JPH03161837A - Multiple do loop operation acceleration source program forming system - Google Patents

Abstract

PURPOSE:To accelerate operation by executing optimization so that the loop length of the innermost DO loop is maximized and the loop length of the inside DO loop is longer than that of an outside DO loop. CONSTITUTION:A loop length sorting part 5 sorts the loop lengths of respective DO loops from evaluation based upon a loop length evaluating part 4 and a control variable change range changing part 6 replaces the change range of control variable among respective DO loops so that the loop length of the inner most DO loop out of a multiple DO loop extracted from a multiple DO loop extracting part 1 is maximized and the loop lengths of the DO loops are gradu ally reduced in the outside direction. An array element changing part 7 changes the storage positions of array element values of an array having control variables as subscripts in the multiple DO loop, an array declaration changing part 8 changes the array declaration description of the array and a DO loop information output part 9 allocates an identification number to the multiple DO loop and outputs a DO loop information list and a changed source program forming part 10 forms and outputs a changed source program 13. Thus, opera tion can be accelerated.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a method for generating a FORTRAN source program for a hector computer including multiple DO loop operations,
In particular, the present invention relates to a method for generating a source program for accelerating multiple DO loop operations (a FORTRAN source program in which the speed of multiple DO loop operations has been achieved) for speeding up multiple DO loop operations included in the FORTRAN source program.

[Conventional technology]

In order to efficiently (at high speed) execute operations expressed by a FORTRAN source program (for example, numerical simulation using the finite difference method), FOl'? T.R.
It is essential to speed up DO loop operations that frequently appear in AN source programs.

In order to speed up DO loop operations, multiple DO
It is important to perform optimization to make the loop length of the inner DO loop larger than the loop length of the outer DO loop.

Conventionally, this type of optimization was performed for F O R i”R
An AN source program (multiple DO loop calculation acceleration source program) is generated by artificially replacing the DO loops in the multiple DO loops to suit the optimization described above. (Although optimization has been carried out, the scope of its application is very limited, and it is only an optimization for object I programs and not for the source program itself).

[Problem to be solved by the invention]

In the conventional technology described above, in order to generate a source program for accelerating multiple DO loop calculations, it is necessary to artificially replace DO loops in multiple DO loops and change the storage positions of array element values based on this. Therefore, in order to speed up multiple DO loop operations, complicated human operations are required, which increases the burden on programmers. The disadvantage is that it is necessary to give up on

In view of the above-mentioned points, an object of the present invention is to be able to automatically perform the above-mentioned optimization for generating a source program for accelerating multiple DO loop calculations, and to eliminate complicated operations such as replacing DO loops manually. An object of the present invention is to provide a source program generation method for accelerating multiple DO loop operations, which enables high speed multiple DO loop operations (achieving efficient operation of a hector computer).

[Means to solve the problem]

In the multiple DO loop calculation speed-up source program generation method of the present invention, the variable name of the first subscript of the array included in the innermost DO loop matches the control variable name of the innermost DO loop, and the control variable of each DO loop is Initial value indicating the range of change. In the generation of a FORTRAN source program including multiple DO loops whose final value and increment value are expressed by constants, a multiple DO loop extraction unit extracts multiple DO loops in the source program before change, and extraction by this multiple DO loop extraction unit is performed. D
an O-loop control variable extraction section; a control variable dependent array extraction section that extracts an array dependent on the control variables extracted by the DO loop control variable extraction section from within the multiple DO loops extracted by the multiple DO loop extraction section; , a loop length evaluation section that calculates the loop length of each DO loop based on the change range of the control variable of each DO loop extracted by the DO loop control variable extraction section; and a loop length sorting unit that sorts the loop lengths of the loop length sorting unit, and a loop length of the innermost DO loop of the multiple DO loops extracted by the multiple DO loop extraction unit based on the sorting result by the loop length sorting unit. A control variable change range changing unit swaps the change range of the control variable between each DO loop 1 so that the loop length decreases as the outer DO loop becomes closer. An array element changing unit that changes the storage position of the value of an array element of an array having a control variable as a subscript for the multiple DO loop according to the permutation, and the storage position of the value of the array element is changed by the array element changing unit. an array declaration change unit that changes the array declaration description of the extracted array; and a DO loop information output unit that assigns an identification number to the multiple DO loop extracted by the multiple DO loop extraction unit and outputs a DO loop information list. , a modified source program that generates and outputs a modified source program that is a multiple DO loop calculation acceleration source program based on the contents modified by the control variable change range changing unit, the array element changing unit, and the array declaration changing unit; It has a raw bottom.

