JPH06290206A - Vectorization systen - Google Patents

Abstract

PURPOSE:To shorten the execution time of a program in a vector processor by enabling the vectorization of a loop including a variable which not conventionally vectrized and quoted before definition. CONSTITUTION:A condition decision part 321 in a vectorization part 32 decides whether or not the loop including the variable quoted before the definition satisfies conditions that are executed in parallel. A loop constitution part 322 generates a work array for containing values that the variable quoted before the definition have at the time of respective rounds of the loop which is decided by the condition decision part 321 so that the parallel execution can be performed and includes the variable quoted before definition and substitutes the quotation of the values in the work array for the quotation of the variable in the loop.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vectorization conversion system for converting a source program into an object program that can be executed in parallel.

[0002]

2. Description of the Related Art In the conventional technique for vectorizing a loop of a source program, a first intermediate text showing the result of syntax analysis shown in FIG. 7 is first generated for the source program shown in FIG. To be done. Next, the parallel executable part is recognized from the first intermediate text. Then, the second intermediate text illustrated in FIG. 8 is generated based on the first intermediate text for the parallel executable part. Then, the sequence of generating the target program from the second intermediate text is performed. The recognition of whether or not they can be executed in parallel is that the variables in the second intermediate text do not conflict with the dependency between their definition and the order in which they are quoted. Further, in the source program illustrated in FIG. 4, since the dependency relation between the variable definition and the citation order is inconsistent, it cannot be vectorized by the conventional method.

However, even in the case of a loop including a variable that is quoted before definition, vectorization can be performed even in the conventional technique when the variable is an index variable or included in a macro operation such as inner product and summation. it can.

[0004]

In general, a vector processor having a vector operation function can shorten the execution time of the target program by increasing the ratio of the portion of the target program executed by the vector instruction. it can. However, in the above-mentioned conventional technique, the loop including the variable quoted before the definition cannot be vectorized and executed by the scalar instruction, so that the execution time cannot be shortened.

An object of the present invention is to convert a vector containing a variable referred to before definition into a vector which can be executed in parallel when a certain condition is satisfied.

[0006]

A vectorizing conversion system for converting a source program described in a high-level language of the present invention into an object program executed by a vector processor having a vector operation function, reads the source program and constructs it. Whether the parsing means for parsing and generating a first intermediate text indicating the analysis result and a loop including a variable quoted before the definition in the first intermediate text satisfy a condition capable of parallel execution Generating a working array in which the variable referred to before the definition of the condition determining means and the loop stores a value for each turn of the loop,
Vectorizing means having loop intermediate means for generating second intermediate text for replacing the variable citations in the loop with the value quotations of the working array; and the object based on the second intermediate text. Compiling means including a code generating means for generating a program.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the present invention will be described with reference to the drawings.

Referring to FIG. 1 showing the system configuration of an embodiment of the present invention, in the vectorization conversion system 1, a compiler section 3 for reading a source program 2 and a first intermediate text 34 generated by the compiler section 3 are included. And the second
Intermediate text 35 and the target program 4 which is a compile output. Further, the compiler unit 3 reads a source program written in a high-level language, performs a syntax analysis, and generates a first intermediate text 34 indicating the analysis result. A vectorization unit 32 for recognizing parallel executable parts and generating a second intermediate text 35 for vectorizing the loop.
And a code generator 33 that generates and outputs the object program 4 from the second intermediate text 35. Furthermore, the vectorization unit 3 further includes a condition determination unit 321 that determines whether or not a parallel execution condition is satisfied for a loop including a variable that is quoted before the definition, and the condition determination unit 3 described above.
For a loop including a variable that is referred to before the definition that is determined to be parallel executable in step 21, the variable that is referred to before the definition is converted into a working array that stores a value for each cycle of the loop, and A loop construct 322 for generating a second intermediate text that vectorizes the loop containing the variable referenced before definition by replacing the variable reference with the value of the working array.

In this embodiment, when the source program illustrated in FIG. 4 is given, the variable referred to before the definition creates a working array for storing the value for each turn of the loop, and this variable in the loop is generated. Convert to a vector that can be executed in parallel by replacing the quotes of with the quotes of the values in the working array. In the source program of FIG. 4, the working array RW illustrated in FIG. 5 is generated, and the variable R in the original loop generates a text portion in which the value for each cycle of the loop is stored in the working array RW. Before defining it in the original loop by transforming the original loop into a text description in which the variable R is excluded from the loop and the reference of the variable R is replaced with the reference of the value of the working array RW. The conflicting dependency of the quoted variable R on the parallel execution is resolved to make the vector executable in parallel.

The first intermediate text illustrated in FIG. 2 is obtained by the syntax analysis unit 31 from the source program illustrated in FIG. 4 by the syntax analysis unit 3 and input to the vectorization unit 32, and the condition determination unit 321 defines the definition. When recognizing whether or not there is a loop including the variable quoted before, the condition determination unit 3
21 determines whether or not the following conditions (1) to (5) are satisfied. (1) The initial value, the final value, and the increment value of the DO variable of the DO statement that constitutes the loop are all constants, and the increment value is a positive value. In the source program illustrated in FIG. 4, since the initial value of the DO variable I is “1”, the final value is “10”, and the increment value is “1”, this condition is satisfied. (2) For variables that are referenced before the definition, the definition in the loop must be made every time the loop is repeated. That is, there is no specific condition for this definition. In the source program illustrated in FIG. 4, the variable R quoted before the definition satisfies this condition because it is always executed each time the loop is executed. (3) For variables that are quoted before the definition, the definition in the loop must be made in only one place in the assignment statement. In the source program illustrated in FIG. 4, the variable R quoted before the definition
Is defined in only one place in the loop, which satisfies this condition. (4) Regarding the variables quoted before the definition, the expression on the right side of the defining statement has no dependency relationship with the elements of other statements in the loop. In the source program illustrated in FIG. 4,
Since the array element B (I) on the right side of the statement defining the variable R quoted before the definition does not appear in another statement in the loop, it has no dependency and satisfies this condition. (5) For variables that are quoted before definition, the citation is defined only before the statement that defines the variable. In the source program illustrated in FIG. 4, the variable R quoted before the definition
Quotation is made only before the definition and satisfies this condition.

