JP2007066128A - Compilation processing method, compilation processor and compilation processing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a new compilation processing technology capable of shortening the execution time of a loop for executing an operation of a non-dense matrix when a source program has the loop. <P>SOLUTION: This compilation processor is constituted so that a multiplex loop sentence in the innermost loop sentence of which the non-dense matrix is included and which has an operation whose execution can be omitted when data about the non-dense matrix is 0 is detected regarding multiplex loop sentences described in the source program; whether or not the data about the non-dense matrix of the operation which the innermost loop sentence has is the data whose value is not updated by rotation of the loop is judged regarding the innermost loop sentence of the detected multiplex loop sentence; whether or not the data about the non-dense matrix is 0 is inspected immediately before the innermost loop sentence in which it is judged that it is the data whose value is not updated; and a statement instructing that the loop proceeds to the next rotation without performing the operation in the innermost loop sentence when the data is 0 is inserted into the multiplex loop sentence and the multiplex loop sentence into which the statement is inserted is parallelized in a cyclic system. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a compile processing method and apparatus for compiling a source program, and a compile processing program used for realizing the compile processing method, and in particular, when the source program has a loop for executing a sparse matrix operation, The present invention relates to a compile processing method and apparatus capable of shortening the execution time, and a compile processing program used for realizing the compile processing method.

  In the conventional technique, when a source program has a loop for executing an operation of a sparse matrix (matrix having many 0s), even though there is an unnecessary operation, the execution time is long. There is a problem of becoming.

For example,
DO J = 1, N
DO I = 1, M
A (I, J) = A (I, J) + B (J) * C
ENDDO
ENDDO
In the double loop that executes the operation of the matrix B, as shown in FIG. 7, the inner loop is executed M times.

  In this case, in the prior art, the inner loop is executed M times as it is without considering whether the matrix B is a sparse matrix. However, when the value of B (J) is 0, the calculation result of the inner loop is “A (I, J)”, which means that the inner loop does not require calculation.

  When the matrix B is a sparse matrix, the value of B (J) is often 0. From now on, if the inner loop is executed M times as it is without considering whether or not the matrix B is a sparse matrix as in the prior art, the inner loop is changed even though there are many cases where the calculation is unnecessary. Since it is executed, there is a problem that the execution time becomes long.

  This problem is a problem that cannot be solved by simply parallelizing the above-described double loop.

  That is, normally, such a multiple loop is parallelized in the outermost dimension. Therefore, when this loop is parallelized for the parallel processors of two CPUs 1 and 2, CPU1 and CPU2 respectively A loop as shown in FIG. 8 is executed.

The execution time when parallelized in this way is as follows:
MAX (CPU1 execution time, CPU2 execution time,..., CPUn execution time)
Can be obtained.

From now on, if the execution time of the inner loop is T, the execution time of the above-mentioned double loop is
When not parallelized: T * N
Parallelization: T * (N / 2)
It becomes.

  Thus, when the source program has a multiple loop such as a double loop, the execution time can be greatly shortened by parallelizing the source program.

  However, even if such parallelization is performed, there is no change in that the inner loop is executed as it is, although there are many cases where calculation is not required in each CPU, and the problem that the execution time becomes longer will be solved. It will not be possible.

  Here, there are Patent Documents 1 and 2 shown below as conventional techniques related to the present invention.

  In the invention described in Patent Document 1, when performing matrix operation of a sparse matrix (sparse matrix) having a characteristic that most matrix elements are 0, all appearing in the sparse matrix operation process in the preprocessing A non-zero element position is detected and a bitmap is created, and only a matrix element position indicated by the bitmap is calculated to perform an iterative matrix solution without waste.

In the invention described in Patent Document 2, the execution of the assigned loop is performed when the execution speed of the in-procedure loop is improved by executing a plurality of tasks in parallel on a plurality of processors. By making the finished processor execute the unprocessed loop, the execution speed is further improved.
JP-A-60-247772 JP-A-3-218556

  As described above, in the conventional technology, when a source program has a loop for executing a sparse matrix operation, an unnecessary operation is included, but the source program is compiled as it is without considering such an operation. Like to do.

