JPS6214243A - Program cost estimating system - Google Patents

Abstract

PURPOSE:To estimate execution cost easily by making the execution cost in a scalar processing device on a part made to vector a base, and making the quotient obtained by dividing it by specified magnification execution cost of a vector processing device. CONSTITUTION:The magnification determining section 21 of a processing device 2 judges vector length by repetition designation value in an original program on the loop of each vector making procedure, reads out magnification from a magnification table 12 and gives it to an execution cost processing section 22. The execution cost processing section 22 calculates the quotient obtained by dividing execution cost value on the execution cost data 11 corresponding to each program statement of vector making procedure loop by magnification making execution cost data 11 on the original program 10 input, and outputs it as execution cost in the case where each program statement is executed by a vector processing device.

Description

[Detailed Description of the Invention] [Summary] This is a method for estimating the cost of program execution by a vector processing device. Regarding the vectorization part, the execution cost in the scalar processing device is used as the base, and the quotient obtained by dividing it by a predetermined magnification is defined as the execution cost in the raw processing device. The magnification is determined as a function of the vector length, and the vector length is a value specifying the number of repetitions in vector repeat operations. With this configuration, the program execution cost of the vector processing device can be easily estimated.

[Industrial application field]

The present invention relates to a method for estimating the cost of program execution by a vector processing device of a computer system.

Estimating the execution cost of a program is extremely effective as data for evaluating the performance of the program or improving the program or the compiler of the program.

For a program scheduled to be executed on a vector processing device, the execution cost on the vector processing device is naturally required, but at the same time, a comparative evaluation with the execution cost on the scalar processing device is also often required.

[Prior Art and Problems to be Solved by the Invention] The cost of executing a program can be considered to be approximately proportional to the time that a processing device is occupied in executing the program.

Therefore, the execution cost can be determined by the time occupied by the processing unit, but usually, in order to obtain a more generalized value, for example, the execution time of the integer addition instruction of the processing unit is
The time expressed as is used as the execution cost.

Conventionally, the execution cost of a program executed on a scalar processing device (herein, a typical central processing unit of a computer system is referred to as a scalar processing device to distinguish it from a vector processing device) is calculated as follows, for example: You can ask for it.

That is, by simulating the execution of a so-called target program obtained by compiling the program to be evaluated, and accumulating the execution cost, which is experimentally determined in advance, for each instruction that constitutes the target program, the program is evaluated. Find the execution cost.

For example, in the case of a program outside the FORTRAN language TR, the result is output as an execution cost corresponding to each program statement of the original program, as shown in FIG.

As is well known, a vector processing device is a processing device that has an arithmetic operation and control section that is configured to continuously and efficiently execute the same type of operations that occur in vector operations, and has a relatively large capacity. High-speed performance is achieved by transferring a block of required vector data to a high-speed vector register and continuously supplying the data to the arithmetic unit.

Therefore, the effect of the high speed achieved by a vector processing device differs depending on the length of the vector to be operated on (the length of the vector and the number of vector element data), and in general, the effect of using the vector processing device is small for relatively small vectors.

For this reason, if the execution cost of a program executed on a vector processing device is determined by accumulating the execution cost of unit instructions as in the case of a scalar processing device, the error will generally become large. had to be actually executed on a vector processing device and measure the execution time.

[Means for solving problems]

FIG. 1 is a block diagram showing the configuration of the present invention.

In the figure, 1 is a storage device that holds the original program and other data, 2 is a processing device that calculates the execution cost, 10 of the storage device 1 is the original program, 11 is the execution cost data of the scalar processing device, and 12 is the magnification It's a table.

In the processing device 2, 20 is a vectorization procedure detection section, 21 is a magnification determining section, and 22 is an execution cost processing section.

[For production]

The vectorization procedure detection unit 20 sequentially scans the program statements of the original program 10, analyzes the program using a known method, identifies a repeating loop for vector calculation, and identifies the program statements included in this loop. Extract the sequence as a vectorization procedure.

The magnification determination unit 21 determines the vector length for each vectorization procedure loop based on the repetition specified value in the original program, reads the magnification from the magnification table 12 using the vector length determination value as an index, and passes it to the execution cost processing unit 22.

The execution cost processing unit 22 takes the execution cost data 11 regarding the original program 10 as input and calculates a quotient by dividing the execution cost value on the execution cost data 11 corresponding to each program statement of each vectorized procedure loop by a magnification factor. , and outputs it as the execution cost when each program statement is executed by a vector processing device.

Here, the contents of the magnification table 12 are based on the magnification obtained using analysis data of the execution state, etc., based on each actual measurement value when a typical program statement is executed by a scalar processing device and a vector processing device. , a table is constructed in advance, and for each vector length, it shows the ratio of the execution cost when such a vector operation is executed on a vector processing device to the execution cost when it is executed on a scalar processing device. .

