JP6810380B2 - Source program conversion system, source program conversion method, and source program conversion program - Google Patents

Description

The present invention relates to an apparatus, method and program for creating another source program optimized for a computer by using a source program as an input.

A program optimization technique is known for the purpose of improving execution performance on a computer by loop unrolling a loop included in a program created in a high-level language. The loop included in the high-level language program is determined for each language, for example, a for loop in C language or C ++ language, and a DO loop in FORTRAN language.

Loop unrolling is a technology that expands a loop so that when the same calculation process is repeatedly executed many times, N times the calculation process that was originally processed in one loop is processed in one loop. is there. Loop unrolling has the effect of widening the scheduling range of instructions in the loop by increasing the amount of processing in one loop, and also has the effect of reducing the overhead of controlling the loop because the number of loops is reduced. There is.

For example, suppose the original source program looks like this:
＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿
for (i = 0; i <n; i ++) {
a [i] = b [i] * c [i];
}
＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿
Here, the asterisk symbol “*” represents multiplication.
The result of loop unrolling for this program is as follows.
＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿
for (i = 0; i <n; i + = 2) {
a [i] = b [i] * c [i];
a [i + 1] = b [i + 1] * c [i + 1];
}
＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿＿
In this loop unrolling, the number of loops is halved by substituting the multiplication result of the array b and the array c into the array a for two elements in the loop.

When loop unrolling a certain program, the execution speed may slow down. For example, when a program obtained by loop unrolling a source program executed by a vector processor is compiled by a compiler for a vector processor, the memory access may become a memory access with one element skipped, so-called stride access.

One measure against this vector computer is described in Patent Document 1. Patent Document 1 describes a data processing device that vectorizes a source program written in a high-level language and, when the conditions are satisfied, expands multiple instruction sequences including vector operation instructions. There is. This data processing device determines whether or not to perform division processing on an instruction sequence including a vectorized vector operation instruction according to a predetermined condition, and when the condition is satisfied, the instruction sequence including the vector operation instruction Expand multiple times.

Further, patent documents 2 to 4 also disclose related techniques. Patent Documents 2 to 4 disclose a method for identifying each loop included in a source program, analyzing the data dependency related to expansion for each loop, and determining whether to expand each loop. There is.

Japanese Unexamined Patent Publication No. 2-4-1562 Japanese Unexamined Patent Publication No. 63-314675 Japanese Unexamined Patent Publication No. 4-307624 Japanese Unexamined Patent Publication No. 4-344535

In the data processing apparatus of Patent Document 1, since the expansion process is performed after the instruction sequence is created by vectorization, the memory access does not become the stride access. Therefore, the data processing device of Patent Document 1 can create a vector instruction sequence for efficient memory access in the vector computer.

However, in the data processing apparatus of Patent Document 1, the expansion process is performed after the instruction sequence is created by vectorization. For this reason, there remains a problem that loop unrolling at the source code level written in a high-level language cannot be realized. Patent Document 2 has a similar problem. Therefore, since it cannot be unrolled at the source code level, it can be pointed out that it is difficult to obtain the effects of duplicate calculation removal that appears with unrolling and instruction scheduling at the time of source code compilation.

With respect to Patent Document 3 and Patent Document 4, it is not possible to obtain a loop unrolling effect suitable for a vector computer as in the invention described below due to differences in restrictions, purposes, and expansion methods.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a source program conversion system capable of creating a program with optimal loop unrolling for a vector computer.

The source program conversion system according to the embodiment of the present invention identifies a program input unit that accepts a source program to be converted and each loop included in the source program input to the program input unit, and puts each loop into each loop. On the other hand, in the loop after expansion, the expansion possibility determination unit that analyzes the data dependency related to expansion and determines whether or not each loop can be expanded and the loop that the expansion possibility determination unit determines that expansion is possible. The number of loop unrolling stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated, and the number of loops of the loop before expansion are calculated in one rotation of the loop after expansion. A program having an expanded loop is generated and processed in response to the number of loop unrolling stages and the loop unrolling width determined by the expansion parameter determination unit that derives the loop unrolling width, which is a positive integer indicating the above. It includes a development program creation unit that outputs as a development program.

