JPH0264766A - Vectorizing processing system - Google Patents

Abstract

PURPOSE:To prevent the processing from being affected by the length of a working vector by assigning a working vector area for an actual length portion dynamically when actual length is longer than the fixed loop length and releasing it at the time when it becomes unnecessary. CONSTITUTION:A source analyzing part 21 in a compiler 2 analyzes a course program 1, detects an involved loop, adding a mark to it and transfers it to a vector analyzing part 22. The analyzing part 22 analyzes the loop to which the mark is added, when a vectoring possible part and a vectoring impossible part are involved in the loop, adds the different mark to the vectorizing possible part and transfers it to a loop dividing processing part 23. At the processing part 23, when loop length necessary at a stage at which the program is executed actually is shorter than the fixed length, an assigned working vector area is used as it is, when the actual loop length is longer than it, it is used with the actual loop length dynamically. Then, when the working vector area becomes unnecessary, a program part to execute releasing is inserted into the program corresponding to the program 1.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a vectorization processing method in a compiler, and particularly to a method for performing loop division using work vectors and vectorizing the same.

[Conventional technology]

In computer systems equipped with vector calculation functions, calculations between multiple regularly arranged data sequences (vector data) can be executed at high speed at once using vector instructions. It is preferable to replace program parts that can be executed with vector instructions with vector instructions. Therefore, when compiling a source program written in a compilation-based high-level language (for example, FORTRAN) to generate a target program, the program is analyzed and vector instructions are used. Replacement with instructions (
vectorization).

Note that portions that can be vectorized are generally expressed as loop structures in the source program.

By the way, in addition to loops that can be vectorized as is,
If you have a loop that has a vectorizable part and a non-vectorizable part, and you want to vectorize such a loop, you can separate the loop into the vectorizable part and the non-vectorizable part and put them into separate loops. After dividing, a program part for vector execution is generated for the vectorizable part, a program part for scalar execution is generated for the part for which vectorization rate is possible, and a target program containing both is output. I try to do that.

However, if there are variables that are defined or referenced across the division points, division cannot be performed as is. By allocating a work vector with a length (the length assumed by the compiler if the loop length is unknown at the time of compilation) and replacing variable definitions and references with work vector definitions and references, variable definitions and references can be performed. It changed the relationships to be suitable for partitioning and promoted vectorization.

Figure 3 is an example of a DO loop in a source program written in the FORTRAN language, which includes a vectorizable part and a non-vectorizable part, and a vectorizable part and a non-vectorizable part. This is an example of a case where there are variables defined and referenced across the dividing points when dividing. Note that this DO loop is executed as follows: “The first element A (1
) , B (1) C (1) , D (I), store the product of elements A (1) and B (1) in variable X, and store the product of elements c (1) and D (1) in variable Store in Y and output the values of variables X and Y. This describes a processing procedure in which the process ``is repeated from the 1st element to the Nth element.

Conventional compilers cannot vectorize this DO loop as is, so the vectorizable part X=A (+) *B (1) Y=C(1)*D (+) and the non-vectorizable part Possible part WRITE (6, *> Divide X, Y into another loop, and set variables X and Y defined across the division points to work vectors WX and W, respectively.
As shown in FIG. 4, the vectorizable part WX (1)=A (1)*B (r)WY (1)=
A program part that vector-executes a loop ■ containing C(+)*D (1);
An objective program is generated that includes a program part that scalar executes a loop (2) including the non-vectorizable part WRITE (6, *) WX (1) and WY (1).

Figure 5 is a diagram conceptually showing the structure of the generated target program, and the program part ■ in Figure 5 corresponds to the loop ■ in Figure 4, and can be executed at high speed with vector instructions. It is a part. Furthermore, the program portion (2) in FIG. 5 corresponds to the loop (2) in FIG. 4, and is a portion that is repeated N times using a scalar instruction.

[Problems to be Solved by the Invention] The conventional vectorization processing method is executed as described above, and is useful for promoting vectorization.
Since a work vector with a length equal to or longer than the loop length is required, it has the following drawbacks.

