WO2015032331A1

WO2015032331A1 - Opencl program compilation method and compiler

Info

Publication number: WO2015032331A1
Application number: PCT/CN2014/085885
Authority: WO
Inventors: 刘颖
Original assignee: 华为技术有限公司
Priority date: 2013-09-06
Filing date: 2014-09-04
Publication date: 2015-03-12
Also published as: CN104424009A; CN104424009B

Abstract

An Open Computing Language (OpenCL) program compilation method and compiler, said method comprising: obtaining OpenCL source files and determining a first data transmission mode for operation data defined in said source files; calculating the execution time of said operation data in said first data transmission mode and in second data transmission mode, said second data transmission mode being different from said first data transmission mode, and said execution time comprising the data transmission time of the operation data as well as the device program execution time; selecting the data transmission mode having a shorter execution time to serve as the compiled data transmission mode for said operation data when said source files are compiled; generating a compile execution code file according to said compiling data transmission mode. The invention effectively ensures program execution efficiency.

Description

OpenCL program compilation method and compiler

Technical field

The present application relates to the field of computer processing technologies, and more particularly to an OpenCL program compilation method and compiler.

Background technique

OpenCL (Open Computing Language) is the first general-purpose parallel programming open and free standard language for heterogeneous systems. It provides software developers with a unified programming environment for high-performance computing servers and desktops. Write efficient and lightweight code for computing systems, handheld devices, and more.

The OpenCL program is mainly divided into two parts: the device program and the host program. For example, when a heterogeneous system is composed of a CPU and a GPU, when the program running on the CPU is a host program, the program running on the GPU is a device program. The execution process of the OpenCL program mainly includes: the host program controls the data transfer from the host end to the device end, the device end executes the device program to process the data, and the host program control transfers the processed result data from the device end to the host end.

According to the execution process of the above OpenCL program, the main efficiency of the OpenCL program is the data transfer phase and the device program execution phase. Therefore, the OpenCL program provides two data transfer modes, namely, a copy mode and a map mode. The copy mode refers to copying data from the host memory to the device memory, or from the device memory to the host memory. Since the data needs to be copied and transferred in the system, the OpenCL program takes a long time in the data transfer phase in the copy mode, but When the device program is executed, because the data is already in the device memory, the device program execution phase takes a short time; the mapping mode means that during the data transfer phase, only the device memory to host memory mapping relationship is established, and the data is still located at the host. In the memory, the data transfer phase takes a short time, but when the device program is executed, it needs to access the data in the host memory, which causes the device execution phase to take a long time.

In the process of implementing the present invention, the inventor has found that in the prior art, in order to ensure the execution efficiency of the OpenCL program, it is usually determined by the technician according to different application scenarios of the system, and the hardware is flat. Taiwan and other features, choose the appropriate data transfer mode to write OpenCL program, but the existing way, the user subjectivity is greater, and can not effectively guarantee the execution efficiency of OpenCL program.

Summary of the invention

The application provides an OpenCL program compilation method and a compiler to solve the technical problem that the efficiency of the OpenCL program cannot be effectively guaranteed in the prior art.

To achieve the above objective, the present application provides the following technical solutions:

In a first aspect, an open computing language OpenCL program compilation method is provided, including:

Obtaining a source program file of the OpenCL program, and determining a first data transmission mode of the operation data defined in the source program file;

Calculating an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, the second data transmission mode is different from the first data transmission mode, and the execution consumption time includes The data transmission time of the operation data and the execution time of the device program;

Selecting, by the execution, a data transmission mode with a small consumption time as a compiled data transmission mode of the operation data when the source program file is compiled;

A compiled execution code file is generated in accordance with the compiled data transfer mode.

In a first possible implementation manner of the first aspect, the calculating a program execution consumption time of the operation data in the first data transmission mode and the second data transmission mode respectively includes:

Verifying whether the operation data is safe when the operation data is processed according to the second data transmission mode;

When the operation data is secure, calculating a program execution consumption time when the operation data is respectively in the first data transmission mode and the second data transmission mode.

In conjunction with the first possible implementation of the first aspect, a second possible implementation manner of the first aspect is further provided. When the first data transmission mode is a replication mode, the second data transmission mode is a mapping mode, the verifying that the operation data is processed according to the second data transmission mode, and whether the operation data is secure comprises:

Analyzing whether there is a write operation of the operation data by the host end during the execution of the program, and if not, determining that the operation data is safe according to the second data transmission mode;

When the first data mode is the mapping mode, and the second data transmission mode is the copy mode, the verifying the operation data is processed according to the second data transmission mode, and whether the operation data is secure comprises:

It is analyzed whether there is a write operation of the operation data by the device end during the execution of the program, and if not, it is determined that the operation data is safe according to the second data transmission mode.