The source program conversion method by the information processing system according to the embodiment of the present invention accepts the source program to be converted, identifies each loop included in the source program, and depends on the data related to the expansion for each loop. The relationship is analyzed to determine whether or not each loop can be expanded, and for the loop that is determined to be expandable, a positive integer indicating how many rotations of the loop before expansion is performed in the loop after expansion. In addition to deriving the number of loop unrolling stages, the loop unrolling width, which is a positive integer indicating how many rotations of the loop before expansion are calculated in one rotation of the loop after expansion, is derived and derived. Based on the number of loop unrolling stages and the loop unrolling width, a program having the expanded loop is generated and processed, and output as an expanded program.

The source program conversion program according to the embodiment of the present invention identifies the information processing resource as a program input means for receiving the source program to be converted and each loop included in the source program input to the program input means. Then, after the expansion, the expansion possibility determination means for analyzing the data dependency related to the expansion for each loop and determining whether or not the expansion of each loop is possible and the loop determined by the expansion possibility determination means to be expandable are performed. In the loop of, the number of loop unrolling stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated, and the number of rotations of the loop before expansion are calculated for the loop after expansion. It has an expansion parameter determining means that derives and processes a loop unrolling width that is a positive integer indicating whether to calculate in one rotation, and an expanded loop that receives the number of loop expansion stages and the loop expansion width determined by the expansion parameter determining means. It operates as a decompression program creation means that generates a program and outputs it as a decompressed program.

The recording medium according to the embodiment of the present invention holds the source program conversion program in a computer-readable manner non-temporarily.

According to the present invention, it is possible to provide a source program conversion system capable of creating a program with optimal loop unrolling for a vector computer.

It is a block diagram which shows the structure of the source program conversion system of one Embodiment of this invention. It is explanatory drawing which showed the program list example of the source program used for operation explanation. It is a flowchart which shows the operation of the expansion parameter determination part. It is a flowchart which shows the operation of the expansion program creation part. It is a figure which showed the expanded program created in the source program conversion system of one Embodiment. It is a figure which showed the other expanded program created in the source program transformation system of one embodiment. It is a flowchart which shows another operation of the expansion parameter determination part. It is a figure which showed the expanded program created by using the flowchart shown in FIG. 7. It is a figure which showed another expanded program created by using the flowchart shown in FIG. 7. It is a block diagram which shows the structural example of the source program conversion system of one Embodiment of this invention.

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

[First Embodiment]
[Description of configuration]
FIG. 1 is a block diagram showing a source program conversion system according to the first embodiment. The loop unrolling device 10 performs loop unrolling on the loop unrollable loop in the loop included in the source program 21 stored in the storage device 20, and stores the loop unrolled program in the storage device 20 as the expanded program 22. It is a device to do. The storage device 20 may be built in the loop unrolling device 10.

The loop unrolling device 10 according to the present embodiment has the following configuration.
・ Program input unit 11
・ Deployment availability judgment unit 12
-Expansion parameter determination unit 13
・ Deployment program creation unit 14
The program input unit 11, the expandability determination unit 12, the expansion parameter determination unit 13, and the expansion program creation unit 14 may be realized by a computer program and a processor operating according to the computer program, or some or all of the components may be realized. It may be formed by an electronic circuit.

FIG. 2 is an example of the source program 21 that is the target of loop unrolling in this embodiment.

The program input unit 11 receives the input of the source program 21 stored in the storage device 20, performs parsing, lexical analysis, and semantic analysis on the program, and passes the result to the expandability determination unit 12. These parsing, lexical analysis, and semantic analysis may be performed using existing techniques.

The expandability determination unit 12 determines whether or not the loop in the passed source program can be expanded. In order to determine whether or not expansion is possible, it is sufficient to determine whether or not there is a dependency of data in the loop and whether or not the counter variable of the loop changes linearly. When there is no dependency of the data in the loop and the counter variable of the loop changes linearly, the expandability determination unit 12 determines that the loop can be expanded. Dependency analysis of the data in this loop may be performed using existing technology. Further, the increment value of the counter variable of the loop (hereinafter referred to as L) is used by the expansion parameter determination unit 13 and the expansion program creation unit 14 described later.

In the case of the source program example as shown in FIG. 2, since the arrays A, B, and C are different arrays, there is no data dependency. Further, since the counter variable i of the loop increases by 1 for each rotation of the loop, it is a linear increase. Therefore, the source program as shown in FIG. 2 is determined to be loop unrollable. Further, the increase value L of the loop counter variable is 1.