(1) When the loop length is very long or when a large number of work vectors are required at the same time, the size of the work vector area becomes large and takes up the memory available to other programs in the system. do.

(2) If the loop length is unknown at compile time, a work vector of the length assumed by the compiler is set, but in actual operation, if the loop length is longer than the set value, the operation will not work normally. do not have.

(3) If the loop length changes during execution, it is necessary to allocate a work vector with the maximum length that the loop can take, so if the loop length fluctuates significantly, use the area allocated as a work vector. ineffective.

The present invention has been proposed in view of the above points, and its purpose is to provide a vectorization processing method that is not affected by the length of a work vector, whether it is long or short, known or unknown, and constant or variable. It is in.

[Means to solve problems]

In order to achieve the above object, the present invention detects and analyzes loops in a source program written in a compiled high-level language and divides them into separate loops corresponding to vectorizable parts and non-vectorizable parts. In a vectorization processing method that allocates work vectors to variables that are defined and referenced across division points, and generates a target program consisting of a program part that executes vectors and a program part that executes scalar, A program part that secures a work vector area of a predetermined length set by the compiler, and a program part that dynamically allocates and allocates a new work vector area when the actual length exceeds the predetermined length. and a program part that releases the newly allocated work vector area when the work vector area is no longer needed.

[Effect]

The vectorization processing method of the present invention detects and analyzes loops in a source program written in a compiled high-level language and separates them into separate loops corresponding to vectorizable parts and non-vectorizable parts. It divides the program, assigns work vectors to variables defined and referenced across the division points, generates an objective program consisting of a vector execution program part and a scalar execution program part. A program part that secures a work vector area of a predetermined length set by the user, a program part that dynamically allocates and allocates a new work vector area when the actual length exceeds the predetermined length, and A program part is inserted that releases the work vector area allocated to the work vector area when the work vector area is no longer needed.

〔Example〕

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

FIG. 1 is a block diagram of an embodiment of a compiler to which the vectorization processing method of the present invention is applied.

In FIG. 1, the compiler 2 includes a source analysis section 21, a vectorization analysis section 22, a loop division processing section 23,
Intermediate text generation unit 24 and intermediate text optimization unit 25
and a target program generating section 26, which has the function of inputting the source program 1, converting it into intermediate text 3, and generating the target program 4 after performing the optimum processing. The loop division processing section 23 includes a loop division section 231 , a loop length determination section 232 , a work vector allocation section 233 , and a dynamic work vector allocation control insertion section 234 .
consists of a vector intermediate text generation section 241 and a scalar intermediate text generation section 242.

The functions and operations of each part of the above embodiment will be explained below.

First, the source analysis unit 21 in the compiler 2 analyzes the supplied source program l, detects loops included in the program, marks them, and passes them to the vectorization analysis unit 22.

The vectorization analysis unit 22 analyzes the marked loop, and if the loop includes a vectorizable part and a non-vectorizable part, adds another mark to the vectorizable part, and It is passed to the division processing unit 23. Also, if the loop contains only vectorizable parts or non-vectorizable parts,
Intermediate text generation unit 2 along with parts that do not form a loop
Pass it to 4.

The loop division unit 231 in the loop division processing unit 23 divides the vectorizable part and the non-vectorizable part so as to form separate loops, and generates a vector intermediate text in the intermediate test test generation unit 24. By calling the section 241 or the scalar intermediate text generation section 242, the intermediate text 3 for the loop in the vectorizable part and the loop in the non-vectorizable part is generated.

Next, if the loop length is known to the compiler 2, the loop length determining unit 232 checks whether the loop length exceeds a certain predetermined length (a value determined by an option or instruction to the compiler), and In this case, the predetermined length is used as the loop length of the divided loop, and in the case that the predetermined length is not exceeded, the original loop length itself is used as the loop length of the divided loop. Furthermore, if the loop length is unknown, the above-described determined length is used as the loop length of the divided loop.

The work vector allocation unit 233 allocates the work vector to the fixed area with the length determined by the loop length determination unit 232.