  From now on, according to the prior art, when the source program has a loop that executes sparse matrix operations, even though there are operations that do not need to be executed, the execution time becomes longer because they are executed There is.

  The present invention has been made in view of such circumstances. When a source program has a loop for executing a sparse matrix operation, the execution time can be shortened when compiling the source program. The purpose is to provide new compilation technology.

  In order to achieve this object, the compile processing apparatus of the present invention (1) includes a sparse matrix in the innermost loop statement of a multiple loop statement described in a source program and includes the sparse matrix. (2) an operation of the innermost loop statement for the innermost loop statement of the multiple loop statement detected by the detecting means; Determining means for determining whether the data of the sparse matrix is data whose value is not updated due to the rotation of the loop, and (3) immediately before the innermost loop statement in which the determining means determines that the data is not updated. Checking whether the data of the sparse matrix is 0, and if it is 0, an insertion means for inserting a command statement instructing to proceed to the next rotation without performing an operation in the innermost loop statement; 4) Insertion means Configured to include a parallelization means to parallelize the multi-loop sentence by inserting the Ryobun in the form of a cyclic manner.

  When configured in this way, the detection means includes a sparse matrix in the single loop statement described in the source program, and the data of the sparse matrix is 0. In some cases, an operation having an operation that can be omitted is detected.

  In this case, for the single loop statement detected by the detection unit, the determination unit determines whether the sparse matrix data of the operation of the single loop statement is data whose value is not updated by the rotation of the loop, The inserting unit checks whether the data of the sparse matrix is 0 immediately before the single loop statement in which the determining unit determines that the data is not updated, and if it is 0, the single loop statement An instruction statement instructing to proceed to the next rotation without performing the operation is inserted, and the parallelizing means parallelizes the single loop statement in which the inserting means has inserted the instruction sentence in a cyclic manner. To process.

  The compile processing method of the present invention realized by the operation of each processing means described above can also be realized by a computer program, and this computer program is provided by being recorded on a suitable computer-readable recording medium, The present invention is realized by being provided via a network, installed when executing the present invention, and operating on a control means such as a CPU.

  In the compile processing device of the present invention configured as described above, for the multiple loop statement described in the source program, the innermost loop statement of the multiple loop statement includes a sparse matrix, and the data of the sparse matrix is 0. If an operation with an operation that can be omitted is detected, the data of the sparse matrix of the operation of the innermost loop statement is updated by the rotation of the loop. Determine whether the data is not processed.

  Subsequently, immediately before the innermost loop statement for which it is determined that the data is not updated, it is checked whether the data of the sparse matrix is 0. If it is 0, the operation in the innermost loop statement is performed. The command statement instructing to proceed to the next rotation without inserting is inserted, and the multiple loop statement in which the command statement is inserted is parallelized in a cyclic manner.

Thus, in the compile processing apparatus of the present invention,
DO J = 1, N
DO I = 1, M
A (I, J) = A (I, J) + B (J) * C
ENDDO
ENDDO
In the case of a double loop statement that executes the operation of the sparse matrix B, as shown in FIG. 1A, when the data B (J) of the sparse matrix B is 0, the inner loop statement is executed. Since the value of this data B (J) is not updated by the rotation of the inner loop, it is checked immediately before this inner loop statement whether the value of B (J) is 0. In some cases, a configuration is adopted in which a command statement (indicated by the * part shown in the figure) that prevents execution of the inner loop statement is inserted.

  As described above, in the compile processing device of the present invention, it is possible to avoid executing unnecessary operations by inserting a statement that makes it possible to omit the execution of unnecessary operations in the source program. A configuration for shortening is adopted.

However, as shown in FIG.
DO I = 1, M
A (I) = A (I) + B (I) * C
ENDDO
Whether the value of B (I) is 0 when the above-mentioned command statement (* command statement shown in the figure: a command statement that allows the execution of unnecessary operations to be omitted) is inserted in the single loop statement Regardless of this, since this statement must be executed every time the loop rotates once, even if the value of B (I) is almost 0, it is executed by the overhead of executing this statement. It becomes difficult to shorten the time.

