is executed as a result of the random selection. The task separating means 22 takes out an operation for randomly separating from the task y = a + b + c + d + e + f, and the task # 1 = a + b, task # 2 = c + d in FIG. 4A. Then, the task is separated into y = # 1 + # 2 + e + f and moved by the task moving means 13.

Next, the task assignment data evaluation means 14
As a result of executing, the evaluation value of the task assignment data of FIG. Next, the task allocation data selection means 15 is executed, and this new task allocation data is selected as the next task allocation data.
Next, the end determination unit 16 returns to the task transformation device 12 again, and this time, the task fusion unit 21 is selected to be executed, and the task of FIG.
# 2 = c + d, task y = # 1 + # 2 + e + f is selected. As a result of the fusion of the tasks, task # 1 = a + b, task y = # 1 + c + d in FIG. 4B.
+ e + f is obtained. The evaluation value of this task allocation data is 9. Therefore, although the evaluation value is worse than the evaluation value 8 of the current task allocation data, as a result of the probabilistic selection by the task allocation data selecting means 15, this new task allocation data is adopted as the next task allocation data. By repeating these operations until the evaluation value does not decrease, it is possible to obtain the task assignment in FIG. 2C having the best evaluation value.

As described above, in the second embodiment, under the constraint that only two intermediate variables can be generated, the optimum parallelism is extracted to obtain the task assignment with the fastest execution time. You can

(Embodiment 3) Next, a third embodiment of the present invention will be described. Also in this embodiment, as shown in FIG. 5, the configuration of the task allocation device is the same as that of the first embodiment shown in FIG.
But the configuration of the task transforming means 12 is different. In FIG. 5, 31 is a loop statement expanding means for expanding a task in the loop statement into a plurality of tasks by expanding the loop statement, and 32 is a loop generating means for collecting a plurality of tasks into one task to generate a loop. is there. The execution time when a loop statement is included is the number of time steps in the loop times the number of loops.

FIG. 6 is a schedule table showing the arithmetic units that execute the tasks and the time steps to be executed in this embodiment.

The operation of the task assigning device having the above configuration will be described below. First, as initial task allocation data, initial task allocation means 1
1, the task containing the loop statement, task for (i = 1,1
5,), task x (i) = a (i) + b (i), task for (i = 1,3,), y (i) = c
It is assumed that (i) + d (i) is assigned as shown in FIG. Further, in this embodiment, the total number of tasks including the loop statement is limited to 8 or less due to the storage capacity constraint. The execution time required to execute the initial task allocation data is {time step 2 of two loop statements} + {first loop number 15} * {time step number 1 in the loop} + {next loop number 15} * {Time step 1 in the loop} gives a total of 32. Here, this is evaluated value 3
Set to 2.

Next, either or both of the loop statement expansion means 31 and the loop generation means 32 in the task transformation means 12 are executed. These choices are randomly determined. Also, how many loops to unroll and how many loops to generate are randomly determined. Here, it is assumed that the loop unrolling means 31 is executed as a result of the random selection. The loop unrolling means 31 is shown in FIG.
There are 5 steps for repeating task for (i = 1,15,) in (b).
It is expanded so that and becomes task for (i = 1,15,5). Then task x (i + 1) = a to interpolate the reduced loop
(i + 1) + b (i + 1), task x (i + 2) = a (i + 2) + b (i + 2), task x
(i + 3) = a (i + 3) + b (i + 3), task x (i + 4) = a (i + 4) + b (i + 4),
Task x (i + 5) = a (i + 5) + b (i + 5) is created and added to the loop. Next, the task moving means 13 is executed to execute the processing shown in FIG.
Tasks are assigned as shown in (b). Next, the task allocation data evaluation means 14 is executed, and {the number of loop statements is 2} * as the evaluation value of the task allocation data.
An evaluation value of 23 is given because a total of 23 is obtained by {number of initial loops 3} * {number of time steps in the loop 2} + {number of next loops 15} * {time step 1 in the loop}. Next, the task allocation data selection means 15 is executed, and the evaluation value has improved, so this new task allocation data is adopted as the next task allocation data. Then, the end judging means 16 returns to the deforming means 12 again to repeat the processing.