In conjunction with the first aspect or any one of the foregoing possible implementation manners of the first aspect, a third possible implementation manner of the first aspect is further provided, where the first data transmission mode is a replication mode, When the second data transmission mode is the mapping mode; or, the first data transmission mode is a mapping mode, and the second data transmission mode is a replication mode;

The calculating the execution consumption time of the operation data in the first data transmission mode and the second data transmission mode respectively includes:

Calculating a data transmission time of the operation data in the copy mode according to the total data amount of the operation data and the data transmission rate;

Calculating a device program execution time of the operation data in the copy mode according to a total amount of memory accesses to the operation data and a memory access rate of the access device end during execution of the device program;

And a sum of a data transmission time calculated in the copy mode and a device program execution time as an execution time of the operation data in the copy mode;

Calculating and eliminating the time according to the mapping relationship between the host end and the device end, and calculating the data transmission time of the operation data in the mapping mode;

Calculating a device program execution time of the operation data in the mapping mode according to a total amount of memory accesses to the operation data and a memory access rate of the access host in the execution of the device program;

The data transfer time calculated in the mapping mode and the sum of device program executions are taken as the execution time of the operation data in the mapping mode.

In conjunction with the third possible implementation of the first aspect, a fourth possible implementation manner of the first aspect is further provided, where the total amount of memory access to the operation data is defined according to a source program file. The number of work items for the device program and the amount of memory access data for the unit work item are calculated.

In conjunction with the third possible implementation of the first aspect, a fifth possible implementation manner of the first aspect is further provided, where the data transmission rate, the memory access rate of the access device end, or the access host end The memory access rate is predetermined based on the hardware characteristics of the current heterogeneous system execution hardware platform.

In a second aspect, a compiler is provided, comprising:

a mode determining module, configured to acquire a source program file of the OpenCL program, and determine a first data transmission mode of the operation data defined in the source program file;

a calculation module, configured to calculate an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, where the second data transmission mode is different from the first data transmission mode, The execution consumption time includes a data transmission time of the operation data and a device program execution time;

And a mode selection module, configured to select a data transmission mode that consumes less time as a compiled data transmission mode of the operation data when the source program file is compiled.

A compiling module is configured to generate a compiled execution code file according to the compiled data transfer mode.

In the first possible implementation manner of the second aspect, the method further includes:

And a verification module, configured to verify whether the operation data is safe when the operation data is processed according to the second data transmission mode, and if so, trigger the calculation module.

In conjunction with the first possible implementation of the second aspect, a second possible implementation manner of the second aspect is further provided, where the verification module is specifically configured to: when the first data transmission mode is a replication mode, The second data transmission mode is a mapping mode, and analyzes whether there is a write operation of the operation data by the host end during the execution of the program, and if not, determining that the operation data is secure; or, when the first data mode is In the mapping mode, when the second data transmission mode is the replication mode, it is analyzed whether there is a write operation of the operation data by the device end during the execution of the program, and if not, the data security is determined.

With reference to the second aspect, or any one of the foregoing possible implementation manners, the first data transmission mode is a replication mode, and the second data transmission mode is a mapping mode; or, the first The data transmission mode is a mapping mode, and the second data transmission mode is a replication mode;

The calculation module includes:

a first transmission time calculation module, configured to calculate a data transmission time of the operation data in the replication mode according to the total data volume of the operation data and the data transmission rate;

a first execution time calculation module, configured to calculate a device program execution time of the operation data in the copy mode according to a total amount of memory accesses to the operation data and a memory access rate of the access device end during execution of the device program;

a first consumption time calculation module, configured to use a sum of a data transmission time calculated in the copy mode and a device program execution time as an execution consumption time of the operation data in the copy mode;

a second transmission time calculation module, configured to calculate and eliminate time according to a mapping relationship between the host end and the device end, and calculate a data transmission time of the operation data in the mapping mode;

a second execution time calculation module, configured to calculate a device program execution time of the operation data in the mapping mode according to a total amount of memory accesses to the operation data and a memory access rate of the access host during execution of the device program;

a second consumption time calculation module, configured to use the data transmission time calculated in the mapping mode and the sum of device program executions as the execution time of the operation data in the mapping mode

According to the foregoing technical solution, the embodiment of the present application provides an OpenCL program compiling method and a compiler, and the compiler obtains a source program file of the OpenCL program, and determines a definition in the source program file. a first data transmission mode of the operation data; calculating an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, selecting the data transmission mode in which the execution consumption time is small as the The compiled data transfer mode of the operation data when the source program file is compiled, and the compiled execution code file is generated according to the compiled data transfer mode. The OpenCL program compiled according to the embodiment of the present application can reduce the program execution time, improve the program execution efficiency, and can effectively ensure the execution efficiency in different heterogeneous systems.

DRAWINGS

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings to be used in the embodiments or the prior art description will be briefly described below. Obviously, the drawings in the following description are only It is an embodiment of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without any creative work.

FIG. 1 is a flowchart of an embodiment of an OpenCL program compiling method according to an embodiment of the present application;

2 is a flowchart of another embodiment of an OpenCL program compiling method according to an embodiment of the present application;

FIG. 3 is a flowchart of still another embodiment of an OpenCL program compiling method according to an embodiment of the present application;

FIG. 4 is a schematic structural diagram of an embodiment of a compiler according to an embodiment of the present disclosure;

FIG. 5 is a schematic structural diagram of another embodiment of a compiler according to an embodiment of the present disclosure;

FIG. 6 is a schematic structural diagram of a computing module in a compiler according to an embodiment of the present disclosure;

FIG. 7 is a schematic structural diagram of an embodiment of a computing device according to an embodiment of the present application.

detailed description

The technical solutions in the embodiments of the present application are clearly and completely described in the following with reference to the drawings in the embodiments of the present application. It is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present application without departing from the inventive scope are the scope of the present application.