  From now on, in the compile processing apparatus of the present invention, the above statement is not inserted in principle for a single loop statement. However, as shown in FIG. If the operation data is invariant because it is scalar data, the above statement can be moved out of the loop in the same way as the innermost loop statement of the multiple loop statement described above. Therefore, it is not necessary to be aware of the overhead caused by executing the command statement (* command statement shown in the diagram).

  Therefore, in the compile processing apparatus of the present invention, the execution of a single loop statement described in a source program is omitted when the single loop statement includes a sparse matrix and the data of the sparse matrix is 0. When an object having an operation that can be detected is detected, it is determined whether the data of the sparse matrix of the operation that the single loop statement has is the data whose value is not updated by the rotation of the loop. Immediately before the single loop statement that is determined to be non-updated data, it is checked whether the data of the sparse matrix is 0. If it is 0, the operation in the single loop statement is not performed. A command statement for instructing to proceed to the rotation is inserted, and the single loop statement into which the command statement is inserted is parallelized in a cyclic manner.

  On the other hand, if the source program has multiple loops, the execution time can be greatly shortened by parallelizing the source program.

For example,
DO J = 1, N
DO I = 1, M
A (I, J) = A (I, J) + B (J) * C
ENDDO
ENDDO
When the double loop for executing the operation of the sparse matrix B is parallelized for the parallel processors of the two CPUs 1 and 2, a loop as shown in FIG. 8 is executed. Depending on the situation, the execution time can be greatly reduced.

  However, for example, if the first half value of the data B (J) of the sparse matrix B is all 0 and the second half value is other than 0, the CPU 2 performs the inner loop operation “N / 2”. However, the CPU 1 does not have to execute the inner loop operation at all.

  In this case, if parallelization is performed by a normal method (a method of dividing and parallelizing evenly) as shown in FIG. 7, the execution time is determined by the longest CPU execution time. If the execution time of the inner loop is T, the execution time is “T * (N / 2)”, and as a result, the execution time can be shortened only by the effect based on the number of CPUs.

  Therefore, in the compile processing device of the present invention, when multiple loop statements and single loop statements are parallelized, the loop statements are cyclically taken into consideration when the appearance position of 0 in the sparse matrix may be biased. Parallelization is made in the form of a method.

  In this cyclic parallelization, when parallelizing for the parallel processors of two CPUs 1 and 2, for example, J = 1, 3, 5,... .., 4, 6... Are executed by the CPU 2, and the other CPU is executed by dividing the loop every rotation.

  Assuming that the values of the first half of the data B (J) of the sparse matrix B are all 0 and the values of the second half are other than 0, the CPU 1 and the CPU 2 can execute the loop on average. Therefore, if the execution time of the inner loop is T, the execution time is “T * (N / 2) * (1/2)”, and parallelization is performed in a normal manner as shown in FIG. The execution time can be shortened compared to the above.

  When the appearance position of 0 in the sparse matrix is not biased, there is no difference between the execution time when parallelization is performed in the normal method and the execution time when parallelization is performed in the cyclic method. There is no penalty for parallelization using the cyclic method.

  As described above, according to the present invention, when a source program is compiled, when the source program has a loop for executing a sparse matrix operation, the execution time can be greatly reduced.

  Hereinafter, the present invention will be described in detail according to embodiments.

  FIG. 2 illustrates an embodiment of a compile processing apparatus 1 having the present invention.

  As shown in this figure, the compile processing apparatus 1 of the present invention includes a program input unit 10 for inputting a source program 2 and a program input unit 10 for generating an object program 3 by compiling the source program 2. An optimization processing unit 11 that executes the optimization processing of the input source program 2 and an object generation unit 12 that generates the object program 3 based on the optimization processing result of the optimization processing unit 11 are provided.

  The optimization processing unit 11 includes a detection unit 110, a determination unit 111, an insertion unit 112, and a parallelization unit 113 in order to realize the present invention.