The expansion parameter determination unit 13 uses two types of parameters for loop unrolling for the loop determined to be loop unrollable, using predetermined conditions based on the array counter variable in the loop and the calculation amount of the calculation formula in the loop. That is, the loop unrolling width and the number of loop unrolling stages are derived.

This loop unrolling width (hereinafter, also referred to as M) indicates whether or not the calculation processing of the loop before expansion, which is separated by how many rotations, is expanded in the same loop. Further, the number of loop unrolling stages (hereinafter, also referred to as N) indicates how many rotations of the loop before expansion are calculated in one rotation of the loop after expansion. For example, the calculation of the first rotation before the loop unrolling is developed so as to be calculated in the same loop as the calculation of the M + 1th rotation, the 2M + 1th rotation, ... (N-1) × M + 1th rotation.

The expansion program creation unit 14 generates a program (expanded program 22) that expands the target loop according to the loop expansion width and the number of loop expansion stages determined by the expansion parameter determination unit 13, and stores the generated expanded program 22. Output to 20. The output destination of the expanded program 22 does not necessarily have to be limited to the storage device 20 in which the source program 21 is stored, and may be configured to record in, for example, a storage connected via a network. .. Further, the output destination of the expanded program 22 may be temporarily displayed on the display or printed on the printer.

[Explanation of operation]
Next, the operations of the expansion parameter determination unit 13 and the expansion program creation unit 14 will be described.

First, the operation of the expansion parameter determination unit 13 will be described with reference to FIG. FIG. 3 is a flowchart showing the operation of the expansion parameter determination unit 13.

The expansion parameter determination unit 13 (loop expansion device 10) derives the loop rotation speed (Z) of the source program 21 and the initial value (K) of the counter variable (S301).

In the case of the loop of the source program 21 shown in FIG. 2, the counter variable i is incremented by 1 from 100 to 10000. Therefore, the expansion parameter determination unit 13 derives the rotation speed Z as “9901” and the initial value K of the counter variable as “100”.

Next, the expansion parameter determination unit 13 derives the cost (C) for one rotation of the loop of the source program 21 (S302). In this embodiment, the cost of one rotation of the loop is the number of operations in the loop. In the case of the loop of the source program 21 shown in FIG. 2, since the addition is once, the expansion parameter determination unit 13 derives the cost C = 1.
In the present embodiment, the cost of one rotation of the loop is used as the number of operations, but the cost is determined by using other factors such as the number of memory accesses (number of loads and number of stores) to the array data in the loop. May be good. The method for determining the cost of one loop rotation is not particularly limited. For example, the number of operations in the loop and the number of memory accesses may be combined.

Next, the expansion parameter determination unit 13 refers to the minimum cost threshold value (C _min ) and derives the minimum integer N _{c such} that C _min ≤ C × N _c (S303). The minimum cost threshold value (C _min ) is a constant value predetermined for each loop unrolling device 10 or for each vector computer that executes an instruction sequence in which the source program 21 is compiled. The setting method is not particularly limited. In this embodiment, it is assumed that C _min = 32. When C _min = 32 and C = 1, the smallest integer N _c = 32 satisfying C _min ≤ C × N _c .

Next, the expansion parameter determination unit 13 refers to the maximum value (N _max ) of the number of loop unrolling stages, and determines the smaller of the previously calculated N _c and the maximum value (N _max ) of the number of loop unrolling stages. Determined to be N (S304). The maximum value (N _max ) of the number of loop unrolling stages is a constant value predetermined for each loop unrolling device 10 or for each vector computer that executes the instruction sequence in which the source program 21 is compiled. The setting method is not particularly limited. In this embodiment, it is assumed that N _max = 4. “4”, which is the smaller value of N _c = 32 and N _max = 4, is determined as the number of loop unrolling stages N.

Next, the expansion parameter determination unit 13 refers to the previously calculated loop rotation speed Z and the loop expansion stage number N, and determines the maximum integer M such that M × N <Z as the loop expansion width M ( S305). In this example, since Z = 9901 and N = 4, it is determined that M = 2475.

Next, the operation of the expansion program creation unit 14 will be described with reference to FIG. FIG. 4 is a flowchart showing the operation of the expansion program creation unit 14.