One of the main ideas of the embodiments of the present application includes:

Compiling a source file of the OpenCL program, and determining a first data transmission mode of the operation data defined in the source program file; calculating the operation data in the first data transmission mode and the second data transmission mode, respectively Execution consumes time, selects the data transfer mode with less execution time as the compiled data transfer mode of the operation data when the source program file is compiled, and generates a compiled execution code file according to the compiled data transfer mode. The OpenCL program compiled according to the embodiment of the present application can reduce the program execution time, improve the program execution efficiency, and effectively ensure the execution efficiency in different heterogeneous systems.

FIG. 1 is a flowchart of an embodiment of an open operation language OpenCL program compiling method according to an embodiment of the present application, which may include the following steps:

101: Obtain a source program file of an OpenCL (Open Computing Language) program, and determine a first data transmission mode of the operation data defined in the source program file.

The OpenCL program is mainly divided into two parts: the host program and the device program (that is, the Kernel program). The host program runs on the host side and the device program runs on the device side. Host program control will operate on data Transfer from the host to the device, and transfer the data from the device to the host after processing the data. The device program is executed by the device side to complete the processing of the operation data.

The main factors affecting the execution efficiency of OpenCL programs are the data transfer phase and the device program execution phase. The OpenCL program provides two data transfer modes, copy mode and map mode.

Copy mode refers to copying operational data from host-side memory to device-side memory, or from device-side memory to host-side memory.

The mapping mode means that only the mapping relationship between device memory and host memory is established during the data transmission phase, and the operation data still exists in the host side memory.

If the data transmission mode of the operation data is the copy mode, the data transmission phase may take a long time, and the device program execution phase takes a short time; if the data transmission mode of the operation data is the mapping mode, the data transmission phase may take a short time, the device The program execution phase takes a long time.

Therefore, the replication mode is mainly applicable to application scenarios in which data is transmitted once and used multiple times. The mapping mode is mainly applicable to application scenarios with large data transmission volume and small amount of access.

The source program file is written by the user, but because the user's subjectivity is large and the user experience is high, the execution according to the data transfer mode of the operation data defined in the source program file cannot effectively guarantee the execution of the program execution. effectiveness.

Furthermore, the inventors have further discovered in the research that since OpenCL programs are well ported and can be executed in different heterogeneous systems, OpenCL programs are more efficient to execute in one heterogeneous system, but perform efficiency in another heterogeneous system. Not necessarily high.

Therefore, the inventor has changed the thinking mode in the process of implementing the present invention. Since the OpenCL program needs to be compiled during the execution process, the source program file is changed into a binary language that can be recognized by the computer. Therefore, the embodiment of the present application utilizes The compiler has improved the compilation process when compiling the source program files.

After the compiler obtains the source program file, the operation data defined in the source program file and the first data transmission mode of the operation data are determined by analyzing the source program file.

The first data transmission mode may be a copy mode or a mapping mode, and is known by a function defined in the source program file. For example, when the operation function corresponding to the operation data is clWriteBuffer, it indicates that the data transmission mode of the operation data is a copy mode, when corresponding The operation function is clEnqueueMapBuffer, which indicates that the data transfer mode of the operation data is the mapping mode.

For example, the source file of a section of OpenCL program includes:

Double h_A[65536]=...;

clWriteBuffer(d_A,65536*8,h_A[0],...);

It can be seen that the defined operation data is A, the data type is Double (double-precision floating-point type), and the data transmission mode is known as the copy mode according to the function clWriteBuffer.

The operation data defined in the source program file may include a plurality of processing operations performed by the embodiments of the present application for each operation data.

102: Calculate an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively.

The second data transmission mode is a data transmission mode different from the first data transmission mode. In the embodiment of the present application, when the first data transmission mode is the replication mode, the second data transmission mode is a mapping mode; the first data transmission When the mode is the mapping mode, the second data transmission mode is the copy mode.

In this embodiment, the execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, that is, the execution consumption time in the copy mode and the mapping mode, respectively, is calculated.

The execution consumption time includes a data transmission time of the operation data and a device program execution time.

In the copy mode, since the data needs to be actually transferred in the host side memory and the device side memory, the data transfer time is related to the total data amount of the operation data and the data transfer rate.

The total amount of data for this operational data can be known from the definition of operational data in the OpenCL source. The data transmission rate can be predetermined in conjunction with the hardware characteristics of the execution platform of the current heterogeneous system.

Since the data needs to be copied and transmitted on the host side and the device side in the copy mode, the data transfer time includes the time when the operation data is transferred from the host side memory to the device side memory, and the time from the device side memory to the host side memory, The secondary transmission time is approximately the same. Therefore, the data transmission time can be equal to twice the product of the total data amount of the operation data and the data transmission rate.

The device program execution time is related to the total amount of memory access to the operation data and the memory access rate to the device-side memory during the execution of the device program.

The total amount of memory access to the operational data can be calculated according to the number of work items of the device program and the amount of memory access data of the unit work item. The number of work items and the amount of memory access data per unit of work items can be passed to the source program file. The analysis was obtained.

In the mapping mode, the operational data is not actually transmitted between the host and the device, but is achieved by establishing a mapping relationship. The data transmission time in the mapping mode is determined by the time when the mapping relationship is established, which includes the mapping relationship establishment time and the mapping relationship elimination time. Generally, the mapping relationship elimination time is the same as the mapping relationship establishment time, so the data transmission time in the mapping mode can be equal to the mapping. The relationship is established or eliminated twice as much.

