JPH0748205B2 - Vectorization processing method - Google Patents

Description

The present invention relates to a vectorization processing method in a compiler that generates an object program to be supplied to a vector processing processor.

[Conventional technology]

The vector processor can execute operations between multiple regularly-arranged data strings (vector data) at a high speed by using vector instructions at a time. Since it is preferable to replace the source program written in a high-level compilation language (for example, FORTRAN) and generate an object program for the vector processor, the program is analyzed and replaced with vector instructions (vector )). The part that can be vectorized is
It is expressed as a loop structure in the source program.

In addition to simple vector operations, even when parallel execution has conflicting data dependencies, special vector instructions applicable to regression type operations such as inner product / summation or recurrence expressions for scalar variables are used. A vector processing processor is provided.

FIG. 4 shows an example of a source program input to the compiler. In a conventional compiler, this source program is analyzed and the first structure having the structure shown in FIG.
The intermediate text is generated, and then the second intermediate text in which a part of the structure shown in FIG. 6 is transformed into a sentence suitable for vectorization is generated, and this is coded to generate the target program. It was That is, the source program shown in FIG. 4 sets the initial value of the scalar variable Q to "1.0",
This shows the process of sequentially adding the value of array element A (I) from the 1st element to the 100th element, and the final result cannot be obtained unless sequential addition is performed, so for the scalar variable Q Although there is a data dependency that contradicts parallel execution, since the vector instruction of the regression type calculation of the sum for the scalar variable Q described above can be applied, it is possible to resolve the conflict of the data dependency and perform vectorization. It was

[Problems to be Solved by the Invention]

As mentioned above, in the conventional compiler, vectorization is possible for regression-type operations such as inner product / summation or recurrence formulas for scalar variables even when there are conflicting data dependencies in parallel execution. However, vectorization is possible only when the regression operation is used independently, and when the value on the left side of the regression operation statement is referenced later in the same loop, the vector There was a drawback that it could not be realized.

That is, when a source program as shown in FIG. 7 is given, the value (evaluation result) for each loop of Q = Q + A (I) is referred to by C (I) = Q + B (I) in the same loop. Therefore, since it is necessary to separately store the value for each loop for the scalar variable Q, the vector instruction corresponding to the regression type operation of the summation for the scalar variable Q cannot be applied, and therefore the scalar variable Q appears. It was forced to process the existing range as a group of statements that cannot be vectorized and to execute only this part sequentially by a scalar instruction.

In a vector processor having a vector operation function, generally, the larger the ratio of the portion of the target program executed by the vector instruction, the shorter the execution time of the program. However, the conventional technique has a drawback in that the performance of the vector processor cannot be fully exhibited.

In addition, since only a group of statements that cannot be vectorized are executed with scalar instructions, the compiler may have to internally divide one loop structure into several loop structures in the source program, increasing the compilation time. There was also the drawback of doing so.

The present invention has been proposed in view of the above points, and an object thereof is to refer to a scalar variable on the left side of a regression operation statement in the same loop after the regression operation statement. It is to provide a vectorization processing method capable of vectorization even in some cases.

[Means for Solving the Problems]

In order to achieve the above object, the present invention inputs a source program written in a high-level language to a vector processor having vector instructions applicable to regression type operations such as inner product / summation or recurrence expressions for scalar variables. In the vectorization processing method of the compiler that generates the target program, it is activated when the loop structure is analyzed and there is a conflicting data dependency in parallel execution, and the structure of the expression is examined to identify the inner product / sum of scalar variables. The recurrence type operation pattern recognition unit that detects the sentence of the regression type operation with the pattern of, and the definition in the loop for the scalar variable of the detected sentence
Definition / reference position relationship investigation section that investigates the reference position relationship, and if the result of the investigation is that there is a reference to a scalar variable on the left side of the regression operation statement after the regression operation statement, that scalar variable is On the other hand, a regression type operation processing unit having a regression type operation vectorization processing unit that allocates a vector work area and transforms a recurrence type regression type operation into an applicable sentence is provided.