This time, the loop generating means 32 is executed,
6 (b), task x (i + 1) = a (i + 1) + b (i + 1), task x (i +
2) = a (i + 2) + b (i + 2), task x (i + 3) = a (i + 3) + b (i + 3), task x (i + 4) = a (i +4) + b (i + 4), task x (i + 5) = a (i + 5) + b (i + 5)
In a loop and task for (i = 1,15,5)
By setting (i = 1,15,), the task assignment data in FIG. 6A is generated again. This time, task moving means 13
Is not executed, the task assignment data evaluation means 14,
The task allocation data selecting means 15 and the end judging means 16 are executed. The evaluation value of the new allocation allocation data becomes 32 again, and the evaluation value of the current allocation allocation data is 2
Although it is worse than 3, the new task assignment data is adopted assuming that new placement data is stochastically adopted. By repeating these processes in this way, it is possible to finally find a way of unrolling the loop that minimizes the execution time. Here, FIG.
The task assignment data shown in (c) can be finally found, and its execution time is {number of loop statements 2} + {initial number of loops 5} * {number of time steps in loop 1}
+ {Next loop number 5} * {number of steps in the loop 1} gives a total of 12, and the evaluation value becomes 12.

As described above, in the third embodiment, the optimum task parallelism can be obtained by taking out the optimum parallelism for the task assignment in which the loop statement exists.

(Fourth Embodiment) Next, a fourth embodiment of the present invention will be described. Also in this embodiment, as shown in FIG. 7, the configuration of the task allocation device is the same as that of the first embodiment shown in FIG.
However, the configuration of the task transformation unit 12 is different. In FIG. 7, reference numeral 41 denotes a condition set generation means for generating a set of tasks to be executed assuming that the condition statement is satisfied from the tasks below the condition statement and a set of tasks to be executed assuming that the condition statement is not satisfied. Reference numeral 42 is a condition set releasing means for releasing a part or all of the task sets generated by the condition set generating means.

FIG. 8 is a schedule table showing the arithmetic units that execute the tasks and the time steps to be executed in this embodiment.

The operation of the task assigning apparatus having the above configuration will be described below. As a result of the conditional statement, the tasks that operate exclusively cannot be executed at the same time, so it is assumed that they can exist in the same arithmetic unit and time step (see FIG. 8A). . In addition, the number of intermediate variables is not particularly limited in this embodiment.

First, as initial task allocation data,
It is assumed that the initial task allocation means 11 has obtained the task allocation data shown in FIG. 8A, and the evaluation value at this time is 4.

Next, either or both of the condition set generation means 41 and the condition set release means 42 in the task transformation means 12 are executed. These choices are randomly determined. Also, how many tasks after the conditional statement are to be set, and how many tasks in the conditional set are to be released are also randomly determined. Here, it is assumed that the condition set generation means 41 is executed as a result of the random selection. The condition set generation means 41 uses task if (x> 0), task # 1 = c + d, task # 1 = e.
+ f, task z = # 1 * g to condition set {task # 1 = c + d, task # 2 =
# 1 * g} and condition set {task # 3 = e + f, task # 4 = # 3 * g} and conditional statement task if (x> 0), task z = # 2, task z = # 4 To do. Next, the task transfer means 13 determines the task assignment as shown in FIG. Next, the task allocation data evaluation means 14 and the task allocation data selection means 1
5 and the end determination means 16 are executed. As a result, it is assumed that the evaluation value becomes 4. The process is repeated from the task transforming means 12 again, and the condition pair releasing means 42 is executed this time. The condition set canceling means 42 uses the condition set {task # 1 = c + d, task # 2 = # 1 * g} and the condition set {task # 3 = e + f, task # 4 = # 3 *.
g} Release each last task from the condition set, and set the condition set to the condition set {task # 1 = c + d} and condition set {task # 3 = e + f}.
And The canceled expression is executed after the conditional statement (Fig. 8).
(See (c)), and the evaluation value 3 is given this time.

As a result of repeating the above processing, FIG.
Since there is no evaluation value equal to or higher than the task allocation data shown in FIG. 8, the task allocation data shown in FIG. 8C is finally obtained.