In the mapping mode, during the execution of the device program, the operation data in the memory of the host is accessed, so the execution time of the device program is related to the total amount of memory access to the operation data during the execution of the device program and the memory access rate of the host memory. .

Similarly, the total amount of memory access to the operational data can be calculated based on the number of work items of the device program and the amount of memory access data per unit of work.

The access rate of the device program to the device side memory and the memory access rate to the host side can be determined in advance in conjunction with the hardware characteristics of the current heterogeneous execution platform.

Thereby, the execution consumption time of the operation data in the copy mode and the map mode can be separately calculated.

103: Select the data transmission mode with less execution time as the compiled data transmission mode of the operation data when the OpenCL source program is compiled.

104: Generate a compiled execution code file according to the compiled data transfer mode.

When the execution consumption time of the operation data in the first data transmission mode and the second data transmission mode is separately calculated, the data transmission mode with less consumption time may be selected as the operation data when the source program file is compiled. Compiling the data transfer mode, so that when compiling, the corresponding compiled execution code file is generated according to the compiled data transfer mode, so that when the OpenCL program is executed, the operation data is transmitted and processed according to the selected compiled data transfer mode, which can be shortened. Execution time is spent and execution efficiency is improved.

In this embodiment, the source program file is acquired, and the first data transmission mode of the operation data defined in the source program is determined; and then the execution of the operation data in the first data transmission mode and the second data transmission mode respectively is respectively calculated. Taking time, selecting a data transfer mode with a small execution time as a compiled data transfer mode of the operation data when the source program file is compiled, thereby generating a compiled execution code file, so that when the program is run in the machine, the program can be executed. The selected compiled data transmission mode processes the operation data, shortens the execution time, can effectively improve the execution efficiency, and when the program is transplanted to another heterogeneous system, the embodiment of the present application is adopted. The program can determine that the operational data conforms to the data transmission mode of the heterogeneous system, thereby effectively ensuring the execution efficiency of the program in different heterogeneous systems.

FIG. 2 is a flowchart of another embodiment of an open computing language OpenCL program compiling method according to an embodiment of the present application, which may include the following steps:

201: Acquire a source program file of the OpenCL program, and determine a first data transmission mode of the operation data defined in the source program file;

202: Verify that the operation data is safe according to the second data transmission mode, and if yes, execute step 203, and if no, end the process.

The second data transmission mode is different from the first data transmission mode. For example, when the first data transmission mode is the replication mode, the second data transmission mode is a mapping mode; when the first data transmission mode is the mapping mode, the first The second data transmission mode is the copy mode.

In this embodiment, when compiling the source program file, it is required to determine whether the operation data is safe if it is processed according to the second data transmission mode of the first data transmission mode, and if it is not secure, the process data is directly ended. Process.

Whether the operation data is safe or not can be determined by judging whether the operation data is processed according to the second data transmission mode, and whether the operation of the program is erroneous, for example, whether the operation data is consistent on the host side and the device side.

As a possible implementation,

When the first data transmission mode is the replication mode, the second data transmission mode is a mapping mode, and the verifying the operation data is processed according to the second data transmission mode, and whether the operation data is secure may be:

It is analyzed whether there is a write operation of the operation data by the host end during the execution of the program, and if not, it is determined that the operation data is safe according to the second data transmission mode.

Since the first data transmission mode of the operation data is the copy mode, the host side and the device side respectively save the operation data, and if the program side performs the write operation on the operation data, the operation data will be operated with the device side. Inconsistent data.

Therefore, if the operation data is processed according to the mapping mode, the operation data exists only with the host side, and during the execution of the program, the operation data processed by the device side is consistent with the operation data of the host side, This will result in different processing than the operation data in the copy mode, making the operation data unsafe and an error in program execution.

When the first data mode is the mapping mode, and the second data transmission mode is the copy mode, the verifying the operation data is processed according to the second data transmission mode, and whether the operation data is secure may be:

Since the first data transmission mode of the operation data is the mapping mode, the operation data exists only with the host side. If the operation data is processed according to the copy mode, the operation data exists on both the host side and the device side. If the device side has a write operation on the operation data, the data of the device side will change, but the data of the host side will not be changed at the same time. This will result in inconsistent data between the device and the host. When the operation data is processed in the copy mode, it is not safe, and an error occurs in program execution.

Therefore, the operation flow of the embodiment is continued only when it is determined that the operation data is processed in accordance with the second data transmission mode.

Wherein, determining whether there is a write operation on the operation data on the host end or the device end, the data flow analysis technology can be used to analyze the definition and usage of the data in the OpenCL source program to determine whether there is a host end or a device end pair. Write operation of the operation data.

203: Calculate an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, where the first data transmission mode is different from the second data transmission mode, and the execution consumption time is Includes data transfer time and device program execution time.

The first data transmission mode may be a replication mode, and the second data transmission mode may be a mapping mode; or the first data transmission mode may be a mapping mode, and the second data transmission mode may be a replication mode. .

Therefore, the calculating the execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, may include:

The data transmission time of the operation data in the copy mode is calculated according to the total data amount of the operation data and the data transmission rate.

The data transmission rate may be represented by a unit data transmission consumption time, and the data transmission time may be equal to twice the total data amount of the operation data and the unit data transmission consumption time.

