JPH0417469B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application of the Invention The present invention relates to a vectorization processing method.

Conventional technology In recent years, in order to make vector processors using parallel pipeline processing realize their true high speed as ultra-high-speed computers for scientific and technical calculations, vector processors that can achieve as high a vectorization rate as possible have been developed. A compiler is required.

Problems with conventional technology Conventionally, if a loop containing a function reference exists in a source program written in a compiler language,
Vectorization of the entire loop becomes impossible except in the following cases, which is one of the causes of a decrease in the vectorization rate.

(1) When the processing procedure of the function can be expanded into the loop processing procedure in a vector-processable format.

(2) The system side has a function processing procedure (vector version of the function) that passes a set of argument data corresponding to the length of the vector and the corresponding function value to the function via a vector register or memory. If available.

(3) When a vector version of the function described in (2) above is prepared by the user and this is known to the compiler.

Loops that include references to functions defined by users can be vectorized by satisfying the condition (3) above. However, in order to realize this condition, it is often necessary to write function processing procedures in assembly language, which poses a problem of placing an excessive burden on the user.

Means for Solving the Problems in the Structure of the Invention The present invention for solving the problems in the prior art described above repeatedly calls a scalar function specified from the outside with elements in an input argument data vector array, and performs corresponding operations. A system that has a relay vectorization function that generates a function value vector by repeatedly calling a process that receives a scalar function: For loop parts that include function references in a source program written in a high-level language, and a compiler that generates a target program, including the step of calling the relay vectorization function at the step of calling the relay vectorization function.

The effects of the present invention will be explained in detail below using examples.

Embodiment FIG. 1 is a functional block diagram of a compiler used in an embodiment of the present invention.

The source analysis unit 21 in the compiler 2 analyzes the supplied source program, passes the vectorization processing part to the vectorization processing unit 22, and passes the other part to the non-vectorization processing unit 27. At this time, the source analysis unit 21 converts the vectorization processing part that includes a function quotation in the loop into the first
It is passed to the vectorization processing unit 26. On the other hand, the source analysis unit 21 performs vectorization processing using the conventional method due to reasons such as parts in which there are no function citations in the vectorization processing portion, or vector version functions are defined even if there are function citations. The portion to be processed is passed to the second vectorization processing unit 25.

FIG. 2 shows an example of a program portion including a function reference in the above-mentioned loop when FORTPAN is used as the source programming language. This program part (so-called DO loop) says, ``For the i-th elements Ai, Bi, Xi, Yi, and Zi of arrays A, B, X, Y, and Z, calculate the product of Ai and Bi, and use this product as an argument. Call the function FUNC defined by the user as data to obtain the function value corresponding to this argument data, calculate the product of the obtained function value and Zi, add Yi to this product, and add the result of this addition.
"Storing to Xi" process starting from the first element.
The processing procedure is described to repeat until the 200th element.

A first vectorization processing unit 26 that receives a program portion including a function reference in a loop as described above.
is the intermediate text 3 to vectorize this
generate.

FIG. 3 shows an example of intermediate text 3 generated based on the DO loop shown in FIG. In this intermediate text 3, V _teo is a register or memory that stores the length of a vector, and vt1 to vt8 are registers or memories that hold vector elements.

In step (1) of intermediate text 3, the vector length is set. In step (2) register vt
1 stores the elements of array A, in step (3) the elements of array B are stored in register vt1, and in step (4) the product between each element in registers vt1 and vt2 is calculated. In the next step (5), the relay vector function ("VECT") configured as shown in Figure 4 is called with the function name FUNC and the contents of register vt3 as arguments, and the relay vectorization function
Function value for vt3 received from VECT is registered
Stored in vt4.

In the above relay vectorization function VECT, a ₁ , a ₂ ...a _o in the argument data vector (array) 12
are the argument data (Ai and Bi) received one after another from the register vt3 in FIG. 3, respectively. The function calling unit 13 repeatedly calls the scalar function specified by the specified function name FUNC using each element ai of the argument data vector, and receives the corresponding scalar function value fi. This scalar function value f1, f
2...Function value vector (array) 1 with fn as an element
4 is generated.

Returning again to the intermediate text in FIG. 3, the function value vectors f1, f2, . next step
In step (6), each element of array Z is stored in register vt5, and in step (7), the product of the contents of register vt4 and register vt5 is calculated, and the result of this calculation is stored in register vt6. In step (8), the elements of array Y are stored in register vt7, and in step (9), addition is performed between each element of vt7 and vt6, and the result of this addition is stored in register vt8. Finally, the result of the operation between the arrays, which is the contents of the register vt8, is stored in X.

The intermediate text 3 as described above is subjected to well-known optimization processing in the intermediate text optimization section 23, and then passed to the target program generation section 24,
A target program 4 is generated. This object program is executed in a processing system that includes a vector calculation function, a relay vectorization function, and scalar functions such as subroutines called from the relay vectorization function.

The relay vectorization function described above is extremely simple, as shown in FIG. 4, and can be easily constructed.

Also, instead of calling the relay vector function, it may be expanded into the intermediate text shown in FIG. 3.

Note that the scalar function called from the relay vectorization function does not need to be dedicated to vectorization processing, and may be a normal program written in a compiler language.

Although we have shown an example of a configuration in which the source analysis section of the compiler detects loops that include function references, it is also possible to have a configuration in which the source analysis section only detects loops and the vectorization processing section detects whether or not there are function references in these loops. .

Effects of the Invention As described in detail above, the present invention adds a relay vectorization function that generates a function value vector by repeatedly calling a scalar function processing function for each component of vector argument data received by the system, and For the function reference loop part in a source program written in a high-level language,
Since the configuration includes a compiler that generates a target program so as to include a step of calling the relay vectorization function at the time of program execution,
The result is the same as if a general-purpose vector version function were prepared, and the vectorization rate is significantly improved.

[Brief explanation of drawings]

FIG. 1 is a functional block diagram of a compiler used in an embodiment of the present invention, FIG. 2 is a diagram showing an example of a loop containing a function reference written in a high-level language, and FIG.
The figure shows an example of intermediate text generated by the vectorization processing section of the compiler, and FIG. 4 is a diagram showing an example of the configuration of the relay vectorization function. 1... Source program, 2... Compiler, 3... Intermediate text, 4... Target program, 21... Source analysis section, 22, 25, 26...
...Vectorization processing unit, 23...Intermediate text optimization unit, 24...Target program generation unit, 27...
Non-vectorization processing section.

Claims (1)

[Scope of Claims] 1. A scalar function specified from the outside is repeatedly called with an element in an input argument data vector array, and a corresponding scalar function value array is created by repeatedly calling a process that receives the corresponding scalar function. For systems that have a relay vectorization function that generates function value vectors consisting of
1. A vectorization processing method, comprising: a compiler that generates a target program so as to include the step of calling the relay vectorization function during program execution.