  The detection unit 110 omits execution of a multiple loop statement described in the source program 2 when the innermost loop statement of the multiple loop statement includes a sparse matrix and the data of the sparse matrix is 0. Executes when a single-loop statement described in the source program 2 is detected and the single-loop statement includes a sparse matrix and the data of the sparse matrix is 0. A process of detecting an object having an operation that can be omitted is performed.

  The determination unit 111 determines, for the innermost loop statement of the multiple loop statement detected by the detection unit 110, whether the data of the sparse matrix of the operation of the innermost loop statement is data whose value is not updated by the rotation of the loop. For the single loop sentence detected by the detection unit 110, a process is performed to determine whether the data of the sparse matrix of the operation of the single loop sentence is data whose value is not updated by the rotation of the loop.

  The insertion unit 112 checks whether the data in the sparse matrix is 0 immediately before the innermost loop statement in which the determination unit 111 determines that the value is not updated. A sparse matrix immediately before a single loop statement in which a command statement instructing to proceed to the next rotation without performing an operation in the loop statement is inserted, or when the determination unit 111 determines that the data is not updated. If the data is 0, if it is 0, a process is performed in which an instruction that instructs to proceed to the next rotation is performed without performing an operation in the single loop sentence.

  The parallelizing unit 113 parallelizes the multiple loop statement in which the insertion unit 112 has inserted the command statement in the form of a cyclic method as shown in FIG. 3, or the single loop in which the insertion unit 112 has inserted the command statement. Processing is performed in which sentences are parallelized in the form of a cyclic method as shown in FIG.

  Here, among the matrices (arrays) included in the matrix described in the source program 2, which is a sparse matrix and which is not a sparse matrix, for example, information described in the source program 2 According to the input information from the programmer.

  FIG. 4 and FIG. 5 illustrate an embodiment of a processing flow executed by the optimization processing unit 11 to realize the present invention.

  Next, compile optimization processing realized by the present invention will be described according to this processing flow.

  When the optimization processing unit 11 receives the source program 2, in order to implement the present invention, as shown in the processing flow of FIGS. 4 and 5, first, in step 10, the source program 2 is unprocessed. When it is determined whether or not a loop statement remains, and when it is determined that an unprocessed loop statement remains, the process proceeds to step 11 where one unprocessed loop statement is described in the source program 2. select.

Subsequently, in step 12, when it is determined whether the selected loop statement is a multiple loop statement or a single loop statement, and it is determined that the selected loop statement is a multiple loop statement, the process proceeds to step 13 and the multiple A = A + B * C, A = A−B * C, A = A + B / C, A = A−B / C are operations on the sparse matrix B in the innermost loop statement of the loop statement.
Check whether there is an operation pattern.

  Here, these calculation patterns indicate calculation patterns that can be omitted when the data of the sparse matrix B is zero.

  Therefore, in this step 13, whether the innermost loop statement of the selected multiple loop statement includes a sparse matrix B and an operation that can be omitted when the data of the sparse matrix B is 0. Is examined.

  Subsequently, when it is determined in step 14 that the above-described operation pattern does not exist in the innermost loop statement of the selected multiple loop statement in accordance with the check processing in step 13, the operation that can be omitted is not included. Return to step 10 to process the next loop statement.

  On the other hand, when it is determined in step 14 that the above-described operation pattern exists, the process proceeds to step 15, and the operation data corresponding to the sparse matrix B described in the innermost loop statement is not detected even if the innermost loop is rotated. It is checked whether the data is invariant data (vector data) whose value is not updated.

  Subsequently, in step 16, when it is determined that the operation data corresponding to the sparse matrix B described in the innermost loop statement is not invariant data according to the check processing in step 15, the instruction statement described in step 17 is inserted. Then, in consideration of the fact that the execution time becomes longer due to the execution of the command statement, the process returns to step 10 to process the next loop statement.

  On the other hand, when it is determined in step 16 that the data is invariant, the process proceeds to step 17 to check whether the value of the operation data is 0 immediately before the innermost loop statement. Then, a command statement for instructing to proceed to the next rotation without performing the operation in the innermost loop statement is inserted.