The device program execution time of the operation data is calculated according to the total amount of memory accesses to the operation data and the memory access rate of the access device end during execution of the device program.

The memory access rate of the access device may be determined according to the hardware characteristics of the current heterogeneous system program execution platform.

The total amount of memory access to the operational data may be equal to: the number of work items of the device program and the amount of memory access data per unit of work items.

The work item work-item is the smallest execution unit. The number of work items indicates how many units the computer is divided into. The amount of memory access data for each work item can be known according to the definition in the OpenCL source program. Analysis of flow analysis techniques.

The sum of the data transfer time calculated in the copy mode and the device program execution time is taken as the execution time of the operation data in the copy mode.

The time is established and eliminated according to the mapping relationship between the host end and the device end, and the data transmission time of the operation data in the mapping mode is calculated.

The mapping relationship establishment and elimination time can be predetermined based on the hardware characteristics of the heterogeneous system execution platform.

The device program execution time of the operation data is calculated according to the total amount of memory access of the operation data and the memory access rate of the access host during execution of the device program.

The memory access rate of the access host may be predetermined according to the hardware characteristics of the heterogeneous system execution platform. The total amount of memory access to the operational data may be equal to the product of the number of work items of the device program and the amount of memory access data per unit of work item.

The sum of the data transmission time calculated in the mapping mode and the device program execution time is taken as the execution time of the operation data in the mapping mode.

204: Select the data transmission mode with less execution time as the compiled data transmission mode of the operation data when the source program file is compiled.

205: Generate a compiled execution code file according to the compiled data transmission mode.

By calculating the execution consumption time when the operation data is processed according to the copy mode and the mapping mode, respectively, selecting a data transmission mode that consumes less time is used as the compiled data transmission mode of the operation data, and the compiled data transmission mode may be the first of the operation data. The data transmission mode, or a second data transmission mode different from the first data transmission mode.

By selecting a compiled data transfer mode that consumes less time, the OpenCL program can be run on the machine, reducing execution time and improving execution efficiency.

In this embodiment, the source file is obtained, and the first data transmission mode of the operation data defined in the source program is determined, and the operation data is verified. If it is processed according to the second data transmission mode, when it is secure, Calculating an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, and selecting a data transmission mode in which the consumption time is small as a compiled data transmission mode of the operation data at the time of compiling, according to which Generate a compiled execution code file, so that when the program is running in the machine, the operation data can be processed according to the selected compiled data transmission mode, the execution time is shortened, the execution efficiency can be effectively improved, and the program is transplanted to another heterogeneous When the system is executed, the technical solution of the embodiment of the present application can be used to determine that the operation data conforms to the data transmission mode of the heterogeneous system, thereby ensuring the execution efficiency of the program in different heterogeneous systems.

The following is a detailed description of the technical solution of the present application in conjunction with a practical application scenario. FIG. 3 is a flowchart of another embodiment of an open computing language OpenCL program compiling method according to an embodiment of the present application. In this embodiment, the source of the following OpenCL program is used. A fragment of the program file is an example:

It can be seen from the segment source file fragment that the operation data defined in the source program file segment includes the operation data A and the operation data B, and the data transmission mode is the copy mode. The following mainly introduces the operation data A as an example. For the operation data B, the processing procedure is similar to the operation data A, and will not be described again.

The method can include the following steps:

301: Acquire a source program file of the OpenCL program, and determine that the first data transmission mode of the operation data A defined in the source program file is a copy mode.

It can be seen from the code "clWriteBuffer(d_A, 65536*8, h_A[0], ...)" that the first data transmission mode of the operation data A is the copy mode.

302: Verify whether the operation data is safe according to the mapping mode, and if yes, execute step 103, and if no, end the process.

According to the data stream analysis technology, in the program, there is no write operation of the operation data A by the host side, and therefore the operation data A is secure.

303: Calculate a data transmission time and a device program execution time of the operation data A in the copy mode, and obtain an execution consumption time of the operation data A in the copy mode.

The data transfer time Ct1=Vt*St*2 of the operation data A in the copy mode.

Where Vt is the total data amount of the operation data A. It can be known from the above program that the quantity type of the operation data A is a double-precision floating-point type, occupies 8 bytes, and the vector length of the operation data A is 65536, therefore, the operation data A The total amount of data is 65536*8B (bytes).

St is the unit data consumption time to indicate the data transmission rate. In this embodiment, it is assumed to be 4 cycles/B (4 clock cycles per byte).

Since the data needs to be transmitted back and forth twice, the data transfer time in the copy mode can be 65536*8*4*2=4Mcycle.

The device program execution time Ca1=Va*Sab of the operation data A in the copy mode.

Wherein, Va is the total amount of memory access of the device program to the operation data A, Va=Ka*Nwi, wherein Nwi is the number of work item work-item, and the work item is the minimum execution unit, the work item can be known as For each data packet in the operational data, Nwi is 65536. Ka is the amount of memory access data for the unit work item of the device program, that is, the amount of memory access corresponding to each work item. It can be known from the program that when the device program is executed, the operation data B and the 1/4 operation data A are accessed once, so Va=1/4*65536*8=128 KB.

Since the operation data A is stored on the device side in the copy mode, the Sab refers to the unit data consumption time when the device program accesses the device-side memory, and is used to indicate the memory access rate to the device-side memory. Assume 4cycle/B.