First, the expansion program creation unit 14 (loop unrolling device 10) creates a double loop structure (S401). The counter variable of the outer loop is set to the same variable as the counter variable of the loop of the input source program 21. In the case of the source program 21 shown in FIG. 2, the counter variable i of the outer loop is set by the expansion program creation unit 14. The initial value of this counter variable is set to "0". Further, the increase value of the counter variable for each rotation of the loop is N × M using the loop expansion width M and the number of loop expansion stages N determined by the expansion parameter determination unit 13. Further, the loop termination condition is i <Z−M × N + 1 using the loop rotation speed Z of the source program. As the counter variable of the inner loop, any variable that does not appear in the source program 21 is used. Here, the counter variable of the inner loop is represented as j. The initial value of the counter variable of this inner loop is set to 0. Further, the increase value of the counter variable for each rotation of the loop of the inner loop is set to the increase value L of the counter variable of the loop of the source program 21. The loop end condition of the inner loop is set to j <M.

Next, the expansion program creation unit 14 creates a calculation formula in the loop of the created double loop structure (S402). First, the expansion program creation unit 14 arranges the calculation formulas in the loop of the source program 21 for the number of loop expansion stages (that is, N). Next, the expansion program creation unit 14 sets the loop variable i in the loop to (i + K + j + 0 × L × M), (i + K + j + 1 × L × M), (i + K + j + 2 × L × M), ..., (I + K + j + (N-1)). Replace with N kinds of terms of × L × M). As described above, K is the initial value of the counter variable of the loop of the original source code 21.

Next, the expansion program creation unit 14 creates a remainder loop (S403). The remainder loop is a loop that processes a loop for the number of rotations less than the number of expansion stages when the loop is expanded. The counter variable of the remainder loop inherits the counter variable of the double loop. The expansion program creation unit 14 sets the initial value of the counter variable of the remainder loop to L × i + K by using the value i at the end of the loop of the counter variable of the double loop. Further, the expansion program creation unit 14 sets the increase value of the counter variable of the remainder loop to L and the end condition to i <L × Z + K. Then, the expansion program creation unit 14 arranges one calculation formula in the loop of the source program as it is in the calculation formula in the loop of the remainder loop.

Here, the operation of the expansion program creation unit 14 will be described as an example when the source program 21 of FIG. 2 is input. FIG. 5 is a program (expanded program 22) that the expansion program creation unit 14 receives and outputs the input of the source program 21 of FIG.

First, the expansion program creation unit 14 creates a double loop structure (double loop structure creation step; S401). Set the counter variable of the outer loop to the same i as the source program. Further, the initial value of the counter variable is "0", and the increase value of the counter variable is set to "2745 x 4" because the loop unrolling width M is "2745" and the number of loop unrolling stages is "4". The loop end condition is "i <9901-2475 x 4 + 1" because the rotation speed Z of the loop of the source program is "9901". The expansion program creation unit 14 integrates these to create the for statement on the first line of FIG. Next, the expansion program creation unit 14 sets the counter variable to j, which does not appear in the source program, for the inner loop. The initial value of j is set to "0", and the increase value is set to "1", which is the increase value L of the counter variable of the source program 21. Further, the loop end condition is set to "2745" of the loop unrolling width M. In this way, the expansion program creation unit 14 creates the for statement on the second line of FIG.

Next, the expansion program creation unit 14 creates an inner loop calculation formula (inner loop calculation formula creation step; S402). The calculation formula of the source program is only one formula in the second line of FIG. The expansion program creation unit 14 arranges four of these equations, which is the number of loop expansion stages N. At this time, the expansion program creation unit 14 sets the counter variables i appearing in the calculation formula as (i + 100 + j + 0 × 1 × 2475), (i + 100 + j + 1 × 1 × 2475), (i + 100 + j + 2 × 1 × 2475), (i + 100 + j + 0 × 1 ×, respectively. Replace with the four types of terms in 2475). The result is the equation in the 3rd to 6th lines of FIG. Then, parentheses indicating the end of the inner loop (lines 5 and 7) and parentheses indicating the end of the outer loop (lines 5 and 8) follow.

