CN115033434B - Method and device for calculating kernel performance theoretical value and storage medium - Google Patents

Abstract

The embodiment of the invention discloses a method, a device and a storage medium for calculating a theoretical value of kernel performance, wherein the method comprises the steps of obtaining a test case set, wherein the test case set is used for simulating different application scenes when the theoretical value of kernel performance is calculated; analyzing the content of each test case in the test case set to obtain each group of metadata, wherein different groups of metadata are hardware operations and related data executed on the kernel under the application scene simulated by different test cases; for each set of metadata: and carrying out theoretical performance calculation on the kernel performance according to the kernel configuration requirements and the group of metadata. The method is suitable for the scene of theoretical analysis of the performance of the kernel in the verification process of the processor chip, and can realize the accuracy, the high efficiency and the flexibility of theoretical performance calculation.

Description

A method, device and storage medium for calculating theoretical value of core performance

technical field

The invention relates to the field of system performance verification, in particular to a method, device and storage medium for calculating a theoretical value of kernel performance.

Background technique

Compared with CPU (central processing unit, central processing unit), GPGPU (General-purpose computing on graphics processing unit, general-purpose graphics processing unit), as a super-large-scale programmable chip, its advantages are mainly reflected in its extremely strong parallel computing capability and data throughput. Therefore, it is required to carefully design the internal pipeline of GPGPU from the performance point of view, and the effectiveness and feasibility of this architecture design need to be tested by performance verification in the front-end development stage of the chip.

The biggest difference between performance verification and functional verification is that performance verification requires a complete set of performance expectations and measurement methods to judge the effectiveness of verification, and the quality of this set of methods is the core component of performance verification.

At present, in the process of performance verification, the calculation of the theoretical value is manually filled in the form by different testers one by one according to the requirements of different test vectors, and then read the form through the script to extract the expected value and provide it to the system for subsequent processing. However, this theoretical value calculation method has the following disadvantages: based on a single test vector, it is scattered and unsystematic and has a large number of repetitive and invalid calculations; the level of developers is uneven and the form is too long and difficult to check, resulting in many omissions and errors; automation The degree is extremely low and unstable. Because the reading of the form cannot be located and tracked to the specific form page position, it is difficult to solve the error.

Contents of the invention

In view of this, an embodiment of the present invention provides a method, device and storage medium for calculating a theoretical value of core performance, so as to realize the accuracy, efficiency and flexibility of theoretical performance calculation in processor chip core performance verification.

In the first aspect, an embodiment of the present invention provides a method for calculating a theoretical value of kernel performance, including:

Obtain a test case set, where the test case set is used to simulate different application scenarios when calculating the theoretical value of kernel performance;

Analyze the content of each test case in the test case set to obtain each group of metadata, where different groups of metadata are the hardware operations and related data performed on the kernel in the application scenarios simulated by different test cases;

For each set of metadata: Calculate the theoretical performance of the kernel performance based on the kernel configuration requirements and this set of metadata.

Further, before performing theoretical performance calculation on kernel performance for each set of metadata: according to the kernel configuration requirements and this set of metadata, the method further includes:

Calculate the corresponding project according to this kernel performance theory, and read the corresponding project description to get the kernel configuration requirements.

Further, according to the kernel configuration requirements and this group of metadata, theoretical performance calculations are performed on the kernel performance, including:

Subfunctions for each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical value of the corresponding kernel module performance based on the kernel configuration requirements and this group of metadata;

After all the sub-function calls are completed, the performance theoretical value calculation results of each module of the kernel under the input of this group of metadata are summarized, and the performance bottleneck distribution and/or bandwidth occupancy are predicted in combination with the preset related requirements.

Further, before the subfunctions of each module of the kernel: before passing the set of metadata and kernel configuration requirements to the subfunctions, the method also includes: searching for the subfunctions of at least one module of the kernel related to the set of metadata;

Subfunctions for each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical performance value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata, including:

For the found subfunctions of each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical performance value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata.