The device program execution time in the copy mode is 128KB*4cycle/B=0.5Mcycle.

Thereby, the execution consumption time C1=Ct1+Ca1=4.5 Mcycle in the copy mode can be calculated.

304: Calculate a data transmission time of the operation data A in the mapping mode and a device program execution time, and obtain an execution consumption time of the operation data A in the mapping mode.

If the data transmission mode of the operation data A is changed to the mapping mode, it is only used to establish and eliminate the mapping relationship in the data transmission phase, and it is assumed that the time for establishing or eliminating the mapping relationship is tm=10Kcycle. Then, the data transmission time in the mapping mode is Ct2=2*tm=0.02 Mcycle.

In the mapping mode, the device program execution time Ca2 = Va * Sam.

As can be seen from the above description, Va = 128 KB.

When Sam accesses the host-side memory, the unit data consumes time, and the memory access rate to the host-side memory is assumed to be 16cycle/B.

Thus, the device program execution time of the mapping mode can be calculated as Ca2=128KB*16cycle/B=2Mcycle.

Then, the execution time in the mapping mode consumes time C2 = Ct2 + Ca2 = 2.022 Mcycle.

305: Compare the execution consumption time of the operation data A in the copy mode and the execution consumption time of the operation data A in the mapping mode.

306: Select to execute a mapping mode with a small consumption time as a compiled data mode of the operation data A at compile time.

307: Generate a compiled execution code file.

It can be known from the calculation result of step 304 that the execution consumption time in the mapping mode is small, so the mapping mode is the compiled data mode of the operation data A at compile time, so that when compiling, the data transmission mode of the operation data A will be operated. Make changes and generate a compiled execution code file corresponding to the mapping mode.

When the OpenCL program is run in the machine, the operation data A is processed in accordance with the mapping mode, so that the execution processing time of the operation data A can be reduced, and the program execution efficiency is improved.

And when the OpenCL program is migrated to another heterogeneous system, the operation data A can be determined according to the solution of the embodiment of the present application to conform to the data transmission mode of another heterogeneous system to ensure the program in the other heterogeneous system. effectiveness.

For the foregoing method embodiments, for the sake of brevity, they are all described as a series of action combinations, but those skilled in the art should understand that the present application is not limited by the described action sequence, because according to the present application, Some steps can be performed in other orders or at the same time. In the following, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

FIG. 4 is a schematic structural diagram of an embodiment of a compiler according to an embodiment of the present disclosure, where the compiler may include:

The mode determining module 401 is configured to acquire a source program file of the OpenCL program, and determine a first data transmission mode of the operation data defined in the source program file.

The first data transmission mode may be a copy mode or a mapping mode, as known by a function defined in the source program file.

The calculating module 402 is configured to calculate an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively.

The second data transmission mode is different from the first data transmission mode. When the first data transmission mode is the copy mode, the second data transmission mode is the mapping mode; when the first data transmission mode is the mapping mode, the second data transmission mode is the copy mode.

The mode selection module 403 is configured to select a data transmission mode with a small consumption time as a compiled data transmission mode of the operation data when the OpenCL source program is compiled.

The compiling module 404 is configured to generate a compiled execution code file according to the compiled data transmission mode.

When calculating the execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, the data transmission mode with less consumption time can be selected as the operation data of the OpenCL source program compiling time. Compiling the data transfer mode, so that when compiling, the corresponding compiled execution code file is generated according to the compiled data transfer mode, so that when the OpenCL program is executed, the operation data is transmitted and processed according to the selected compiled data transfer mode, and the execution can be shortened. It takes time to improve execution efficiency.

In this embodiment, when the compiler acquires the source program file for compiling, first determines a first data transmission mode of the operation data defined in the source program; and then separately calculates the operation data in the first data transmission mode and the second data respectively. The execution time in the transfer mode consumes time, and the data transfer mode in which the consumption time is small is selected as the compiled data transfer mode of the operation data at the time of compiling, whereby the compiled execution code file can be generated, so that when the program is run in the machine, The operation data can be processed according to the selected compiled data transmission mode, the execution time is shortened, the execution efficiency can be effectively improved, and the program is transplanted to another heterogeneous system, and the technical solution of the embodiment of the present application can be used. It is determined that the operational data conforms to the data transmission mode of the heterogeneous system, thereby ensuring the execution efficiency of the program in different heterogeneous systems.

FIG. 5 is a schematic structural diagram of another embodiment of a compiler according to an embodiment of the present disclosure, where the compiler may include:

The mode determining module 501 is configured to acquire a source program file of the OpenCL program, and determine a first data transmission mode of the operation data defined in the source program file.

The verification module 501 is configured to verify whether the operation data is safe when the operation data is processed according to the second data transmission mode.

As a possible implementation,

The verification module may be specifically configured to: when the first data transmission mode is a replication mode, and when the second data transmission mode is a mapping mode, analyze whether a host side writes the operation data during a program execution process. If not, determining that the operation data is safe according to the second data transmission mode; when the first data mode is a mapping mode, and the second data transmission mode is a replication mode, analyzing is performed during program execution Whether there is a write operation of the operation data by the device end, and if not, determining that the operation data is safe according to the second data transmission mode.