Next, the expansion program creation unit 14 creates a remainder loop (remainder loop creation step; S403). As the counter variable of the remainder loop, i, which is the counter variable of the previous double loop, is used. Then, the initial value of the counter variable is "1 x i + 100" by using "1" which is the increase value L of the counter variable of the source program and "100" which is the initial value K of the counter variable of the source program. Further, the increase value of the counter variable is “1” which is L, and the end condition of the loop is “9901” which is the rotation speed Z of the loop of the source program, and “i <1 × 9901 + 100”. The 9th line in FIG. 5 is a for statement showing the loop structure of the remainder loop. Then, the calculation formula of the loop of the source program 21 is arranged as it is (FIGS. 5 and 10), followed by parentheses indicating the end of the loop (FIGS. 5 and 11).

In this way, the expansion program creation unit 14 creates an expansion loop.

Further, the term using the constant in the expansion loop may be a calculated value as shown in FIG.

As can be seen with reference to FIGS. 5 and 6, due to the above series of loop unrolling processes, the number of calculation formulas in the loop was one in the original source program, but a loop structure having four calculation formulas A program (expanded program 22) has been created. That is, it can be seen that the expanded program 22 can loop unroll the source program 21 at the source program level.

Further, when the expanded program 22 is compiled and executed so that it can be executed by a vector computer, the memory access of each vector operation becomes continuous access, which is a vector operation for efficient memory access.

That is, according to the present invention, it is possible to provide a source program conversion system capable of creating a program with optimal loop unrolling for a vector computer.

Here, the addition and replacement functions of the loop unrolling device 10 (expansion parameter determination unit 13) will be described. The expansion parameter determination unit 13 sets the loop expansion width to an integer equal to or less than (loop rotation speed ÷ loop expansion stage number). On the other hand, the expansion parameter determination unit 13 having the addition / replacement function has the loop expansion width and the vector length of the vector computer (computer that executes the source program 21) that executes the expanded program 22 (hereinafter referred to as VL). Notation) is set.

In the following description, it is assumed that the vector computer that executes the source program is determined, and the vector length of the computer is held in either the loop unrolling device 10 or the source program. Further, the operator (programmer) may manually input the vector length to the loop turning device 10.

Further, the loop turning device 10 (expansion parameter determination unit 13) may be configured to determine by preprocessing whether the vector length or model of the vector computer that executes the source program is described in the source code. When the model can be specified, the value of the vector length of the model may be obtained by referring to the model-vector length database prepared in advance.

FIG. 7 is a flowchart showing the operation of the expansion parameter determination unit 13 according to this operation example. Since the steps (S301 to S304) until the number of loop unrolling stages N is determined are the same as the flowchart of FIG. 3 described above, the description thereof will be omitted. The expansion parameter determination unit 13 determines the loop expansion width M after determining the number of loop expansion stages N (after executing S304). The loop unrolling width M determines the vector length VL of the computer that executes the source program as the loop unrolling width M (S705).

As an example of the source program 21 according to this modification, the operation will be described using the loop shown in FIG. 2 described above. Further, it is assumed that the vector length VL of the computer that executes the source program is "256". Since the process up to determining the number of loop unrolling stages N is the same as the flowchart of FIG. 3 described above, the expansion parameter determination unit 13 sets the loop unrolling number Z of the source program to "9901" and the initial value of the counter variable. K is determined to be "100", and the number of loop unrolling stages N is determined to be "4". Finally, the loop unrolling width M is determined to be "256", which is the vector length VL of the computer that executes the source program.

FIG. 8 shows a program (expanded program 22) created by the expansion program using the parameters determined by the expansion parameter determination unit. FIG. 9 is an expanded program 22 in the case where a term including a plurality of constant values is further calculated to simplify the term.

As can be seen with reference to FIGS. 8 and 9, due to the above series of loop unrolling processes, the number of calculation formulas in the loop was one in the original source program, but a loop structure having four calculation formulas A program (expanded program 22) has been created. That is, it can be seen that the expanded program 22 can loop unroll the source program 21 at the source program level.

Further, when the loop unrolling program developed by the loop unrolling device 10 of the first modification of the present embodiment is vectorized, the memory areas accessed by a plurality of vector operations in the loop become continuous. For example, when considering the calculation formulas in the third and fourth lines of FIGS. 8 and 9, the term "B [i + j + 100]" in the calculation formula in the third line is vectorized. , "B [i + j + 100]" to "B [i + j + 355]", which is a continuous 256 elements on the memory. The term “B [i + j + 356]” on the 4th line is vectorized from “B [i + j + 356]” to “B [i + j + 611]”, which is a continuous 256 in memory. It is an element. All of these 512 elements are continuous in the memory (the same applies to the terms in the 5th and 6th lines, and all are continuous in the memory).

