JPH04286031A - Parallel execution processing method for loop of program - Google Patents

Abstract

PURPOSE:To realize parallelling and to improve execution performance even when there is possibility that the value of a variable substituted within a loop is used beyond the loop on the loop parallel execution processing method of a program, in which the rotation of the loop in the program is executed in parallel in plural computers. CONSTITUTION:In the computers except for a computer 14-N controlling the final rotation of the loop for the variable A whose value is substituted within the loop, the copy areas A1, A2,... of the variable are used for respective computers. The computer 14-N controlling the final rotation of the loop uses the original variable A as it is. Thus, the value of the variable A is guaranteed beyond the loop.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] The present invention is applicable to, for example, FORTRAN
This invention relates to a loop parallel execution processing method for a program in which each rotation of the loop is executed on each computer in order to execute one loop in a program in parallel on multiple computers.

[0002] In scientific and technical calculations using FORTRAN, DO loops are often used when processing arrays. Therefore, if multiple computers share the rotation of this DO loop and execute it in parallel, the processing time can be shortened. This kind of parallel processing is called a loop slicing method. If the technology associated with this loop slicing method can be further improved, it will be possible to execute programs more efficiently.

0003

2. Description of the Related Art FIG. 4 is an explanatory diagram of the prior art. Assuming that there are two computers that can execute in parallel, the loop slicing method uses F as shown in Figure 4 (a), for example.
The loop of the program 40 written in ORTRAN is parallelized to create a procedure 41-1 as shown in FIG. 4(b).
, 41-2 and executed on two computers.

In the example shown in FIG. 4, odd-numbered rotations are
The rotations are executed by one computer, and the even-numbered rotations are executed in parallel by a second computer. Here, the variable A used in the loop in (a) is used as temporary data for each rotation of the loop. When executing this in parallel, it is necessary to prepare areas for A for each computer, such as A1 and A2 shown in (b), to prevent area contention. The data being processed by the first computer is
This is because if the second computer uses it, a contradiction will occur. The same applies to the opposite case.

[0005]

[Problem to be solved by the invention] In a general program, variables such as A shown in Figure 4 are rarely used only in this loop, and are usually the same in various parts of the program. variables are used (with different meanings).

[0006] At this time, if it is found through data flow analysis using the compiler's optimization function that the use of variable A is closed within a loop, that is, it can be treated as a separate variable from other uses of A. , as in the example shown in FIG. 4, it can be seen that correct results can be obtained without assigning a value to the original variable A, so parallelization is easily possible.

However, when there is a possibility that the value of variable A used in this loop will be used outside the loop (even if it is not actually referenced), it is necessary to assign the value to the original variable A. Parallelization as shown in FIG. 4(b) is impossible, and complicated processing must be performed to guarantee the value of A.

In fact, in the example shown in FIG. 4, if the value of variable A is to be used after the loop ends, which is the value A1 of the processing result by the first computer or the value A2 of the second processing result? Complex processing is required to check whether the result is the final result and to assign that value to the original variable A at the end of the loop.

[0009] The present invention aims to solve the above problems, and even when there is a possibility that the value of a variable assigned within a loop may be used outside the loop, parallelization can be executed and unnecessary processing is not performed. The purpose is to make it executable.

0010

[Means for Solving the Problems] FIG. 1 is a diagram illustrating the principle of the present invention. As in the program 10 shown in FIG.
Assume that a value is assigned to variable A in a loop, and the same variable A is referenced after the loop ends.

[0011] The compiler 11 uses D for parallel processing.
Outputs an object program in which the rotation of the O-loop is divided. The execution image of the object program is as follows. Assume that there are N computers capable of parallel processing that constitute the multiprocessor 12. calculator 1
One DO loop execution image 14-1 obtained by dividing the rotation of the DO loop is assigned to 3-1, and the computer 13-2
, assigns another DO loop execution image 14-2 in which the rotation of the DO loop is divided, ..., assigns another DO loop execution image 14-N in which the rotation of the DO loop is divided to the computer 13-N. and execute them in parallel on each computer.

[0012] Here, if the computer 13-N is in charge of the final rotation part of the loop, in the DO loop execution image 14-N of the computer 13-N, the location where the value of variable A is stored is The area of original variable A is used as is. Other computers 13-1, 13-2,...
Now, copy area A1 of variable A prepared for each computer,
Use A2,... This guarantees the value of variable A outside the loop.

[0013]

[Operation] For example, when using the loop slicing method described as the prior art, if the execution order on each computer does not break the rotational relationship in the original loop, the original loop at the end of execution on any one computer The value of the variable (A1 or A2 shown in FIG. 4B) corresponding to the variable is equal to the value that the original variable A should take.

Utilizing this, the area of the original variable A is directly used instead of the copy of A (A1 or A2) that is equal to the value that the original variable A should take. That is, as shown in FIG. 1, N computers 13-1 to 13-
If there is N, the computer variable that governs the final rotation is used as it is as the original variable A instead of the duplicate area AN. As a result, at the end of a full rotation,
The final result is guaranteed as is, and parallel processing can be performed at high speed.