Wherein, determining whether the host side or the device end has a write operation on the operation data, the data stream analysis technology may be used to analyze the definition and usage of the data in the source program file to determine whether there is a host end or a device end pair operation. Data write operation.

The calculating module 502 is configured to calculate a program execution consumption time when the operation data is in the first data transmission mode and the second data transmission mode, respectively, when the verification module 501 verifies the operation data security.

The second data transmission mode is different from the first data transmission mode, and the execution consumption time includes a data transmission time of the operation data and a device program execution time.

Therefore, referring to FIG. 6, the computing module may specifically include:

The first transmission time calculation module 601 is configured to calculate a data transmission time of the operation data in the replication mode according to the total data amount of the operation data and the data transmission rate.

The first execution time calculation module 602 is configured to calculate a device program execution time of the operation data in the replication mode according to a total amount of memory accesses to the operation data and a memory access rate of the access device end during execution of the device program.

The total amount of memory accesses to the operation data is calculated according to the number of work items of the device program defined in the source program file and the amount of memory access data of the unit work item.

The total amount of memory access to the operational data may be equal to the product of the number of work items of the device program and the amount of memory access data per unit of work item.

The work item work-item is the smallest execution unit. The number of work items indicates how many units the computer is divided into. The amount of memory access data of each work item can be known according to the definition in the source program file. Analysis of flow analysis techniques.

The first consumption time calculation module 603 is configured to use the sum of the data transmission time calculated in the copy mode and the device program execution time as the execution consumption time of the operation data in the copy mode.

The second transmission time calculation module 604 is configured to calculate and eliminate the time according to the mapping relationship between the host end and the device end, and calculate the data transmission time of the operation data in the mapping mode.

The second execution time calculation module 605 is configured to calculate a device program execution time of the operation data in the mapping mode according to a total amount of memory accesses to the operation data and a memory access rate of the access host end during execution of the device program.

The second consumption time calculation module 606 is configured to use the sum of the data transmission time calculated in the mapping mode and the device program execution as the execution time of the operation data in the mapping mode.

The mode selection module 503 is configured to select a data transmission mode with a small consumption time as a compiled data transmission mode of the operation data when the OpenCL source program is compiled.

The compiling module 504 is configured to generate a compiled execution code file according to the compiled data transmission mode.

In this embodiment, the compiler obtains the source program file, and determines a first data transmission mode of the operation data defined in the source program file, and verifies the operation data. If it is processed according to the second data transmission mode, the security time is And calculating an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, and selecting a data transmission mode in which the consumption time is small as a compiled data transmission mode of the operation data at the time of compiling, According to this, the compiled execution code file can be generated, so that when the program is running in the machine, the operation data can be processed according to the selected compiled data transmission mode, the execution time is shortened, the execution efficiency can be effectively improved, and the program can be transplanted to another When performing in a heterogeneous system, the technical solution of the embodiment of the present application can be used to determine that the operational data conforms to the data transmission mode of the heterogeneous system, thereby ensuring the execution efficiency of the program in different heterogeneous systems.

The compiler described in the foregoing embodiment is applied to the computing device in a practical application. The computing device that deploys the compiler in the embodiment of the present application can implement the compilation of the source program file, and compile the source program file into a machine-recogable code. The data transfer mode with low consumption time can be compiled for the operation data defined in the source program file, so that the execution time of the program is reduced and the program execution efficiency is improved.

From the above description, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Therefore, referring to FIG. 7, the embodiment of the present application further provides a computing device, where the computing device includes at least a memory 701 and a processor 702 connected to the memory 701 via a bus.

The memory 701 stores a set of program instructions. The memory 701 may be a high speed RAM memory or a non-volatile memory such as at least one disk memory or the like.

The processor 702 is configured to invoke a program instruction stored by the memory 701, and perform the following operations:

The processor 702 may be a central processing unit CPU, or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

Optionally, the computing device can be used to execute any of the OpenCL program compilation methods shown in FIG. 1 to FIG. 2 provided by the embodiments of the present application.

The various embodiments in the present specification are described in a progressive manner, and each embodiment focuses on differences from other embodiments, and the same similar parts between the various embodiments may be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

Finally, it should also be noted that in this context, relational terms such as first and second are used merely to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply these entities. There is any such actual relationship or order between operations. Furthermore, the term "comprises" or "comprises" or "comprises" or any other variations thereof is intended to encompass a non-exclusive inclusion, such that a process, method, article, or device that comprises a plurality of elements includes not only those elements but also Other elements, or elements that are inherent to such a process, method, item, or device. An element that is defined by the phrase "comprising a ..." does not exclude the presence of additional equivalent elements in the process, method, item, or device that comprises the element.

For the convenience of description, the above devices are described separately by function into various units. Of course, the functions of each unit may be implemented in the same software or software and/or hardware when implementing the present application.

It will be apparent to those skilled in the art from the above description of the embodiments that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such understanding, the technical solution of the present application may be embodied in the form of a software product in essence or in the form of a software product, which may be stored in a storage medium such as a ROM/RAM or a disk. , an optical disk, etc., includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to perform the methods described in various embodiments of the present application or portions of the embodiments.

The above description of the disclosed embodiments enables those skilled in the art to make or use the application. Various modifications to these embodiments are obvious to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application is not limited to the embodiments shown herein, but is to be accorded the broadest scope of the principles and novel features disclosed herein.