When all of the above conditions (1) to (5) are satisfied, it is determined that the loop including the variable quoted before the definition can be executed in parallel.

The loop constructing unit 322 performs the following operation for the loop including the variable that is referred to before the definition, which is determined by the condition determining unit 321 to be executable in parallel.

First, the loop construction unit 322 creates a working array for the variables quoted before the definition. The working array of this is quoted before defining the number of dimensions as "1", the lower limit of the dimension as "initial value of DO variable", the upper limit of dimension as "final value of DO variable + increment value of DO variable" and data type. It has the same data type as the variable. In the source program illustrated in FIG. 4, a real number type array RW having a lower limit of size “1” and an upper limit of size “11” is generated. Next, the loop configuration unit 322 includes an assignment statement for storing the value of the variable quoted before the definition at the time of the first iteration of the original loop into the working array.
Generates the intermediate text of. The right side of the generated assignment statement is the working array whose subscript is the initial value of the DO variable, and the left side is the variable that is referenced before the definition in the original loop.

In the source program illustrated in FIG. 4, the second intermediate text illustrated in FIGS. 3 and 5 including the assignment statement RW (1) = R is generated.

Further, the loop construction unit 322 includes a second loop including a loop for storing the value of the variable quoted before the definition in the second and subsequent iterations of the original loop in the working array.
Generates the intermediate text of. The DO variable of the generated loop and the initial value, the final value, and the increment value of the DO variable are the same as those of the original loop, and the assignment statement in the loop has "DO variable +
It is a working array having a subscript of "DO variable increment value", and the left side is the same as the left side of the assignment statement in which the variable quoted before the definition in the original loop is defined. In the source program illustrated in FIG. 4, a second intermediate text that executes the assignment statement WR (I + 1) = B (I) in the loop illustrated in FIGS. 3 and 5 by a vector instruction is generated.

Then, except for the statement in which the variable that is referenced before the definition is defined from the original loop, the reference of the variable that is referenced before the definition is replaced with the reference of the working array.
Generates the intermediate text of. The subscript of its working array is the DO variable of the original loop. In the source program illustrated in FIG. 4, the statement R = B (I) defining the variable R in the original loop is removed from the loop, and the variable R is quoted A (I) = R 3 and the example illustrated in FIG. 5 A second statement in which a statement replacing A (I) = RW (I) is executed by a vector instruction
Generates the intermediate text of. Then, further, an assignment statement for guaranteeing the final value of the variable quoted before the definition in the original loop is generated in the second intermediate text. The left side of this assignment statement is a variable that is quoted before definition, and the right side is a working array whose subscript is "DO variable end value + DO variable increment value". In the source program illustrated in FIG. 4, the second intermediate text including the assignment statement R = RW (11) illustrated in FIGS. 3 and 5 is generated.

[0017]

As described above, according to the present invention, it is possible to vectorize a loop including a variable that is quoted before the definition that cannot be vectorized conventionally. Thereby, the execution time of the target program can be shortened in the vector processor.

[Brief description of drawings]

FIG. 1 is a diagram showing a system configuration of an embodiment of the present invention.

FIG. 2 is a diagram illustrating a first intermediate text of FIG.

FIG. 3 is a diagram illustrating a second intermediate text of FIG.

FIG. 4 is a diagram illustrating a source program that cites variables before definition.

FIG. 5 is a diagram showing an example of vectorization in which the source program of FIG. 4 is modified and which can be executed in parallel.

FIG. 6 is a diagram illustrating a source program that can be executed in parallel in a conventional technique.

FIG. 7 is a diagram illustrating a first intermediate text for the source program of FIG.

FIG. 8 shows a second intermediate text for the source program of FIG.

[Explanation of symbols]

 1 Vectorization conversion system 2 Source program 3 Compiler 4 Object program 31 Syntax analysis unit 32 Vectorization unit 33 Code generation unit 34 First intermediate text 35 Second intermediate text 321 Condition determination unit 322 Loop configuration unit

Claims (1)

[Claims] 1. A vectorization conversion system for converting a source program written in a high-level language into an object program executed by a vector processor having a vector operation function, the source program being read, syntax-analyzed, and an analysis result being shown. Syntactic analysis means for generating a first intermediate text, and condition determination means for determining whether or not a loop including a variable quoted before definition in the first intermediate text satisfies a condition for parallel execution The variable referenced before defining the loop creates a working array that stores a value for each iteration of the loop, and replaces the variable references in the loop with the value references of the working array. A vectorizing means having a loop forming means for generating a second intermediate text; and the object program based on the second intermediate text. A vectorization conversion system comprising: a code generation unit that generates a.