[Effect]

In the multiple DO loop calculation speed-up source program generation method of the present invention, the multiple DO loop extraction section extracts multiple DO loops in the source program before change, and the DO loop control variable extraction section is extracted by the multiple DO loop extraction section. The control variable of each DO loop in the multiple DO loop is extracted, the change range of the control variable is detected, and the control variable dependent array extraction unit extracts the DO loop control variable from the multiple DO loop extracted by the multiple DO loop extraction unit. The loop length evaluation section calculates the loop length of each DO loop based on the change range of the control variable of each DO loop extracted by the DO loop control variable extraction section. , the loop length sorting section sorts the loop lengths of each DO loop based on the evaluation by the loop length evaluation section, and the control variable change range changing section extracts the loop lengths using the multiple DO loop extraction section based on the results of sorting by the loop length sorting section. multiplex D
Regarding the O-loop, the change range of the control variable is swapped between each DO loop so that the loop length of the innermost DO loop is maximum and the loop length decreases for the outer DO loops, and the array element change section changes the control variable change range change section. For a multiple DO loop in which the range of change of a control variable has been swapped, the storage position of the array element value of the array that has the control variable as a subscript is changed in accordance with the swap, and the array declaration changing section uses the array element changing section to change the storage position of the array element value of the array. The array declaration description of the array whose element value storage position has been changed is changed, and the DO loop information output section assigns an identification number to the multiple DO loop extracted by the multiple DO loop extraction section and generates the DO loop information list. Output and generate a modified source program based on the contents changed by the control variable change range changing section, array element changing section, and array declaration changing section, which is a source program for accelerating multiple DO loop calculations. and output.

〔Example〕

Next, the present invention will be explained with reference to the drawings.

FIG. 1 is a block diagram showing the structure of an embodiment of a source program generation method for accelerating multiple DO loop calculations according to the present invention. The multiple DO loop calculation speed-up source program generation method of this embodiment is based on a multiple DO loop extraction unit 1 that extracts multiple DO loops from a source program 11 before change.
(If a plurality of multiple DO loops exist, all multiple DO loops are extracted), and the multiple DO loop extraction unit 1 extracts the control variables of each DO loop in the extracted multiple DO loops. A DO loop control variable extraction unit 2 detects the change range (initial value, final value, and incremental value), and extracts an array (an array dependent on the control variable) that includes the DO loop control variable as a subscript in each multiple DO loop. Control variable dependent array extraction unit 3 and each D in the multiple DO loop
A loop length evaluation unit 4 calculates the loop length of each DO loop based on the change range of the control variable of the O-loop, and the loop length of each DO loop in each multiple DO loop is calculated in descending order based on the evaluation by the loop length evaluation unit 4. The loop length sorting section 5 performs sorting, and the control variable is set so that the loop length of the innermost DO loop becomes maximum for each multiple DO loop based on the sorting result by the loop length sorting section 5, and the loop length decreases toward the outer DO loop. The control variable change range change unit 6 switches the change range (initial value, final value, and incremental value) between each DO loop in the multiple DO loop, and the control variable change range change unit 6 changes the change range of the control variable into each DO loop. Array element change that changes the storage position of the value of an array element of an array that has the control variable to be swapped as a subscript for multiple DO loops swapped between DO loops in accordance with the swap (swap regarding the control variable that is the subscript) 7, the control variable change range changing unit 6, and the array element changing unit 7 to reflect the changes and reset the array size as necessary.
An identification number is assigned to each multiple DO loop extracted by the array declaration modification unit 8 that modifies the array declaration description using the ENSION statement and the multiple DO loop extraction unit 1, and the DO loop information list 12 is output to file 1. DO loop information output unit 9 and control variable change range changing unit 6. A modified source program generation section 10 generates a modified source program 13 based on the contents modified by the array element modification section 7 and the array declaration modification section 8 and outputs it to a file, and one or more multiple DO loops (containing array (including multiple DO loops) in which the variable name of the first subscript matches the control variable name of the innermost DO loop in the multiple DO loop, and the initial value, final value, and increment value of the control variable are expressed only by constants. FO R TR AN
Source program before change 11 which is a source program
(original source program) and the control of each DO loop within the multiple DO loop for each multiple DO loop.
A DO loop information list 12 showing variable names, change ranges (initial values, final values, and incremental values) of control variables, loop lengths (vector lengths), etc. regarding the source program before change 111 and the source program after change 13, and multiple DO loops. The modified source program 13 (new FORTRAN source program 12), which is a calculation speed-up source program, is included.