Claims

An open computing language OpenCL program compiling method, which comprises:

Obtaining a source program file of the OpenCL program, and determining a first data transmission mode of the operation data defined in the source program file;

Calculating an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, the second data transmission mode is different from the first data transmission mode, and the execution consumption time includes The data transmission time of the operation data and the execution time of the device program;

Selecting, by the execution, a data transmission mode with a small consumption time as a compiled data transmission mode of the operation data when the source program file is compiled;

A compiled execution code file is generated in accordance with the compiled data transfer mode.
The method according to claim 1, wherein the calculating the program execution consumption time of the operation data in the first data transmission mode and the second data transmission mode respectively comprises:

Verifying whether the operation data is safe when the operation data is processed according to the second data transmission mode;

When the operation data is secure, calculating a program execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively.
The method according to claim 2, wherein when the first data transmission mode is a copy mode, the second data transmission mode is a mapping mode, and the verifying the operation data is performed according to the second data transmission Mode processing, whether the operation data is safe or not includes:

Analyzing whether there is a write operation of the operation data by the host end during the execution of the program, and if not, determining that the operation data is safe according to the second data transmission mode;

When the first data mode is the mapping mode, and the second data transmission mode is the copy mode, the verifying the operation data is processed according to the second data transmission mode, and whether the operation data is secure comprises:

It is analyzed whether there is a write operation of the operation data by the device end during the execution of the program, and if not, it is determined that the operation data is safe according to the second data transmission mode.
The method according to any one of claims 1 to 3, wherein, when the first data transmission mode is a copy mode, the second data transmission mode is a mapping mode; or the first data transmission mode When the mode is mapped, the second data transmission mode is a copy mode;

The calculating the execution consumption time of the operation data in the first data transmission mode and the second data transmission mode respectively includes:

Calculating a data transmission time of the operation data in the copy mode according to the total data amount of the operation data and the data transmission rate;

Calculating a device program execution time of the operation data in the copy mode according to a total amount of memory accesses to the operation data and a memory access rate of the access device end during execution of the device program;

And a sum of a data transmission time calculated in the copy mode and a device program execution time as an execution time of the operation data in the copy mode;

Calculating and eliminating the time according to the mapping relationship between the host end and the device end, and calculating the data transmission time of the operation data in the mapping mode;

Calculating a device program execution time of the operation data in the mapping mode according to a total amount of memory accesses to the operation data and a memory access rate of accessing the host end during execution of the device program;

The data transfer time calculated in the mapping mode and the sum of device program executions are consumed as execution time of the operation data in the mapping mode.
The method according to claim 4, wherein the total amount of memory accesses to the operation data is calculated according to the number of work items of the device program defined in the source program file and the amount of memory access data of the unit work item. get.
The method according to claim 4, wherein the data transmission rate, the memory access rate of the access device, or the memory access rate of the access host is predetermined according to hardware features of the current heterogeneous system execution hardware platform. of.
A compiler, comprising:

a mode determining module, configured to acquire a source program file of the OpenCL program, and determine a first data transmission mode of the operation data defined in the source program file;

a calculation module, configured to calculate an execution consumption time of the operation data in the first data transmission mode and the second data transmission mode, respectively, the second data transmission mode and the first data transmission The transmission mode is different, and the execution consumption time includes a data transmission time of the operation data and a device program execution time;

And a mode selection module, configured to select a data transmission mode that consumes less time as a compiled data transmission mode of the operation data when the source program file is compiled.

A compiling module is configured to generate a compiled execution code file according to the compiled data transfer mode.
The compiler according to claim 7, further comprising:

And a verification module, configured to verify whether the operation data is safe when the operation data is processed according to the second data transmission mode, and if so, trigger the calculation module.
The compiler according to claim 8, wherein the verification module is specifically configured to: when the first data transmission mode is a copy mode, and the second data transmission mode is a mapping mode, and the analyzing is performed during program execution. Whether there is a write operation of the operation data by the host side, if not, determining that the operation data is secure; or, when the first data mode is a mapping mode and the second data transmission mode is a copy mode, analysis During the execution of the program, whether there is a write operation of the operation data by the device side, and if not, the data is determined to be safe.
The compiler according to any one of claims 7 to 9, wherein the first data transmission mode is a copy mode, and the second data transmission mode is a mapping mode; or, the first data transmission The mode is a mapping mode, and the second data transmission mode is a copy mode;

The calculation module includes:

a first transmission time calculation module, configured to calculate a data transmission time of the operation data in the replication mode according to the total data volume of the operation data and the data transmission rate;

a first execution time calculation module, configured to calculate a device program execution time of the operation data in the copy mode according to a total amount of memory accesses to the operation data and a memory access rate of the access device end during execution of the device program;

a first consumption time calculation module, configured to use a sum of a data transmission time calculated in the copy mode and a device program execution time as an execution consumption time of the operation data in the copy mode;

a second transmission time calculation module, configured to calculate and eliminate time according to a mapping relationship between the host end and the device end, and calculate a data transmission time of the operation data in the mapping mode;

a second execution time calculation module, configured to calculate a device program execution time of the operation data in the mapping mode according to a total amount of memory accesses to the operation data and a memory access rate of the access host during execution of the device program;

And a second consumption time calculation module, configured to use the data transmission time calculated in the mapping mode and the sum of device program executions as the execution time of the operation data in the mapping mode.