That is, as shown by * 1 in FIG.
IF (B (k-1) .EQ .0) GOTO 10
Or as shown by * 2 in FIG.
IF (B (k) .EQ .0) GOTO 20
Is inserted.

  Subsequently, in step 18, the loop statement of the multiple loop selected in step 11 (the instruction statement is inserted according to the processing in step 17) is parallelized in a cyclic manner, and then the step to process the next loop statement is performed. Return to 10.

On the other hand, when it is determined in step 12 that the selected loop sentence is a single loop sentence, the process proceeds to step 19 where A = A + B * C, which is an operation on the sparse matrix B. , A = A−B * C, A = A + B / C, A = A−B / C
Check whether there is an operation pattern.

  Here, these calculation patterns indicate calculation patterns that can be omitted when the data of the sparse matrix B is zero.

  Therefore, in this step 19, it is checked whether or not the selected single loop statement includes a sparse matrix B and an operation that can be omitted when the data of the sparse matrix B is 0. .

  Subsequently, when it is determined in step 20 that the above-described operation pattern does not exist in the selected single loop statement in accordance with the check processing in step 19, since the operation that can be omitted is not included, the next loop statement Return to step 10 to process.

  On the other hand, when it is determined in step 20 that the above-described operation pattern exists, the process proceeds to step 21 and the operation data corresponding to the sparse matrix B described in the single loop statement is not detected even if the single loop is rotated. Check if the value is invariant data (scalar data) that is not updated.

  Subsequently, in step 22, when it is determined that the operation data corresponding to the sparse matrix B described in the single loop statement is not invariant data according to the check processing in step 21, the instruction statement described in step 23 is inserted. Then, in consideration of the fact that the execution time becomes longer due to the execution of the command statement, the process returns to step 10 to process the next loop statement.

  On the other hand, when it is determined in step 22 that the data is invariant, the process proceeds to step 23 to check whether the value of the operation data is 0 immediately before the single loop statement. Insert a command statement that indicates to proceed to the next rotation without performing the operation in the single loop statement.

  Subsequently, in step 24, the loop statement of the single loop selected in step 11 (the instruction statement is inserted according to the processing in step 23) is parallelized in a cyclic manner, and then the next loop statement is to be processed. Return to Step 10.

  When it is determined that all the loop statements described in the source program 2 have been processed in step 10 by repeating the processing in steps 10 to 24 in this way, the optimization processing for compilation according to the present invention is performed. Exit.

  As described above, the compile processing apparatus 1 according to the present invention is unnecessary by inserting a statement that makes it possible to omit execution of unnecessary operations in the source program 2 in consideration of the nature of a sparse matrix. This eliminates the need to execute computations, thereby reducing the execution time.

  The compile processing apparatus 1 according to the present invention, when the execution time is shortened by parallelizing the source program 2 in which this statement is inserted, when the occurrence position of 0 of the sparse matrix is biased. In consideration of the fact that the execution time may not be shortened, instead of performing parallel processing based on the normal equally dividing method, parallel processing based on the cyclic method is performed, As a result, when the execution time is shortened by parallelization, the execution time can be shortened more greatly.

  (Supplementary note 1) In a compile processing method for compiling a source program, for a multiple loop statement described in the source program, the innermost loop statement of the multiple loop statement includes a sparse matrix, and the data of the sparse matrix is 0 In the process of detecting an operation having an operation that can be omitted in the case of, and for the innermost loop statement of the detected multiple loop statement, the sparse matrix data of the operation possessed by the innermost loop statement is obtained by the rotation of the loop. It is 0 by checking whether the sparse matrix data is 0 immediately before the process of determining whether the data is not updated and immediately before the innermost loop statement that has determined that the data is not updated. In some cases, a process of inserting a command statement instructing to proceed to the next rotation without performing an operation in the innermost loop statement, and a multiple loop statement in which the command statement is inserted are cyclically selected. Further comprising a step of parallelism in the form of a method, the compilation process wherein.