As described above, in the fourth embodiment, the optimum task parallelism can be obtained by taking out the optimum parallelism for the task assignment in which the conditional statement exists.

(Fifth Embodiment) Next, a fifth embodiment of the present invention will be described. Also in this embodiment, as shown in FIG. 9, the configuration of the task allocation device is the same as that of the first embodiment shown in FIG.
But the configuration of the task transforming means 12 is different. In FIG. 9, 51 is a task transformation rule storage means 5
Task transformation rule execution means for selecting and executing one transformation rule from the task transformation rules stored in No. 2, and task transformation rule storage means 52 for storing task transformation rules.

FIG. 10 shows an algorithm of the parallelism extraction rule of the sum difference calculation, which is one of the task modification rules in this embodiment.

The operation of the task assigning apparatus having the above configuration will be described below. Note that FIGS. 2A and 2C used in the first embodiment are used as the schedule table showing the arithmetic units that execute the tasks and the time steps to be executed in the present embodiment.

First, as initial task allocation data,
It is assumed that the initial task allocation means 11 has obtained the task allocation data shown in FIG. Next, the task modification rule execution means 51 in the task modification means 12 randomly selects one task modification rule from the task modification rules stored in the task modification rule storage means 52, and the task modification rule execution means 51 , As a task transformation rule,
Select and execute the parallelism extraction rule of the sum difference operation. Hereinafter, this process will be described with reference to FIG.

First, at step 61, one task is randomly selected from a plurality of tasks consisting of only difference and sum operations. Here, the task y = a + b + c + d + e + f will be selected. Next, in step 62, the number of usable intermediate variables is checked, and the task selected in step 61 is separated into the number of intermediate variables. That is, task y = a + b +
c + d + e + f will be separated into task # 1 = a + b + c and task # 2 = d + e + f. Next, in step 63, step 6
Each of the tasks separated in 2 contains only one operation and is separated so that only one intermediate variable is used. That is, task # 1 = a + b + c is the same as task # 1 = a + b,
Task # 2 = d + e + f is separated into # 1 = # 1 + c, task # 2 = d + e, and task # 2 = # 2 + f. Next, step 6
In 4, a task y = # 1 + # 2 is generated that fuses the intermediate variables of the tasks separated in step 62. Therefore, as the execution result of the task transformation rule executing means 51, task # 1 = a + b, task # 1 = # 1 + c, task # 2 = d + e, task # 2 = # 2 + f, task y = # 1
You get + # 2. Subsequently, the task moving means 13, the task allocation data evaluating means 14, the task allocation data selecting means 15 and the end judging means 16 are executed, the movement of the task is repeated, and finally the optimum task allocation shown in FIG. Can be obtained.

As described above, in the fifth embodiment, the task allocation with the fastest execution time can be obtained by extracting the optimum parallelism using the knowledge base.

[0046]

As described above, the present invention is the first task allocation device important in a computer using parallel processing.
Initial task assignments that generate period task assignment data
Based on means and generated initial task assignment data
A single operation that computes an intermediate variable for a task that contains multiple operations
Task transformation means that transforms into multiple tasks composed of
And move the decomposed multiple tasks to multiple computing means.
Depending on the task moving means that moves and the task moving or transforming
The new task assignment data
Evaluation by giving an evaluation value based on the mustep
Task allocation data evaluation method and new task allocation
Probability of hit data and current task assignment data
Task assignment data to select task assignment data
Data selecting means, task transforming means, task moving means
And the processing operation by the task assignment data evaluation means
Repeat until the evaluation value does not change and the evaluation value decreases.
End determination means for determining the end of the process when no more
And the task transforming means formulates the plurality of tasks.
Task fusion method that fuses to one task by using substitution of
And separate tasks with multiple operations into several tasks
Since it is configured to include the task separation means that performs the task allocation, the allocation is started from the task allocation data given by the initial task allocation means, and the movement and modification of the tasks are repeated, so that the optimum hardware resource can be obtained. The effect that the parallelism can be taken out and the execution speed can be increased can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a task allocation device according to a first exemplary embodiment of the present invention.