Further, find subfunctions of at least one module of the kernel related to this group of metadata, including:

Pre-establish the mapping relationship between keywords of multiple kernel modules and multiple sub-functions;

Match this group of metadata to the keyword of each kernel module one by one. When the keyword of a kernel module is successfully matched, it is determined that the metadata belongs to the stimulus of the kernel module, and the subfunction corresponding to the keyword of the kernel module is the metadata A related sub-function.

In a second aspect, an embodiment of the present invention provides a device for calculating a theoretical value of kernel performance, including:

The test case set acquisition unit is used to obtain the test case set, wherein the test case set is used for the simulation of different application scenarios when calculating the theoretical value of the kernel performance;

The metadata generation unit is used to analyze the content of each test case in the test case set to obtain each group of metadata, wherein different groups of metadata are the hardware operations and related data performed on the kernel under the application scenarios simulated by different test cases;

The theoretical performance calculation unit is used for each set of metadata: to calculate the theoretical performance of the kernel performance according to the kernel configuration requirements and the set of metadata.

Further, the device also includes a kernel configuration requirement acquisition unit, configured to:

For each set of metadata in the theoretical performance calculation unit: according to the kernel configuration requirements and this set of metadata, before performing theoretical performance calculations on the kernel performance, calculate the corresponding items according to this kernel performance theory, and read the corresponding item description to get the kernel configuration Require.

Further, the theoretical performance calculation unit is used to calculate the theoretical performance of the kernel performance according to the kernel configuration requirements and this group of metadata, including:

Subfunctions for each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical value of the corresponding kernel module performance based on the kernel configuration requirements and this group of metadata;

After all the sub-function calls are completed, the performance theoretical value calculation results of each module of the kernel under the input of this group of metadata are summarized, and the performance bottleneck distribution and/or bandwidth occupancy are predicted in combination with the preset related requirements.

Further, the theoretical performance calculation unit is also used to: before the sub-functions of each module of the kernel: before passing this group of metadata and kernel configuration requirements to the sub-functions, find the information of at least one module of the kernel related to this group of metadata subfunction;

The theoretical performance calculation unit is used for the sub-functions of each module of the kernel: passing this group of metadata and kernel configuration requirements to the sub-functions, and calling the sub-functions to calculate the performance theoretical value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata calculations, including:

For the found subfunctions of each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical performance value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata.

Further, the theoretical performance calculation unit is used to find subfunctions of at least one module of the kernel related to this set of metadata, including:

Pre-establish the mapping relationship between keywords of multiple kernel modules and multiple sub-functions;

Match this group of metadata to the keyword of each kernel module one by one. When the keyword of a kernel module is successfully matched, it is determined that the metadata belongs to the stimulus of the kernel module, and the subfunction corresponding to the keyword of the kernel module is the metadata A related sub-function.

In the third aspect, the embodiment of the present invention also provides a computer-readable storage medium, the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more central processing units , so as to realize the method for calculating the theoretical value of the kernel performance described in the aforementioned first aspect.

In the technical solution provided by the embodiment of the present invention, the intelligent automatic processing method is used to replace the old manual script processing method, which greatly improves the stability and brings smaller maintenance requirements. The calculation of the theoretical value of the core performance is processed in layers, which improves the accuracy, convenience and versatility of the calculation.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a flow chart of a method for calculating a core performance theoretical value provided by an embodiment of the present invention;

FIG. 2 is an applicable architecture diagram of a method for calculating a theoretical value of kernel performance provided by an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a device for calculating a theoretical value of kernel performance provided by an embodiment of the present invention.

Detailed ways

Embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings.

It should be clear that the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

An embodiment of the present invention provides a method for calculating a theoretical value of core performance, which can be executed by a corresponding device for calculating a theoretical value of core performance, and the device can be integrated in an intelligent device deployed with a processor core. Referring to FIG. 1 , the method specifically includes the following steps 101-103.

Step 101. Obtain a test case set, wherein the test case set is used for simulating different application scenarios when calculating the theoretical value of kernel performance.