  (Supplementary note 2) In the compiling method according to supplementary note 1, in the detection process, for the single loop statement described in the source program, the single loop statement includes a sparse matrix, and the sparse matrix In the process of detecting the one having an operation that can be omitted when the data is 0, and in the above determination process, the data of the sparse matrix of the operation that the single loop statement has is detected in the loop. It is determined whether or not the data is not updated by rotation, and in the above insertion process, whether or not the sparse matrix data is 0 immediately before the single loop statement that determines that the data is not updated. If the result of the check is 0, a command statement instructing to proceed to the next rotation without performing an operation in the single loop statement is inserted, and in the parallelization process, the command statement is inserted 1 Heavy To parallelize the-loop sentence in the form of a cyclic system, compilation method characterized.

  (Additional remark 3) In the compile processing apparatus which compiles a source program, about the multiple loop sentence described in the source program, the innermost loop sentence of the multiple loop sentence contains a sparse matrix, and the data of the sparse matrix is 0 The detection means for detecting an operation having an operation that can be omitted when the condition is, and the innermost loop statement of the multiple loop statement detected by the detection means, the data of the sparse matrix of the operation possessed by the innermost loop statement is a loop. Whether the data of the sparse matrix is 0 immediately before the innermost loop statement in which the determination means determines that the data is not updated by the rotation and the determination means determines that the data is not updated. Inserting means for inserting a command statement instructing to proceed to the next rotation without performing an operation in the innermost loop statement when the value is 0, and the inserting means In that it comprises a parallelism means to parallelize the multi-loop sentence by inserting the statement in the form of a cyclic system, compilation and wherein.

  (Additional remark 4) In the compile processing apparatus according to Additional remark 3, the detection means includes a sparse matrix in the single loop sentence described in the source program, and data of the sparse matrix When the value of 0 is 0, an object having an operation that can be omitted is detected, and the determination means has a loop of the sparse matrix data of the operation of the single loop statement detected by the detection means. The insertion means determines that the data in the sparse matrix is 0 immediately before the single loop statement in which the determination means determines that the data is not updated. If it is 0, an instruction statement instructing to proceed to the next rotation without performing an operation in the single loop statement is inserted. Insert a sentence To parallelize the singlet loop statement in the form of a cyclic system, compilation and wherein.

  (Additional remark 5) In the compile processing program which compiles a source program, about the multiple loop statement described in the source program, the innermost loop statement of the multiple loop statement contains a sparse matrix, and the data of the sparse matrix is 0 If there is a processing that detects an operation that can be omitted if it is, and for the innermost loop statement of the detected multiple loop statement, the sparse matrix data of the operation that the innermost loop statement has is the value of the rotation of the loop It is 0 by checking whether the sparse matrix data is 0 immediately before the process of determining whether the data is not updated and immediately before the innermost loop statement that has determined that the data is not updated. In this case, a process for inserting a command statement instructing to proceed to the next rotation without performing an operation in the innermost loop statement, and a multiple loop statement in which the command statement is inserted are Compilation program for executing a process of parallelization in the form of click system to a computer.

  (Supplementary note 6) In the compile processing program according to supplementary note 5, in the processing to detect, the single loop statement described in the source program includes a sparse matrix in the single loop statement, and the sparse matrix When the data is 0, an operation having an operation that can be omitted is detected, and in the above determination process, for the detected single loop statement, the sparse matrix data of the operation of the single loop statement is It is determined whether the data is not updated by rotation, and in the above insertion process, it is determined whether the sparse matrix data is 0 immediately before the single loop statement that is determined to be data not updated. If the result is 0, a command statement instructing to proceed to the next rotation without performing an operation in the single loop statement is inserted, and in the parallel processing, the command statement is inserted. To parallelize the singlet loop statement in the form of a cyclic system, compilation program characterized.