Further, when the expanded program 22 is compiled and executed so that it can be executed by the corresponding vector computer, the memory access of each vector operation by the corresponding vector computer becomes continuous access, and the vector operation performs more efficient memory access. It becomes.

On the other hand, the expansion method shown above has an advantage that loop unrolling for a general-purpose vector computer can be performed.

As described above, the source program conversion system to which the present invention is applied can create a program with optimal loop unrolling for a vector computer.

Each part of the source program conversion system may be realized by using a combination of hardware and software of a computer system as illustrated in FIG. This computer system includes one or more processors and memory tailored to the desired form. Further, in the form of this computer system, in each part, a source program conversion program is expanded in the above memory, and based on this program, hardware such as one or more processors is operated by an execution instruction group or a code group. It should be realized. At this time, if necessary, this program may realize each part in cooperation with the functions provided by the software such as the operating system, the microprogram, and the driver.

The program data expanded in the memory appropriately includes an execution instruction group, a code group, a table file, content data, and the like that operate the processor as one or more of the above-mentioned parts.

In addition, this computer system does not necessarily have to be constructed as a single device, but is constructed by combining multiple servers, computers, virtual machines, etc., so-called thin clients, distributed computing, and cloud computing. May be good. Further, a part or all parts of the computer system may be replaced with hardware or firmware (for example, one or more LSIs: Large-Scale Integration, FPGA: Field Programmable Gate Array, a combination of electronic elements). Similarly, only a part of each part may be replaced with hardware or firmware.

Further, this program may be temporarily recorded on a recording medium and distributed. The program recorded on the recording medium is read into the memory via wired, wireless, or the recording medium itself, and operates the processor or the like.

In the present specification, the recording medium also includes a storage medium, a memory device, a storage device, and the like of similar terms. Examples of this recording medium include optical disks, magnetic disks, semiconductor memory devices, hard disk devices, tape media, and the like. Further, it is desirable that the recording medium is non-volatile. Further, as the recording medium, a combination of a volatile module (for example, RAM: Random Access Memory) and a non-volatile module (for example, ROM: Read Only Memory) may be used.

To explain the above embodiment in another expression, the information processing resource that operates the processor as a loop unrolling device is the program input unit and the expandability determination shown in FIG. 1 based on the source program conversion program expanded in the memory. As a result, the source program conversion system can be realized by operating as a unit, an expansion parameter determination unit, and an expansion program creation unit. Each component can also be regarded as a functional unit (software module) of a software program. However, the division of each part shown in these drawings is a configuration for convenience of explanation, and various configurations can be assumed at the time of mounting.

Similarly, if the above embodiment is described in yet another expression, the recording medium includes a source program conversion program that is expanded in the memory and operates on the information processing resource, and the information processing resource includes a program input step and a deployment possibility determination step. A source program conversion system (loop unrolling device) can be constructed by executing the unrolling parameter determination process and the unrolling program creation process in a timely manner.

The present invention has been described by exemplifying embodiments. However, the specific configuration of the present invention is not limited to the above-described embodiment, and is included in the present invention even if there is a change within a range that does not deviate from the gist of the present invention. For example, changes such as separation and merging of block configurations and replacement of procedures of the above-described embodiments are free as long as the gist of the present invention and the functions described are satisfied, and the above description does not limit the present invention.

In addition, some or all of the above embodiments may also be described as follows. The following notes do not limit the present invention in any way.
[Appendix 1]
A program input section that accepts the source program to be converted,
Expandability determination unit that identifies each loop included in the source program input to the program input unit, analyzes the data dependency related to expansion for each loop, and determines whether or not each loop can be expanded. When,
With respect to the loop determined to be expandable by the expansion possibility determination unit, the number of loop expansion stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated in the loop after expansion, and the number of loop expansion stages before expansion. An expansion parameter determination unit that derives the loop expansion width, which is a positive integer indicating how many rotations of the loop the loop is calculated in one rotation of the loop after expansion.
An expansion program creation unit that receives the number of loop expansion stages and the loop expansion width determined by the expansion parameter determination unit, generates and processes a program having an expanded loop, and outputs it as an expanded program.
A source program conversion system that includes.

