JPH06250988A - Automatic parallel processing method - Google Patents

Abstract

PURPOSE:To suppress the generation of excess tasks by obtaining the execution condition where a conditional formula is established in parallely executing a loop including the conditional formula by the automatic parallel function of a compiler and generating a parallel code to be executed under the execution condition. CONSTITUTION:A data dependence analyzing procedure 22 analyzes the data dependence of a loop recognized by a loop recognition procedure 21. When an execution condition discrimination procedure 24 obtains the execution condition where the conditional formula is established against the loop including the conditional formula, the loop which is recognized as being parallely executed by a parallel discrimination procedure 23 is deformed into a loop which can be parallely executed under the execution condition. A memory allocation means 25 discriminates the data hierarchy, allocating memories according to the hierarchies. A parallel code generation means 26 generates a code for parallel execution.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic parallel processing method for automatically converting a target program into a parallel processing program and executing the program.

[0002]

2. Description of the Related Art In a scientific and technological calculation field such as fluid analysis, meteorological analysis, plasma analysis, structural analysis, resource exploration, optimization problem, etc., a large array is often processed using some loops. In the method of extracting such data one by one from the array according to the loop structure and sequentially executing it by the scalar instruction, it takes several days to several tens of days even if the high-performance ultra-large computer is occupied and executed. This is not uncommon, and is not practical considering cost performance. In addition, even in the case of a single processor, even in the method of parallel execution with vector instructions by a super computer equipped with a vector processor having a vector operation function, there is a limit to the speedup of the processor itself, and it is possible to dramatically improve the processing speed. I can't hope. For this reason, a multitasking method that allows parallel processing by a multiprocessor is used. In addition, in parallel processing, there is macrotasking that processes a subroutine or a group of subroutines as a unit,
There is microtasking that processes loops, sentences, or a group of sentences as a unit.

When a user uses macrotasking, it is necessary to top-down review whether the program algorithm is suitable for parallel processing and analyze the parallelism of processing and the independence of data. Requires advanced technology. Also, when using microtasking, it is necessary to focus on the local part of the program from the bottom up and analyze the parallelism of the processing for each loop iteration and each group of statements, data independence, etc. Yes,
Since very careful attention is required, automatic parallelization by the compiler is essential.

In the automatic parallelization processing, information such as a control structure such as a loop in a source program and a definition reference relation of data is analyzed, a parallel executable part having no data dependency is determined, and a code for the parallel execution is determined. To generate.
In addition, even if there is a data dependency, exclusion /
Parallel control is made possible by performing synchronous control and modifying the source program. In general, since the microtask has a small unit of instruction processing, the overhead has a large influence on the processing time of the entire program. Therefore, reduce unnecessary task generation as much as possible,
Overhead needs to be suppressed. In addition, since a task is created for a loop that contains a conditional expression in a loop that can be executed in parallel and has no executable statement other than the part that is executed when the conditional expression is satisfied, that conditional expression If is not established, useless task generation / execution / termination is performed and the parallelization processing efficiency decreases. Referring to FIGS. 2 to 5 exemplifying the conventional parallelization processing, the source program 1 calculates the value of each element of the given array A in the loop 27 of the execution statement as illustrated in FIG. Sequentially, the product with the corresponding element of the array Y is obtained only when the value is a positive value, and the result is stored in the corresponding element of the array X. When the program is translated and executed by the conventional technique, the RESERVE routine (31 in FIG. 3) is called to secure the available processors 1 to N by the task scheduler as illustrated in FIGS. 3 to 5. (51 in FIG. 5), the PARALLUL DO routine (3 in FIG. 3)
2) is called to assign each loop (33 in FIG. 3) that is made into a task to each of the N number of processors, and is executed in parallel (52 in FIG. 5), and the result is the element of array A (41 in FIG. 4). Has a positive value, array A (41 in FIG. 4) and array Y
The product of the corresponding elements of (44 in FIG. 4) is array X (4 in FIG. 4).
3) is stored in the corresponding element, and at the end of the loop (33 in FIG. 3), all tasks (in this embodiment, one task is generated for each iteration processing unit of the loop structure and executed by each processor) Waits for the end of parallel execution (53 in FIG. 5). Finally, the RELEASE routine (3 in Figure 3
4) is called to release the processor (5 in FIG. 5).
4).