Next, the operation of the multiple DO rube calculation high-speed source program generation system of this embodiment configured as described above will be explained. In addition, here, FORTR as shown in Figure 2 is used.
AN source program (FORT including triple DO loop)
RAN source program) is the source program before change
The operation when the value is 1 will be explained.

Regarding the triple DO loop in the FORTRAN source program (source program 11 before change) shown in FIG. 2, the loop length of each DO loop is 120.50 and 20 in order from the outermost DO loop, and Since the loop length of is the minimum, it is inefficient from the point of view of effective utilization of the vector computer's capacity. Therefore, the FORT that can speed up multiple DO loop operations (in this case, triple DO loop operations) can be achieved by following the steps shown below.
RAN source program (source program 13 after modification)
) will be produced.

First, the multiple DO loop extraction unit 1 extracts multiple DO loops from the entire pre-change source program 11,
A triple DO loop consisting of DO loops with statement numbers 40, 50 and 6o is extracted.

Next, the DO loop control variable extraction unit 2 extracts the multiple D
The control variables K, J, and I of each DO loop in the O loop (triple DO loop) are extracted, and the initial values of the control variables are used to define the range of change of these control variables. A closing value and an increment value are detected for each DO rube.

Furthermore, the control variable dependent array extraction unit 3 extracts the multiplex D
In the range of the O-loop, an array (here, array DF) that depends on the control variable is extracted.

Changing the order of control variables is achieved as shown below.

First, the loop length evaluation unit 4 determines the initial value and final value of the control variable of each DO loop for each DO rule-7'' (DO loops with statement numbers of 40.50 and 60) in the multiple DO roof. and the increment value, the loop lengths (vector lengths) are calculated.Here, it can be seen that each loop length is 120.50 and 20 in order from the loop length of the outer DO loop.

Next, the loop length sorting unit 5 sorts the DO loops in the multiplex DO loop in descending order of loop length.

Based on the result of this sorting, the control variable change range changing unit 6
As a result, the change range of the control variable is changed to maximize the loop length of the innermost DO loop. That is, the innermost D
The loop length of the O loop is 120, and Ml and MK, which indicate the final value of the control variable, are exchanged so that the loop length of the outer DO loop decreases to 50 and 20 (
In this example, the initial value and increment value of each control variable remain 1).

In response to this exchange, the array element change unit 7 changes the control variables 1, J within the scope of the multiple DO loop.
In order to change the storage position of the value of the array element of the array DF that depends on K and rDF (IJ K) - DSQ
RT (DBLE (3*I+2*15 J+K)) Variable for the right side of J (subscript of array DF)■
and variable K are exchanged.

Since it is necessary to change the array declaration description of array DF due to the exchange of variable I and variable K, the array declaration modification unit 8 changes the array declaration description (rDIMENS) of array DF.
ION DF (20.50120)J) changes are made.

In order to provide information indicating the above-mentioned changes to the programmer, etc., the DO loop information output unit 9 automatically adds information regarding each multiple DO loop (here, only information regarding the triple DO loop) to the multiple DO loop. The DO loop information list 12 is output to a file after being given a loop identification number.

FIG. 4 is a diagram showing an example of the DO loop information list 12 output in this manner. This DO loop information list 12 includes the multiple DO loop identification number, the statement number of each DO loop in the multiple DO loop, and the control variable for the triple DO loop extracted from the source program 11 before change (original source program). The change range 16 (initial value, final value, and increment value) of the control variable and the loop length (hector length) are different before and after the source program change according to this embodiment (both the source program 11 before the change and the source program 13 after the change). is shown.