[Appendix 2]
In the source program conversion system described in Appendix 1 above,
The program input unit receives the source program used in the vector computer and receives it.
The expansion program creation unit generates and processes a program having a loop expanded for the vector computer based on the number of loop expansion stages and the loop expansion width determined by the expansion parameter determination unit, and produces the expanded program. The source program conversion system according to claim 1 for output.

[Appendix 3]
In the source program conversion system described in Appendix 1 or 2 above,
The expansion parameter determination unit calculates the loop processing cost (C) based on at least one of the number of calculations and the number of memory accesses in the loop to be expanded in the received source program, and calculates a predetermined threshold value (Cmin). ) Is used to determine the smallest integer N satisfying Cmin ≤ C × N as the number of loop unrolling stages.

[Appendix 4]
In the source program conversion system described in any one of the above appendices 1 to 3,
The expansion parameter determination unit calculates the loop rotation speed of the loop developed by the source program that has been received, refers to the determined loop expansion stage number, and sets the loop expansion width as (the loop rotation speed of the source program ÷ the loop expansion stage number). ) A source program conversion system that sets the following positive integers.

[Appendix 5]
In the source program conversion system described in any one of the above appendices 1 to 3,
The expansion parameter determination unit is a source program conversion system that sets the vector length of the vector computer that executes the expanded program in the loop expansion width.

[Appendix 6]
In the source program conversion system described in any one of the above appendices 1 to 3,
The expansion parameter determination unit determines whether the vector length or model of the vector computer that executes the source program is described in the source code, and if the vector length or model is described, the described vector. A source program conversion system that identifies the vector length determined by the length or model and sets the identified vector length in the loop expansion width.

[Appendix 7]
In the source program conversion system according to any one of the above appendices 1 to 6,
In the expansion program creation unit, the outer loop increases the counter variable from the initial value 0 by the product of the loop expansion width and the number of loop expansion stages for each loop rotation, and the number of rotations of the loop of the source program and the said. The loop is looped until it is less than or equal to the difference between the product values, and the inner loop increases the counter variable from the initial value of 0 by the increase value of the counter variable of the source program loop for each loop rotation. Create a double loop that loops for less than the loop expansion width.
A source program conversion system that expands the calculation formula in the loop of the source program in the double loop by the number of loop unrolling stages.

[Appendix 8]
Accept the source program to be converted,
Each loop included in the source program is identified, the data dependency related to expansion is analyzed for each loop, and whether or not each loop can be expanded is determined.
For the loop determined to be expandable, in the loop after expansion, the number of loop expansion stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated, is derived, and what is in the loop before expansion. The loop unrolling width, which is a positive integer indicating whether to calculate the loop unrolled loop in one rotation of the loop after expansion, is derived.
A source program conversion method by an information processing system that generates and processes a program having expanded loops based on the number of derived loop unrolling stages and the loop unrolling width, and outputs it as an expanded program.

[Appendix 9]
Information processing resources,
Program input means that accepts the source program to be converted,
Expansion propriety determination means for identifying each loop included in the source program input to the program input means, analyzing the data dependency related to expansion for each loop, and determining whether or not each loop can be expanded. When,
With respect to the loop determined to be expandable by the expansion possibility determination means, the number of loop expansion stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated in the loop after expansion, and the number of loop expansion stages before expansion. An expansion parameter determining means for deriving the loop expansion width, which is a positive integer indicating how many rotations of the loop the loop is calculated in one rotation of the loop after expansion.
An expansion program creating means that receives the number of loop expansion stages and the loop expansion width determined by the expansion parameter determining means, generates and processes a program having an expanded loop, and outputs it as an expanded program.
Source program conversion program that operates as.

[Appendix 10]
The source program conversion program according to Appendix 9, wherein the information processing resource is a source program conversion program that uses a combination of a processor, a memory, and a bus.

[Appendix 11]
A recording medium that non-temporarily holds the source program conversion program according to Appendix 9 or 10 in a computer-readable manner.

According to the present invention, it can be applied to an application such as a software optimization device that creates a software program optimized for a target vector computer by using a software program as an input.