[Action]

In the vectorization processing method of the present invention, when the loop structure is analyzed during compilation and parallel execution has a conflicting data dependency, the regression-type arithmetic processing unit is activated, and the regression-type arithmetic processing unit performs the regression. The pattern operation pattern recognition unit examines the structure of the expression and detects a regression type operation sentence with a specific pattern of inner product and sum for the scalar variable, and the definition / reference positional relationship investigation unit detects it as a scalar variable of the detected sentence. In contrast, if the definition / reference position relationship in the loop is investigated, and as a result of the investigation, there is a reference to the scalar variable on the left side of the regression operation statement after the regression operation statement, the regression operation vector The conversion processing unit allocates a vector work area to the scalar variable and transforms it into a sentence to which the recurrence-type regression operation can be applied.

〔Example〕

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

FIG. 1 is a block diagram showing an embodiment of a compiler to which the vectorization processing method of the present invention is applied. In FIG. 1, a compiler 2 reads a source program 1 written in a high-level language, performs a syntax analysis to generate a first intermediate text 24, and a first intermediate text 24, and reads the source program 1. The vectorization unit 22 that detects the loop structure in 1 to recognize and process the vectorizable portion and generates the second intermediate text 25, and the second intermediate text 25 are read and coded to obtain the target program 3 And a code generation unit 23 for generating and outputting. Further, the vectorization unit 22 includes a loop structure analysis unit 221 that collects information for the array element of each statement in the loop structure, a scalar variable, and the like to examine a data dependency relationship that contradicts parallel execution, and a loop structure analysis unit. There is a data dependency relationship that contradicts parallel execution by the data dependency relationship determination unit 222 and the data dependency relationship determination unit 222 that determines whether there is a data dependency relationship that conflicts with parallel execution based on the information obtained by the unit 221. A regression type arithmetic processing unit 223 that performs transformation suitable for vectorization when possible for a sentence group that is determined to be true, and a vector text generation unit that generates a second intermediate text including a sentence corresponding to a vector instruction to be applied 224 and. Further, the regression type arithmetic processing unit 223 examines the structure of the expression for the sentence group that is determined to have the data dependency that contradicts the parallel execution, and detects the regression type arithmetic operation having the specific pattern of the inner product / sum for the scalar variable. Regression type calculation pattern recognition unit 2231 and a loop structure analysis unit 221 for the scalar variable of the sentence detected by the regression type calculation pattern recognition unit 2231.
The definition / reference position relationship investigation unit 2232 that investigates the definition / reference position relationship in the loop based on the information obtained in When there is a reference to a scalar variable on the left side of a type operation statement, a vector work area of a one-dimensional array element is assigned to that scalar variable and the regression type operation of recurrence formula is transformed into an applicable statement. Regression type calculation vectorization processing unit
It is composed of 2233 and.

The operation of the above embodiment will be described below. The source program shown in FIG. 7, which could not be vectorized by the conventional compiler, is used as a specific example.

First, the syntactic analysis unit 21 reads the source program 1,
The first intermediate text 24 is generated. That is, the first intermediate text having the structure shown in FIG. 8 is generated for the source program shown in FIG.

Next, the vectorization unit 22 reads the first intermediate text 24, and the loop structure analysis unit 221 reads the first intermediate text 24.
The loop structure is detected from this, and information is collected for each statement in the loop structure to investigate the data dependency that contradicts the parallel execution.

Next, the data dependency relationship determination unit 222 determines whether or not there is a data dependency relationship that contradicts parallel execution, based on the information obtained by the loop structure analysis unit 221. For the source program in Figure 7 (the first intermediate text in Figure 8),
Assignment statement Q = Q + A (I) It is determined that there is a data dependency relationship inconsistent with parallel execution for the scalar variable Q on the left side. Then, when it is determined that there is a data dependency that contradicts parallel execution, control is passed to the regression-type arithmetic processing unit 223.

In the regression-type calculation processing unit 223, the regression-type calculation pattern recognition unit 2231 examines the structure of the expression and detects a regression-type calculation statement having a specific pattern of inner product / sum for scalar variables.

Next, the definition / reference positional relationship investigation unit 2232, based on the information obtained by the loop structure analysis unit 221, in the loop for the scalar variable Q on the left side of the sentence detected by the regression type operation pattern recognition unit 2231. Investigate the definition and reference position relationship of. At this time, for the source program of FIG. 7, after the regression type operation statement Q = Q + A (I), C (I) = Q + B (for the scalar variable Q on the left side of the regression type operation statement. It turns out that there is a reference in I).