[0015]

Embodiment FIG. 2 shows an example of a language processing system for implementing the present invention, and FIG. 3 shows an embodiment of the present invention. In FIG. 2, 20 is a source program written, for example, in FORTRAN, 21 is a computer consisting of a CPU and memory, 22 is a source program analysis section in the FORTRAN compiler 11, 23 is an optimization section that performs program optimization, 24 is a loop detection unit that detects a DO loop,
24 is a data flow analysis unit that analyzes the flow of data in a program; 25 is a parallelism analysis unit that analyzes whether a certain part of the program can be executed in parallel when outputting an object program for a multiprocessor;
27 is a parallel intermediate code generator that generates intermediate code for each computer when it can be executed in parallel; 28 is a code generator that allocates registers and areas and generates object program code; 29 is a compiler 11 represents the object program that is the output result of .

An example will be described in which the compiler 11 inputs, for example, the source program 20 shown in FIG. 3 and compiles it. As a parameter to compiler 11,
Suppose that triple parallelization is specified. The source program analysis unit 22 analyzes the source program 20 and passes the analysis result to the optimization unit 23 in the form of intermediate text. In the optimization section 23, the loop detection section 24 first detects D.
Detect O-loop. In this example, D with the following content:
It is an O-loop.

The data flow analysis unit 25 examines the relationships between variables inside and outside the loop. Based on the result, the parallelism analysis unit 26 determines whether parallel execution of this DO loop is possible. If parallel execution is possible, the parallel intermediate code generation unit 27 generates intermediate code for generating an object program in which the rotation of the DO loop is divided into three parts.

In this program, a value is unconditionally assigned to the variable T in the DO loop. In this case, it is clear that the value that the variable T should take is the value that is assigned at the final rotation of the loop. Therefore, the parallel execution portion assigned to each computer is fixedly divided in advance, and the computer responsible for the final rotation of the loop uses the area of the original variable T. This makes it possible to easily guarantee the final value of variable T.

From this, the parallel intermediate code generator 27
Now, as shown in Figure 3 (b), (a) the first 1/3 of the rotation of the DO loop, (b) D
Divide into the middle 1/3 of the rotation of the O loop, (c) the last 1/3 of the rotation of the DO loop, and so on.

In (a) and (b), which do not include the final rotation, the replication regions T1 and T2 of variable T are respectively
In (c), which includes the final rotation, the area of the original variable T is used as is. Note that in (b) of Figure 3, the execution image is shown in the form of a source program to simplify the explanation, but when the program is executed,
The content is expanded into machine language code that is used by computers.

[0021] When distributing this code to multiple computers for execution, for example, the well-known UNIX system (US AT
A process control function such as "fork" in the operating system developed by Bell Laboratories, Inc., may be used. There are various other distribution methods, but their explanations are omitted because they are not related to the gist of the present invention.

After executing the execution images (a) to (c) shown in (b) of FIG. 3, the processing results are “joi”
By using a control function such as "n", even if the variable T is referenced thereafter, the correct variable value assigned at the final rotation of the loop can be used.

In particular, in the FORTRAN DO loop,
Since the loop rotation number is calculated immediately before execution, it is possible to fix the combination of rotation orders during parallel execution, and this method is effective. Instead of dividing the loop into three parts such as the first part, middle part, and last part, if the value of the loop variable i is divided by 3 and there is a remainder of 1, or if the value of the loop variable i is divided by 3 and the remainder is 2 , loop variable i
For example, when the value of is divisible by 3, the rotation of the loop may be divided and assigned to each computer. However, when dividing in this way, the determination process of the computer responsible for the final rotation may become complicated.

In the example shown in FIG. 3, the rotation of the loop is divided into three parts, but it is of course possible to divide the rotation according to the number of computers that can execute the process in parallel, and the present invention can be similarly implemented with other numbers of divisions. Can be done.

[0025]

[Effect of the invention] When executing general-purpose programs in parallel,
The easiest way to parallelize is to execute each rotation of a loop in parallel. There are many programs in which variables used within a loop appear to be referenced outside the loop, and in such cases, with conventional technology, automatic parallelization is either suppressed or unnecessary synchronization processing is required. , execution performance was decreasing. According to the present invention, execution performance is improved because parallel execution can be performed in exactly the same execution time as when variables are used closed in a loop.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the present invention.

FIG. 2 is a diagram showing an example of a language processing system for implementing the present invention.

FIG. 3 is a diagram showing an embodiment of the present invention.

FIG. 4 is an explanatory diagram of a prior art.

[Explanation of symbols]

10 Program 11 Compiler 12 Multiprocessor 13-1 to 13
-N calculator

Claims (1)

[Claims] [Claim 1] A plurality of computers (13-1 to 13-N
), in a loop parallel execution processing method for a program in which the loop in program (10) is divided into certain rotation units and executed in parallel, the final value of the loop is For computers other than the computer that controls rotation, each computer has its own variable replication area (A
1, A2, ...), and in the computer that controls the final rotation of the loop, the original variable (A) is used as is, thereby guaranteeing the value of the variable outside the loop. Parallel execution processing method.