With the above configuration, the execution cost of executing a program on a vector processing device can be reduced without actually executing it.
It can be determined while maintaining reasonable accuracy.

〔Example〕

In FIG. 1, a storage device 1 holds the original program and other data, and a processing device 2 executes a process to determine the execution cost using the data in the storage device 1, and if necessary, displays it on a display device, a printing device, etc. An output list 3 of the determined execution costs is output.

A vectorization procedure detection unit 20 executed by the processing device 2 sequentially scans the program statements of the original program 10, analyzes the program using a known method, identifies a repeat loop for vector calculation, and The sequence of program statements included in the loop is extracted as a vectorization procedure.

For example, when the original program 10 is described below as a program written in FORTRAN, each loop starts with the rDOJ statement of the program statement illustrated in FIG. The statement number indicating the end of the loop (in the example in Figure 3, fDOJ is followed by <“10”)
) ends with the program statement.

The magnification determining unit 21 determines the vector length for each vectorization procedure loop by identifying the repetition specified value in the original program.

That is, the repetition specification [I
・1.1004 is I=1.2.3, -.
99.100, which specifies to execute the loop 100 times, so the vector length is determined to be 100.

An example of the magnification table 12 is shown in the table below.

TableSuch a magnification table is created by calculating the length of each vector using analysis data of the execution state, etc., based on actual measurement values when a typical program statement is executed on a scalar processing device and a vector processing device. It is configured in advance by setting a magnification factor for each vector length. Show the cost ratio.

In the table, if the vector length is relatively short, the scaling factor is less than 1, so when executed on a vector processing device,
This indicates that the execution cost is higher than in the case of a scalar processing device.

The range from the maximum to the minimum is provided in consideration of the fact that the magnification varies depending on the contents of vector calculations.

The magnification determining unit 21 reads the magnification from the magnification table 12 using the determined vector length, and the execution cost processing unit 22
give it to

The execution cost processing unit 22 inputs the execution cost data 11 of the scalar processing device regarding the original program 10, and
Execution cost data corresponding to each program statement of each vectorized procedure loop] Calculate the quotient by dividing the execution cost value above 1 by the magnification factor, and use it as execution cost data when each program statement is executed by a vector processing device 13 Output as .

In this case, the average value is used as the magnification value applied to each individual program statement, and the maximum and minimum values are applied, for example, to the total execution cost of each loop to calculate the range of the estimated execution cost. used to indicate

Thereafter, the output unit 23 edits the execution cost data 1] and the execution cost data 13, and outputs an output list 3 of execution costs to a display device or the like.

Figure 2 shows an example of the output list, and the FORTl?? displayed in the center? Regarding the AN program statement, the execution cost (V-CO3T) on the vector processing device is shown on the left, and the execution cost (S-C
O3T) is shown to make it convenient to compare and examine the effects of vector processing devices and scalar processing devices.

Note that the rVJ display to the right of the number in the V-CO3T column indicates that it is a vectorization procedure part, and the part without this display is not a hectorization procedure, so this part contains the same information as 5-C03T. The execution cost is shown.

With the above configuration, the execution cost when executing a program on a vector processing device can be determined while maintaining reasonable accuracy without actually executing it, and the execution cost of a vector processing device and a scalar processing device can be calculated without actually executing it. It also becomes easier to display costs at the same time to make it easier to compare costs.

〔Effect of the invention〕

As is clear from the above description, according to the present invention, it is possible to easily estimate the execution cost of a program in a vector processing device, and it is also easy to compare this with the execution cost of a scalar processing device. Therefore, there is a significant industrial effect of improving the efficiency of program development and the processing efficiency of information processing systems.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the configuration of an embodiment of the present invention, FIG. 2 is a diagram showing an example of an output of the execution cost of the present invention, and FIG. 3 is a diagram showing an example of the output of the execution cost of a scalar processing device. be. In the figure, ■ is a storage device, 2 is a processing device, 3 is an output list, 10 is an original program, 11.13 is execution cost data, 12 is a magnification table, 20 is a vectorization procedure detection unit, 21 is a magnification determination unit, 22 is an execution cost processing unit;
3 indicates an output section.

Claims (1)

[Claims] A program (10
), a means for inputting the execution cost (11) when the program (10) is executed by a scalar processing device, and an execution cost ratio of the vector processing device and the scalar processing device are calculated based on the vector length. cost ratio determination means (12) determined as a function of
), and the vector length is estimated based on the specified value of the number of repeated operations in the vector operation loop of the program (10), the execution cost ratio is determined from the cost ratio determination means based on the vector length, and the operation is inputted by the input means. A program cost estimation method comprising means (21, 22) for calculating an execution cost estimate when executed by the vector processing device from the execution cost of a loop and the determined execution cost ratio.