In this step, the kernel may be a central processing unit or a core chip in a GPGPU, which is internally composed of multiple modules (called kernel modules). A test case set is a group of test cases, and different test cases in this group of test cases are used to simulate different application scenarios when calculating the theoretical value of the same performance of the kernel. For the calculation of the same performance theoretical value of the kernel, it is necessary to form a set of test cases by combining various test configurations and kernel loads. For example, when calculating the theoretical value of the data throughput capacity of a certain module in the kernel, various Different data read commands are tested in combination with different configuration parameters of the module.

Step 102: Analyze the content of each test case in the test case set to obtain each group of metadata, wherein different groups of metadata are hardware operations and related data performed on the kernel under different application scenarios simulated by the test cases.

In this step, a set of metadata is obtained by analyzing the content of a test case in the test case set, and the set of metadata is the hardware operation and related data performed on the kernel in the application scenario simulated by the corresponding test case. Wherein, the parsing process may be realized by using existing text parsing technology, and details are not repeated here.

Step 103 , for each set of metadata: according to the kernel configuration requirements and this set of metadata, theoretical performance calculation is performed on the kernel performance.

In this step, a set of metadata corresponds to an application scenario and has its own kernel load information. The kernel is configured according to the kernel load information in the metadata, and after the configuration is completed, this group of metadata is input into the relevant modules of the kernel in sequence, and the performance output generated in the relevant modules is obtained. Preferably, the performance output generated by the group of metadata in different modules can be further summarized, and the performance bottleneck distribution and bandwidth occupancy can be predicted in combination with preset related requirements.

Exemplarily, according to the kernel configuration requirements and this set of metadata, theoretical performance calculations are performed on the kernel performance, including:

Subfunctions for each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical value of the corresponding kernel module performance based on the kernel configuration requirements and this group of metadata;

After all the sub-function calls are completed, the performance theoretical value calculation results of each module of the kernel under the input of this group of metadata are summarized, and the performance bottleneck distribution and/or bandwidth occupancy are predicted in combination with the preset related requirements.

Preferably, before the subfunctions of each module of the kernel: before passing the set of metadata and kernel configuration requirements to the subfunctions, the method further includes: searching for the subfunctions of at least one module of the kernel related to the set of metadata;

Subfunctions for each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical performance value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata, including:

For the found subfunctions of each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical performance value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata.

During specific implementation, searching for subfunctions of at least one module of the kernel related to this group of metadata may further include:

Pre-establish the mapping relationship between keywords of multiple kernel modules and multiple sub-functions;

Match this group of metadata to the keyword of each kernel module one by one. When the keyword of a kernel module is successfully matched, it is determined that the metadata belongs to the stimulus of the kernel module, and the subfunction corresponding to the keyword of the kernel module is the metadata A related sub-function.

Wherein, in the above-mentioned embodiment, the mapping relationship between keywords of multiple kernel modules and multiple sub-functions is established in advance, for example: the keyword of kernel module 1 corresponds to sub-function 1; the keyword of kernel module 2 corresponds to sub-function 2; the keyword of kernel module 3 corresponds to subfunction 3, and so on. Each kernel module has its own keyword, and the keyword is a description of the characteristics of the kernel module. After obtaining a set of metadata, match the set of metadata with the keywords of each kernel module one by one. When the keyword of a kernel module is successfully matched, it is determined that the set of metadata belongs to the incentive of the kernel module. The key of the kernel module The sub-function corresponding to the word is a sub-function related to this group of metadata. The sub-functions corresponding to the keywords of the unmatched kernel modules do not need to be called again. The theoretical value calculation results in this implementation mode are accurate and concise, which can avoid the need to call all sub-functions every time the theoretical value is calculated, which consumes a lot of resources. Disadvantages, there will be no problem of covering up key information due to too many output results.

On the basis of the above scheme, in the method provided by the embodiment of the present invention, for each set of metadata: according to the kernel configuration requirements and this set of metadata, before performing theoretical performance calculation on the kernel performance, it also includes

Step 100 , calculate the corresponding items according to the current kernel performance theory, and read the corresponding item description to obtain the kernel configuration requirements.

Different projects have specific kernel configuration requirements, which can be manually written by technicians in advance and stored in the form of project descriptions. In this step, you can read the corresponding project description according to this project to obtain the kernel configuration requirements. This method can improve the versatility of the embodiments of the present invention, and is applicable to various projects.