Furthermore, the modified source program generation unit 10 generates a modified source program 13 (multiple DO loop calculation speed-up source program) as shown in FIG. 3, reflecting the contents modified as described above. Output and save to file.

FIG. 5 shows the time (CPU time) required to execute the modified source program 13 (see FIG. 3) generated as described above and the unmodified source program 11 (see FIG. 2) in a certain vector computer. FIG.

As shown in this figure, the loop length of the innermost DO loop is 20
By changing from to 120 to speed up the multiple DO loop (triple DO loop), the CPU time was reduced to 0.023
2 to 0.008 seconds (approximately one-third), and it can be seen that this embodiment (and thus the present invention) is very effective in increasing the speed (improving efficiency) of vector operations in vector computers.

Note that due to constraints on internal processing in changing the storage position of array element values in the array element changing section 7, the present invention is suitable for FORTRAN source programs including up to three multiple DO loops.

〔Effect of the invention〕

As explained above, the present invention automatically optimizes the loop length of the innermost DO loop in multiple DO loops so that the loop length of the inner DO loop becomes larger than the loop length of the outer DO loop. By doing so, it is possible to speed up multiple DO loop calculations without having to manually perform complicated operations such as replacing DO loops (achieving efficient operation of vector computers). effective.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, FIG. 2 is a diagram showing an example of the source program before the change in FIG. 1, and FIG. 3 is the source program after the change in FIG. 1. FIG. 4 is a diagram showing an example of the DO rube information list in FIG. FIG. 3 is a diagram showing a comparison with the CPU time required to execute the modified source program shown in FIG. 3; In the figure, 1...Multiple DO loop extraction unit, 2...DO loop control variable extraction unit, 3...Control variable dependent array extraction unit, 4...Loop length evaluation unit, 5...Loop length Sorting section, 6... Control variable change range changing section, 7... Array element changing section, 8 ・ 9 ・ 1 0 ・ 1 1 ・ 1 2 ・ 1 3 ・ ・Array declaration changing section, ・DO lube Information output unit; - source program generation unit after change; - source program before change; - DO loop information list; - source program after change.

Claims (1)

[Claims] An initial value and a final value in which the variable name of the first subscript of the array included in the innermost DO loop matches the control variable name of the innermost DO loop and indicate the change range of the control variable of each DO loop. and F containing multiple DO loops whose increment values are represented by constants
In generating an ORTRAN source program, there is a multiple DO loop extractor that extracts multiple DO loops in the source program before modification, and a control variable for each DO loop in the multiple DO loops extracted by this multiple DO loop extractor. a DO loop control variable extraction unit that detects a change range of the control variable; and an array that depends on the control variables extracted by the DO loop control variable extraction unit from within the multiple DO loops extracted by the multiple DO loop extraction unit. a control variable dependent array extraction unit to extract, and each DO extracted by the DO loop control variable extraction unit;
a loop length evaluation section that calculates the loop length of each DO loop based on the range of change of the control variable of the loop; a loop length sorting section that sorts the loop length of each DO loop based on the evaluation by the loop length evaluation section; For the multiple DO loops extracted by the multiple DO loop extraction unit based on the results of sorting by the length sort unit, the control variables are changed so that the loop length of the innermost DO loop is maximized and the loop length decreases as the outer DO loops become closer. A control variable change range changing unit that exchanges the range between each DO loop, and a control variable change range changing unit that changes the value of an array element of an array having a control variable as a subscript for multiple DO loops in which the change range of a control variable is changed by this control variable change range changing unit. an array element changing unit that changes the storage position in accordance with the exchange; an array declaration changing unit that changes the array declaration description of the array whose storage position of the array element value has been changed by the array element changing unit; a DO loop information output unit that assigns an identification number to the multiple DO loops extracted by the DO loop extraction unit and outputs a DO loop information list; the control variable change range change unit; the array element change unit; and the array declaration. A source program for accelerating multiple DO loop calculations, comprising: a modified source program generation unit that generates and outputs a modified source program that is a source program for accelerating multiple DO loop calculations based on the content changed by the modification unit. Generation method.