10 Loop unrolling device (information processing device, information processing resource)
11 Program input unit 12 Expandability judgment unit 13 Deployment parameter determination unit 14 Deployment program creation unit 20 Storage device 21 Source program 22 Expanded program (optimized source program)

Claims (9)

A program input section that accepts the source program to be converted,
Expandability determination unit that identifies each loop included in the source program input to the program input unit, analyzes the data dependency related to expansion for each loop, and determines whether or not each loop can be expanded. When,
With respect to the loop determined to be expandable by the expansion possibility determination unit, the number of loop expansion stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated in the loop after expansion, and the number of loop expansion stages before expansion. An expansion parameter determination unit that derives the loop expansion width, which is a positive integer indicating how many rotations of the loop the loop is calculated in one rotation of the loop after expansion.
An expansion program creation unit that receives the number of loop expansion stages and the loop expansion width determined by the expansion parameter determination unit, generates and processes a program having an expanded loop, and outputs it as an expanded program.
A source program conversion system that includes. In the source program conversion system according to claim 1,
The program input unit receives the source program used in the vector computer and receives it.
The expansion program creation unit generates and processes a program having a loop expanded for the vector computer based on the number of loop expansion stages and the loop expansion width determined by the expansion parameter determination unit, and produces the expanded program. The source program conversion system according to claim 1 for output. In the source program conversion system according to claim 1 or 2.
The expansion parameter determination unit calculates the loop processing cost (C) based on at least one of the number of calculations and the number of memory accesses in the loop to be expanded in the received source program, and calculates a predetermined threshold value (Cmin). ) Is used to determine the smallest integer N satisfying Cmin ≤ C × N as the number of loop unrolling stages. In the source program conversion system according to any one of claims 1 to 3,
The expansion parameter determination unit calculates the loop rotation speed of the loop developed by the source program that has been received, refers to the determined loop expansion stage number, and sets the loop expansion width as (the loop rotation speed of the source program ÷ the loop expansion stage number). ) A source program conversion system that sets the following positive integers. In the source program conversion system according to any one of claims 1 to 3,
The expansion parameter determination unit is a source program conversion system that sets the vector length of the vector computer that executes the expanded program in the loop expansion width. In the source program conversion system according to any one of claims 1 to 3,
The expansion parameter determination unit determines whether the vector length or model of the vector computer that executes the source program is described in the source code, and if the vector length or model is described, the described vector. A source program conversion system that identifies the vector length determined by the length or model and sets the identified vector length in the loop expansion width. In the source program conversion system according to any one of claims 1 to 6.
In the expansion program creation unit, the outer loop increases the counter variable from the initial value 0 by the product of the loop expansion width and the number of loop expansion stages for each loop rotation, and the number of rotations of the loop of the source program and the said. The loop is looped until it is less than or equal to the difference between the product values, and the inner loop increases the counter variable from the initial value of 0 by the increase value of the counter variable of the source program loop for each loop rotation. Create a double loop that loops for less than the loop expansion width.
A source program conversion system that expands the calculation formula in the loop of the source program in the double loop by the number of loop unrolling stages. Accept the source program to be converted,
Each loop included in the source program is identified, the data dependency related to expansion is analyzed for each loop, and whether or not each loop can be expanded is determined.
For the loop determined to be expandable, in the loop after expansion, the number of loop expansion stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated, is derived, and what is in the loop before expansion. The loop unrolling width, which is a positive integer indicating whether to calculate the loop unrolled loop in one rotation of the loop after expansion, is derived.
A source program conversion method by an information processing system that generates and processes a program having expanded loops based on the number of derived loop unrolling stages and the loop unrolling width, and outputs it as an expanded program. Information processing resources,
Program input means that accepts the source program to be converted,
Expansion propriety determination means for identifying each loop included in the source program input to the program input means, analyzing the data dependency related to expansion for each loop, and determining whether or not each loop can be expanded. When,
With respect to the loop determined to be expandable by the expansion possibility determination means, the number of loop expansion stages, which is a positive integer indicating how many rotations of the loop before expansion are calculated in the loop after expansion, and the number of loop expansion stages before expansion. An expansion parameter determining means for deriving the loop expansion width, which is a positive integer indicating how many rotations of the loop the loop is calculated in one rotation of the loop after expansion.
An expansion program creating means that receives the number of loop expansion stages and the loop expansion width determined by the expansion parameter determining means, generates and processes a program having an expanded loop, and outputs it as an expanded program.
Source program conversion program that operates as.

Images