Based on the foregoing embodiments, a preferred embodiment is provided below by taking the processor type as GPGPU as an example.

In order to solve various problems of current technology and improve the efficiency and reliability of performance verification, this solution has made the following improvements and enhancements:

The method described in pure programming language (Python) is used to replace the original method of manually inputting Excel tables plus perl scripts to read and process;

Carry out two-layer modeling 1) The test case set model focuses on extracting the characteristic values and data of the test case itself and processes the output value generated after calling the core model (kernel theory model) for subsequent use 2) The core model covers each sub- The module models and provides a unified API for the test case set model to call.

Specifically, referring to the framework provided in FIG. 2 , a method for calculating a theoretical value of kernel performance includes the following steps 200-203.

Step 200, pre-store a plurality of test case sets used for calculating the theoretical value of GPGPU kernel performance.

In this step, multiple test case sets can be created in advance. Among them, different test case sets are used for the application scenario simulation when calculating the theoretical value of different performances of the GPGPU kernel, for example: the test case set a is used for the application scenario simulation when the theoretical value calculation of the GPGPU kernel performance The application scenario simulation for the calculation of the theoretical value of the GPGPU kernel performance 2, the test case set c is used for the application scenario simulation of the theoretical value calculation of the GPGPU kernel performance 3, etc. Among them, the performance of the GPGPU kernel includes: the receiving and transmitting rate of the front-end controller responsible for distributing threads, the execution rate of various operations of the mathematical operation execution module, the throughput capacity of the data throughput module, and so on. Application scenario simulation may include assigning specific loads and test configurations to the GPGPU kernel, and the loads and test configurations are usually different in different application scenarios.

During specific implementation, the content of each test case set includes: test case configuration, test case source code and execution instruction set. Among them, the source code of the test case is a part of the implementation code of an application scenario simulation in the calculation process of a performance theoretical value of the GPGPU kernel. This part of the implementation code can be realized by a high-level programming language, and typically includes the following code: Initialization of the theoretical value calculation process , GPGPU kernel load configuration, the low-level language implementation code used in the calculation of the theoretical value of the GPGPU kernel execution instructions in the execution instruction set file storage location, the storage location of the manipulated data required by the instruction register content, GPGPU Reading and writing of configuration registers necessary to perform specified operations, data acquisition during theoretical value calculations, etc.

The execution instruction set file is a low-level language implementation code collection file for all execution instructions of the GPGPU kernel. Wherein, the execution instruction may include: a read instruction, a write instruction, a shift instruction, an exclusive OR instruction, and the like.

Typically, the source code of the test case set of HLSL (High Level Shading Language, Advanced Rendering Language) is implemented by C++ programming, and the format is cpp file; the file format of the execution instruction set is SP3 (rendering instruction file format).

Step 201, the theoretical model of the test case set obtains the target test case set required for the calculation of the theoretical value of the GPGPU kernel performance in multiple pre-stored test case sets, and calls the theoretical model tool set to analyze the configuration and source code of each test case in the target test case set and the execution instruction set to obtain various sets of metadata necessary for the kernel theoretical model.

In this step 201, after the test case set theoretical model finds the target test case set required for the calculation of a theoretical value of a performance of the GPGPU kernel in a plurality of pre-stored test case sets, for each test case in the target test case set: Read the configuration, source code, and execution instruction set files of the test case, and perform text analysis on the read results to obtain the hardware operations and related data performed on the GPGPU kernel in the application scenario simulated by the test case, as the kernel theoretical model. A required set of metadata. Among them, for any test case in the target test case set, there is a set of metadata corresponding to the simulated application scenario, and the set of metadata includes:

The number of instructions represented by the low-level language implementation code for the execution instructions of the GPGPU kernel, the operation type of each instruction, the type and size of the operand, etc.;

The number of various storage units of the GPGPU core, the modules and interfaces through which instructions are executed, the total amount and type of data requests, etc.

Step 202 , the kernel theoretical model acquires the GPGPU kernel configuration requirements required for this calculation of the theoretical value of the GPGPU kernel performance.