[0005]

According to the method of the present invention,
For a parallel executable loop that includes a conditional expression, find the execution condition that satisfies the conditional expression before generating a task for parallel execution, and transform the parallel executable primitive loop into a loop that executes parallel under that execution condition. As a result, an optimum parallel code that does not generate / execute / end useless parallel execution tasks is generated, and the processing efficiency of the entire program is improved.

[0006]

SUMMARY OF THE INVENTION An automatic parallelization processing method of the present invention is a multitasking system in which a source program is translated to generate a target program for parallel execution, and the target program and a runtime routine are combined. In a memory sharing type multiprocessor computer system that executes an operating system that includes a linker that generates an executable program of a configuration and a task scheduler that manages tasks for parallel execution, After inputting the source program, the loop recognition means for recognizing the control structure of the loop of the executable statement repeatedly executed in the source program, and the order relation for defining and quoting the recognized data in the loop are analyzed, and the data are dependent on each other. Data dependency analysis means that recognizes whether there is a relationship Based on the recognized control structure of the loop and the dependency relationship of the data, the determination of the loop that can be executed in parallel and the data at the time of parallel execution are defined, and the exclusive / synchronous control that guarantees the order relation to be recognized is recognized. The parallelism determining means and the conditional expression in the parallel executable loop are recognized, and it is determined that there is no executable statement other than the portion executed when the conditional expression is satisfied in the parallel executable portion. An execution condition determining means for generating a loop that can be executed in parallel, deforming the loop, and determining an execution condition; and, based on the declaration information of the source program and the dependency information of data during parallel execution, A memory allocator that determines the shared data between tasks generated by the linker and the unique data of each task in the data hierarchy, and allocates it to the memory. When, wherein for a parallel executable said loop obtained by the parallel determination unit, and a parallel code generation means for generating a code for parallel execution also include code for exclusive / synchronous control, and,
The task scheduler includes a task generation unit that generates a parallel execution task based on an object program including the parallel code generated by the compiler and information for executing the object program generated by the compiler. A task allocation unit that secures an available processor when parallel execution is requested and allocates a parallel execution task to the processor, and task control that performs scheduling of the parallel execution task and exclusive / synchronous control between the tasks Means and a task releasing means for releasing the task and the processor at the end of parallel execution of the task, the parallel execution code generation by the compiler and the parallel execution task by the task scheduler and automatic memory allocation thereof. To do.

[0007]

The automatic parallelization processing method of the present invention recognizes a conditional expression in a loop that is determined to be parallel executable based on the control structure of the loop of the executable statement and the data dependency relationship, and executes the task for parallel execution. Optimal parallel code generation and memory allocation are automatically performed by finding the execution condition that satisfies the conditional expression before generating it, and transforming the above loop into a loop that does not include a conditional expression that can be executed in parallel under that execution condition. To do.

[0008]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1 is a diagram showing a configuration of an embodiment of an automatic parallelization processing method of the present invention, FIG. 2 is a diagram illustrating a program in FORTRAN language as an example of a source program 1, and FIG. 3 is a diagram showing the program of FIG. FIG. 4 is a diagram illustrating a program in the case of parallel conversion according to the conventional technique, and FIG. 4 is a diagram illustrating a data hierarchy of shared data and task-specific data and a state of memory allocation when the program in FIG. 2 is parallelized according to the conventional technique. FIG. 5 is a diagram illustrating the execution status of each task when the program of FIG. 2 is parallelized by a conventional technique, and FIG. 6 illustrates a program when the program of FIG. 2 is parallelized by the method of the present invention. FIG. 7 is a diagram illustrating a data hierarchy of shared data and task-specific data and a state of memory allocation when the program of FIG. 2 is parallelized by the method of the present invention, and FIG. FIG. 9 is a diagram illustrating an execution status of each task when a program is parallelized by the method of the present invention, and FIG. 9 is a diagram showing a flow of operations of the execution condition determining means 24 included in the compiler, which is a feature of the present invention. .