FIG. 2A is a list diagram showing an initial state of task assignment in the first embodiment of the present invention. FIG. 2B is a list diagram showing an intermediate state of task assignment in the first embodiment of the present invention. List view showing the task allocation result state in the first embodiment

FIG. 3 is a block diagram showing a configuration of a task allocation device according to a second exemplary embodiment of the present invention.

FIG. 4A is a list diagram showing an initial state of task assignment in the second embodiment of the present invention. FIG. 4B is a list diagram showing an intermediate state of task assignment in the second embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a task allocation device according to a third exemplary embodiment of the present invention.

FIG. 6A is a list diagram showing an initial state of task allocation including a loop statement in the third embodiment of the present invention; and FIG. 6B shows an intermediate state of task allocation including a loop statement in the third embodiment of the present invention. List diagram (c) List diagram showing a task assignment result state including a loop statement in the third embodiment of the present invention

FIG. 7 is a block diagram showing a configuration of a task allocation device according to a fourth exemplary embodiment of the present invention.

FIG. 8A is a list diagram showing an initial state of task allocation including a conditional statement in the fourth embodiment of the present invention. FIG. 8B shows an intermediate state of task allocation including a conditional statement in the fourth embodiment of the present invention. List diagram (c) List diagram showing a task assignment result state including a conditional statement in the fourth example of the present invention

FIG. 9 is a block diagram showing a configuration of a task allocation device according to a fifth exemplary embodiment of the present invention.

FIG. 10 is a flowchart showing the algorithm of the sum / difference parallelism extraction rule in the fifth embodiment of the present invention.

FIG. 11 is a block diagram showing the configuration of a conventional task allocation device.

[Explanation of sign]

11 Initial task allocation means 12 task transformation means 13 task transfer means 14 Task assignment data evaluation means 15 Task allocation data selection means 16 Termination determination means 21 Task fusion means 22 means for separating tasks 31 Loop sentence expansion means 32 loop generation means 41 condition set generation means 42 Condition group releasing means 51 task transformation rule execution means 52 task transformation rule storage means

─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Hitoshi Araki 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (72) Inventor, Kato 1006 Kadoma, Kadoma City, Osaka Matsushita Electric Industrial Incorporated (56) References JP-A-3-260837 (JP, A) JP-A-6-301742 (JP, A) JP-A-4-70931 (JP, A) Matou et al., Parallel compilation method, information processing Proceedings of the 44th Annual Meeting of the Japanese Society of the Society, published on February 24, 1992, Information Processing Society of Japan, Volume 5, p. 5-97-98 Mato et al., Based on genetic rules Parallel rule-based annealing, IEICE technical report, Vol. 92, No. 185, Published August 21, 1992, The Institute of Electronics, Information and Communication Engineers, p. 1-8 Boundaries et al., New methods of loop unrolling and software piping, Proceedings of the 44th IPSJ National Conference (published on February 24, 1992), Information Processing Society of Japan, Fifth volume, p. 5-101-102 Takeuchi et al., Improvement of microscopic parallelism by optimizing expression evaluation, Proc. Of the 44th national conference of Information Processing Society of Japan (first half of the 4th year), February 24, 1992, incorporated foundation. IPSJ, Volume 5, p. 5-103 to 104 (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06F 17/50 G06F 9/45

Claims (2)

(57) [Claims] 1. A plurality of initial task assignment means for generating initial task assignment data, and a plurality of tasks each including a plurality of operations based on the generated initial task assignment data Based on the time step for executing the task transforming means for transforming into a task, the task moving means for moving the decomposed plurality of tasks to the plurality of computing means, and the task assignment data updated by the movement or modification of the task. Task assignment data evaluating means for evaluating by giving an evaluation value; task assignment data selecting means for probabilistically selecting task assignment data from new task assignment data and current task assignment data; Processing operations by means of movement and task assignment data evaluation Repeatedly until the value does not change, and comprises an end determination means for determining the end of the process when the evaluation value does not decrease, the task transformation means, the plurality of tasks into one task using the substitution of the expression A task allocating device comprising: a task merging means for merging and a task separating means for separating a task having a plurality of operations into some tasks. 2. The task modification means stores a task modification rule in a task modification rule storage means, and a task modification rule selected and executed from the task modification rules stored in the task modification rule storage means. The task assignment device according to claim 1, further comprising rule execution means.