The dynamic work vector allocation control insertion unit 234 determines whether the actual loop length required when the program being compiled is actually executed is equal to or greater than the length determined by the loop length determination unit 232. If it is shorter, the work vector area allocated by the work vector allocation unit 233 is used as is, and if the actual loop length is longer, the work vector area is dynamically allocated using the actual loop length and used. is used to insert into the main program corresponding to the source program l a program part (control text) that performs an operation to release the work vector area when it is no longer needed at the end of the execution of the loop.

Next, the intermediate text 3 generated by the intermediate text generation unit 24 and the loop division processing unit 23 is subjected to well-known optimization processing in the intermediate text optimization unit 25, and then passed to the target program generation unit 26, The program generation unit 26 generates a target program 4 from the passed intermediate text 3.

FIG. 2 is a conceptual diagram of the structure of a target program generated by the compiler 2 to which the present invention is applied to the source program of FIG. 3 used in the explanation of the conventional example.

As can be seen by comparing with Figure 5, which shows the target program according to the conventional example, the program part ■ sets the base address of the work vectors WX and WY to point to the area allocated at the time of compilation, and the actual loop length N is determined at the time of compilation. A program part that checks whether the length is longer than the set length m, and if it is longer, dynamically secures an area for the work vectors WX and WY using the length N, and resets the base address for the work vectors wx and wy.■ and a program part (2) for releasing the secured area are inserted into the target program by the dynamic work vector allocation control insertion unit 234 of the loop division processing unit 23. Therefore, when this target program is executed, the work vector area is dynamically allocated and released as necessary, and the available memory in the system is effectively used while performance is improved through partial vectorization. Can be done.

[Effects of the Invention] As explained above, in the vectorization processing method of the present invention, when dividing a loop using a work vector, the loop length of each divided loop is determined appropriately, and the actual If the loop length is longer than the specified loop length, the target program is generated with a function to dynamically allocate a work vector area for the actual length and release it when it is no longer needed. , even when the loop length is long and the size of the work vector area is large, or when the loop length fluctuates, there is no pressure on the memory usage of the entire system and there is no reduction in usage efficiency. .

Additionally, even if the actual loop length is longer than the loop length determined by the compiler, a work vector area for the actual length is secured from the system's free memory, ensuring normal operation. be.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of a compiler to which the vectorization processing method of the present invention is applied. FIG. 2 is a conceptual diagram of the structure of a target program generated by the present invention. FIG. 4 is a diagram showing an example of a small loop, FIG. 4 is a diagram showing a state in which a DO small loop is divided, and FIG. 5 is a conceptual diagram of the structure of a target program generated by a conventional compiler. In the figure, ■・・・・・・・・・Source program 2・・・・・・
... Compiler 21 ... Source analysis section 22 ... Vectorization analysis section 23 ... Loop division processing section 231 ... Loop division section 232 ... Loop length determination Section 233...Work vector allocation section 234...Dynamic work vector allocation control insertion section 24...
...Intermediate text generation unit 241...Vector intermediate text generation unit 242...
Scalar intermediate text generation unit 25...Intermediate text optimization unit 26...Object program generation unit 3.
・・・・・・Intermediate text 4・・・・・・・・・Objective program student F8. Conceptual diagram of Oshitsuzukuri of the blind 7-chome, white sibling program Figure 2 Configuration diagram of the main stream A sequence Figure 1 Do +0 1=1. N X = A(1)*B (1) Y=C(1)*O(1) 10 WdayITE(6,*) Third
Diagram Do/Rape is divided into sections to show the power of the book. Figure 4. Generated T

Claims (1)

[Claims] It detects and analyzes loops in a source program written in a compiled high-level language, divides them into separate loops corresponding to vectorizable parts and non-vectorizable parts, and defines and references them across the division points. In a vectorization processing method that allocates a work vector to a variable in the program and generates a target program consisting of a program part to be executed as a vector and a program part to be executed as a scalar. a program part that secures a work vector area; a program part that dynamically allocates and allocates a new work vector area when the actual length exceeds the predetermined length;
A vectorization processing method characterized by inserting a program part that releases a newly allocated work vector area when the work vector area is no longer needed.