The regression type operation vectorization processing unit 2233 allocates a vector work area to the scalar variable and transforms it into a sentence to which the recurrence type regression type operation can be applied. The following processing is specifically performed on the source program shown in FIG.

1 using vector work area for scalar variable Q
A process of providing a temporary array element Q '(I) of dimension is performed. At this time, the size of the array is the vector length + 1 element.

Immediately before the vector instruction corresponding to the recurrence-type regression operation is executed, the initial value of the scalar variable Q is set to the temporary array element Q ′.
Processing for generating a sentence to be stored in the first element of (I) is performed.

A temporary array element Q ′ (I) is used instead of the scalar variable Q in the assignment statement of the sum total regression type operation (the left side is Q ′ (I +
1), Q ′ (I)) on the right side, and then a process for applying a vector instruction corresponding to the recurrence type regression operation is performed.

After the sum total regression type assignment statement, processing for using the temporary array element Q ′ (I) is performed on the statement that refers to the scalar variable Q. At this time, the temporary array element Q ′
(I) is referred to from the second element Q ′ (I + 1).

FIG. 2 shows the concept of the program modified by the processing of the regression calculation vectorization processing unit 2233 described above.

After that, the vector text generation unit 224 generates the second intermediate text 25 based on the processing result of the regression calculation vectorization processing unit 2233. A second intermediate text having a structure as shown in FIG. 3 is generated for the program of FIG.

Finally, the code generation unit 23 uses the vector text generation unit 224.
The second intermediate text 25 generated in step 2 is read, coded to generate the target program 3, output is performed, and the processing of the compiler 2 ends.

〔The invention's effect〕

As described above, in the vectorization processing method of the present invention, since it is possible to vectorize a sentence group that was conventionally impossible to vectorize, in the vector processing processor,
There is an effect that the execution speed of the program can be improved. Further, since the processing for dividing the non-vectorizable portion is also reduced, there is an effect that the compilation time can be shortened.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an embodiment of a compiler to which the vectorization processing method of the present invention is applied, and FIG. 2 is a conceptual diagram of a program modified by a regression type operation vectorization processing unit in the embodiment of FIG. , FIG. 3 is a conceptual diagram of a second intermediate text corresponding to the program of FIG. 2, FIG. 4 is a diagram showing an example of a source program, and FIG. 5 is a first intermediate text corresponding to the source program of FIG. 6 is a conceptual diagram of a second intermediate text corresponding to the first intermediate text of FIG. 5, FIG. 7 is a diagram showing an example of a source program, and FIG. 8 is a source program of FIG. It is a conceptual diagram of the 1st intermediate text corresponding to. In the figure, 1 ... Source program 2 ... Compiler 21 ... Syntax analysis unit 22 ... Vectorization unit 221 ... Loop structure analysis unit 222 ... Data dependency relationship determination unit 223 ... Regression type arithmetic processing unit 2231 ... Recurrence-type operation pattern recognition unit 2232 …… Definition / reference position relationship investigation unit 2233 …… Regression-type operation vectorization processing unit 224 …… Vector text generation unit 23 …… Code generation unit 24 …… First intermediate text 25 …… No. 2 Intermediate text 3 ... Objective program

Claims (1)

[Claims] 1. A target program is generated by inputting a source program described in a high-level language to a vector processor having a vector instruction applicable to regression type operations such as inner product / summation or recurrence expression for scalar variables. In the vectorization processing method of the compiler, it is activated when the loop structure is analyzed and there is a conflicting data dependency in parallel execution, and the structure of the expression is examined and regression with a specific pattern of inner product / sum for scalar variables Regression type operation pattern recognition part to detect the sentence of type operation and definition in the loop for the scalar variable of the detected sentence
Definition / reference position relationship investigation section that investigates the reference position relationship, and if the result of the investigation is that there is a reference to a scalar variable on the left side of the regression operation statement after the regression operation statement, that scalar variable is On the other hand, a vectorization processing method characterized by comprising a regression-type operation processing section having a regression-type operation vectorization processing section for allocating a vector work area to transform a recursive regression-type operation into an applicable sentence .