Referring to FIG. 1, this embodiment includes a compiler 2, a linker 5 and a task scheduler 7. The source program 1 is translated by the compiler 2 to generate an object program 3 for parallel execution, The target program 3 and the runtime routine 4 are linked by the linker 5, and the executable target program 6 having a multitask structure
And executes the target program 6 in parallel under the control of the task scheduler 7. According to the present invention, the following processes are performed in order to reduce the number of unnecessary tasks to be executed in parallel and increase the execution efficiency of the entire program. Referring to FIG. 1 showing the configuration of an embodiment of the present invention, the compiler 2 executes a loop recognition procedure 21 and a data dependency analysis procedure 2
2, parallelism determination procedure 23, and execution condition determination procedure 24
, Memory allocation procedure 25, and parallel code generation procedure 2
6 is included. Each of the above procedures will be described with reference to the source program illustrated in FIG.
1 analyzes the control structure of the input source program 1 and recognizes a loop (27 in FIG. 2). The data dependency analysis procedure 22 defines the data in the loop 27, analyzes the order relation to be referred to, and recognizes that each element of the arrays A, X, and Y does not have a dependency that spans the repetition of the loop. The parallelism determination procedure 23 recognizes that the loop 27 is a loop that can be executed in parallel because each data in the loop 27 has no dependency. Execution condition judgment procedure 2
4, referring to the flow chart of the operation illustrated in FIG. 9, the loop 27 is taken out (step 90 in FIG. 9), it is determined that the loop can be executed in parallel (91), and whether there is a conditional expression in the loop. It is judged (92) whether or not there is an executable statement other than the execution part when the condition of the conditional expression is satisfied in the parallel executable loop (93), and the loop 27 is executed only when the conditional expression is satisfied. Recognize that the executable statement should be executed. Next, the execution condition number storage area I where the conditional expression is satisfied
N (76 in FIG. 7) and the execution condition value storage area IM (7 in FIG. 7)
Before the loop 27, a declaration statement (61 in FIG. 6) for securing 5) in the inter-task shared data space (77) and an assignment statement 63 that sets the initial value of the number of execution conditions to 0 are generated (94). A loop 64 for obtaining the execution condition is generated (95), and the loop 27
Is transformed into a loop 66 for parallel execution under the execution conditions (96). The memory allocation procedure 25 classifies and allocates each data appearing in the program of FIG. 6 into a data hierarchy of inter-task shared data 77 and in-task unique data 78 as illustrated in FIG. 7. Parallel code generation procedure 2
6 generates a code for parallel execution including the parallel executable part recognized in the parallelism determination procedure 23 and the part generated or modified in the execution condition determination procedure 24.

According to the above description, in the method of the present invention, when the program illustrated in FIG. 2 is translated and executed, a loop to be executed (27 in FIG. 2) as illustrated in FIGS. The transformation that separates the conditional expression and the execution statement of
A loop (64 in FIG. 6) and a loop (66) are generated.
The loop (64) is executed as one task by the CALL RESERVE (62). Then the loop (66)
Calls the RESERVE routine (62) to secure the available processors 1 to N by the task scheduler (7 in FIG. 1) (81 in FIG. 8), and then PARALLLE
Loop 6 by calling the L-DO routine (65 in FIG. 6)
6 in parallel (82 in FIG. 8), and as a result, array A (7 in FIG. 7) when the elements of array A (71 in FIG. 7) have positive values.
1) and the corresponding element of array Y (74 of FIG. 7) are stored in the corresponding element of array X (73 of FIG. 7), and at the end of the loop (66 of FIG. 6) the parallel execution of the previous task Wait for the end (83 in FIG. 8). Here, since the loop (66 in FIG. 6) that is executed in parallel is executed under the execution condition where the conditional expression of the loop (27 in FIG. 2) is satisfied, useless task generation / execution / termination is not performed. Then, the RELEASE routine (67 in FIG. 6) is called to release the processor (84 in FIG. 8).