It is explanatory drawing of the source program for demonstrating this invention. It is an example of 1 embodiment of the compilation processing apparatus of this invention. It is explanatory drawing of the parallelization by a cyclic system. It is one Embodiment of the processing flow which an optimization process part performs. It is one Embodiment of the processing flow which an optimization process part performs. It is explanatory drawing of the command sentence insertion by this invention. It is explanatory drawing of the source program for demonstrating a prior art. It is explanatory drawing of the source program for demonstrating a prior art.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Compile processing apparatus 2 Source program 3 Object program 10 Program input part 11 Optimization process part 12 Object generation part 110 Detection part 111 Judgment part 112 Insertion part 113 Parallelization part

Claims (5)

In a compile processing method for compiling a source program,
For multiple loop statements described in a source program, the inner loop statement of the multiple loop statement contains a sparse matrix and has an operation that can be omitted when the data of the sparse matrix is 0 The process of
Regarding the innermost loop statement of the detected multiple loop statement, a process of determining whether the sparse matrix data of the operation of the innermost loop statement is data whose value is not updated by the rotation of the loop,
Check whether the sparse matrix data is 0 immediately before the innermost loop statement in which it is determined that the data is not updated. If it is 0, the operation in the innermost loop statement is not performed. Inserting a command to instruct to proceed to the next rotation,
A process of parallelizing a multiple loop statement in which the above-mentioned imperative statement is inserted in a cyclic manner,
A featured compilation method. The compile processing method according to claim 1,
In the detection process, the single loop statement described in the source program includes an operation that can be omitted when the single loop statement includes a sparse matrix and the data of the sparse matrix is 0. Detect things,
In the determination process, for the detected single loop statement, it is determined whether the sparse matrix data of the operation of the single loop statement is data whose value is not updated by the rotation of the loop,
In the above insertion process, it is checked whether the sparse matrix data is 0 immediately before the single loop statement in which it is determined that the data is not updated. Insert a statement to go to the next rotation without performing the operation in
In the process of parallelization, parallelizing a single loop statement in which the above-mentioned imperative sentence is inserted in a cyclic manner,
A featured compilation method. In a compilation processing device for compiling a source program,
For multiple loop statements described in a source program, the inner loop statement of the multiple loop statement contains a sparse matrix and has an operation that can be omitted when the data of the sparse matrix is 0 Detecting means for
Determination means for determining whether the data of the sparse matrix of the operation of the innermost loop statement is data whose value is not updated by the rotation of the innermost loop sentence of the multiple loop sentence detected by the detection means;
Immediately before the innermost loop statement in which the determination means has determined that the data is not updated, it is checked whether the sparse matrix data is 0. If it is 0, the operation in the innermost loop statement is performed. An insertion means for inserting a command statement instructing to proceed to the next rotation without performing
The insertion means comprises parallel means for parallelizing a multiple loop statement into which a statement is inserted in a cyclic manner;
A featured compilation processing device. The compile processing apparatus according to claim 3,
The detecting means includes a single loop statement described in a source program, the single loop statement including a sparse matrix, and an operation that can be omitted when the data of the sparse matrix is 0 Detect
The determination means determines, for the single loop sentence detected by the detection means, whether the sparse matrix data of the operation of the single loop sentence is data whose value is not updated by the rotation of the loop,
The inserting means checks whether the data of the sparse matrix is 0 immediately before the single loop statement in which the determining means determines that the data is not updated. Insert a statement that indicates to proceed to the next rotation without performing the operation in the loop statement,
The parallelizing means parallelizes the single loop statement in which the insertion means has inserted the imperative sentence in a cyclic manner.
A featured compilation processing device. In a compilation processing program that compiles a source program,
For multiple loop statements described in a source program, the inner loop statement of the multiple loop statement contains a sparse matrix and has an operation that can be omitted when the data of the sparse matrix is 0 Processing to
For the innermost loop statement of the detected multiple loop statement, a process for determining whether the data of the sparse matrix of the operation of the innermost loop statement is data whose value is not updated by loop rotation;
Check whether the sparse matrix data is 0 immediately before the innermost loop statement for which it is determined that the data is not updated. If it is 0, the operation in the innermost loop statement is not performed. Processing to insert a command to instruct to proceed to the next rotation,
A compiling process program for causing a computer to execute a process of parallelizing a multiple loop statement in which the above-mentioned imperative sentence is inserted in a cyclic manner.

Images