[0012]

As described above, according to the present invention,
When the loop containing the conditional expression is executed in parallel by the automatic parallelization function of the compiler, the execution condition that satisfies the conditional expression is calculated before the task for parallel execution is generated, and the primitive loop that can be executed in parallel is executed under the execution condition. By creating a loop for parallel execution in
Parallel code that does not execute / terminate is automatically generated,
The processing efficiency of the entire program can be improved.

[Brief description of drawings]

FIG. 1 is a diagram showing a functional configuration of an embodiment of an automatic parallelization processing method of the present invention.

FIG. 2 shows FORTRA as an example of the source program 1.
It is a figure which illustrates the program by N language.

FIG. 3 is a diagram illustrating a program when the program of FIG. 2 is parallelized by a conventional technique.

FIG. 4 is a diagram illustrating a data hierarchy and a state of memory allocation when the program of FIG. 2 is parallelized by a conventional technique.

FIG. 5 is a diagram illustrating an execution status of each task when the program of FIG. 2 is parallelized by a conventional technique.

FIG. 6 is a diagram illustrating a program when the program of FIG. 2 is parallelized by the method of the present invention.

FIG. 7 is a diagram exemplifying a data hierarchy and a state of memory allocation when the program of FIG. 2 is parallelized by the method of the present invention.

FIG. 8 illustrates the execution status of each task when the program of FIG. 2 is parallelized by the method of the present invention.

FIG. 9 is a diagram showing a flow of an execution condition determination procedure 24 provided in the compiler 2, which is a feature of the present invention.

[Explanation of symbols]

 1 Source Program 2 Compiler 3 Target Program 4 Runtime Routine 5 Linker 6 Executable Program 7 Task Scheduler 21 Loop Recognition Procedure 22 Data Dependency Analysis Procedure 23 Parallelism Determination Procedure 24 Execution Condition Determination Procedure 25 Memory Allocation Procedure 26 Parallel Code Generation Procedure

Claims (1)

[Claims] 1. A compiler for translating a source program to generate an object program for parallel execution, a linker for combining the object program and a runtime routine to generate an executable program having a multitask structure, and for parallel execution. In a memory sharing type multiprocessor computer system that executes an operating system including a task scheduler that manages the task of, the compiler executes the source program to be translated, and then repeatedly executes the source program. Loop recognition means for recognizing the control structure of a sentence loop, and data dependency relationship analysis means for analyzing the order relation for quoting and recognizing the recognized data in the loop, and recognizing whether or not there is a mutual dependency relationship between the data. And the control structure of the recognized loop and the dependence of the data. Parallelism determination means for recognizing the determination of the loop that can be executed in parallel based on the relationship and the exclusive / synchronous control that guarantees the order relationship for defining and quoting data at the time of parallel execution; Recognizing the conditional expression, determining that there is no executable statement other than the part to be executed when the conditional expression is satisfied in the part that can be executed in parallel, and creates a loop that can be executed in parallel and transforms the loop. Based on the execution condition determination means that determines the execution condition, and the dependency information of the source program declaration information and the data at the time of parallel execution, the inter-task shared data generated by the linker and the tasks are A memory allocating unit that discriminates the unique data that it has into a data hierarchy and allocates it to a memory, and parallel execution that is obtained by the parallelism determining unit. Parallel code generation means for generating code for parallel execution including code for exclusive / synchronous control with respect to the loop, and the task scheduler includes the parallel code generated by the compiler. Secures a task generation unit that generates a task for parallel execution based on an object program including the object program and information for executing the object program generated by the compiler, and a processor that can be used when parallel execution is requested. Task allocation means for allocating a task for parallel execution to the processor, scheduling of the parallel execution task and exclusion / interval between the tasks
Task control means for performing synchronous control, and task release means for releasing the task and the processor at the end of parallel execution of the task, the parallel execution code generation by the compiler, and the task scheduler of the task to be executed in parallel An automatic parallelization processing method characterized by executing and automatically allocating it.