Exemplarily, the GPGPU kernel configuration requirements may specifically include: parameters, data width of each internal port, number and configuration of different sub-modules, ideal peak capacity of each sub-module, and the like.

Step 203, Kernel theoretical model For each set of metadata input by the test case set theoretical model this time: according to this set of metadata and the obtained GPGPU kernel configuration requirements, call the theoretical model tool set to perform theoretical performance calculation and analysis, and obtain this set of elements Calculation results of theoretical values of GPGPU kernel performance under data input.

Wherein, the calculation result of the theoretical value may include: theoretical value, possible performance bottleneck distribution and bandwidth occupation. Specifically, call the theoretical model toolset for theoretical performance calculation and analysis, and obtain the calculation results of the theoretical value of the GPGPU kernel performance under the input of this group of metadata, including:

Sub-step 1, the kernel theoretical model searches the subfunction of at least one module of the GPGPU kernel related to this group of metadata in the theoretical model tool set;

Sub-step 2, for the sub-functions of each module of the found GPGPU kernel:

① Pass this group of metadata and GPGPU kernel configuration requirements to the sub-function;

②Call the sub-function to calculate the theoretical value of the corresponding GPGPU kernel module performance based on the received metadata of this group and the GPGPU kernel configuration requirements;

Among them, the calculation results show what kind of performance output this kind of stimulus (this set of metadata input by the theoretical model of the test case set and the obtained GPGPU kernel configuration requirements) will produce in its own module;

Sub-step 3. After all sub-function calls are completed, the top-level function summarizes the performance theoretical value calculation results of each module of the GPGPU kernel under the input of this group of metadata obtained in sub-step 2, and predicts the performance bottleneck distribution based on the preset related requirements and bandwidth occupancy are returned to the kernel theoretical model after converting dimensions and formats, and the kernel theoretical model is passed to the test case set theoretical model.

Optionally, after the test case set theoretical model obtains the analysis result of the core theoretical model, it calls a pre-customized GUI graphical user interface to display the analysis result. Wherein, the analysis results may include the following contents: calculation results of theoretical performance values of each module of the GPGPU kernel under different sets of metadata input; performance bottleneck distribution and bandwidth occupation.

After the improvement scheme given in the above example, it can solve the instability and low efficiency of the existing method and liberate manpower. More importantly, the new scheme uses modeling thinking to model the core of the chip from the perspective of theoretical value calculation. It brings higher accuracy, reduces double counting and brings wider application prospects.

Correspondingly, an embodiment of the present invention provides a device for calculating a theoretical value of core performance, which can be used to execute the method for calculating a theoretical value of core performance described in the embodiment of the present invention, and the device can be integrated in a processor core deployed in smart devices. Referring to Figure 3, the device specifically includes the following units:

The test case set obtaining unit 301 is used to obtain the test case set, wherein the test case set is used for simulation of different application scenarios when calculating the theoretical value of kernel performance;

The metadata generation unit 302 is used to analyze the content of each test case in the test case set to obtain each group of metadata, wherein different groups of metadata are hardware operations and related data performed on the kernel under the application scenarios simulated by different test cases;

The theoretical performance calculation unit 303 is configured to, for each set of metadata: perform theoretical performance calculation on the kernel performance according to the kernel configuration requirements and the set of metadata.

Further, the device for calculating the theoretical value of kernel performance provided by the embodiment of the present invention also includes a kernel configuration requirement acquisition unit 300, configured to:

For each set of metadata in the theoretical performance calculation unit: according to the kernel configuration requirements and this set of metadata, before performing theoretical performance calculations on the kernel performance, calculate the corresponding items according to this kernel performance theory, and read the corresponding item description to get the kernel configuration Require.

Further, the theoretical performance calculation unit 303 is used to perform theoretical performance calculations on the kernel performance according to the kernel configuration requirements and this group of metadata, including:

Subfunctions for each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical value of the corresponding kernel module performance based on the kernel configuration requirements and this group of metadata;

After all the sub-function calls are completed, the performance theoretical value calculation results of each module of the kernel under the input of this group of metadata are summarized, and the performance bottleneck distribution and/or bandwidth occupancy are predicted in combination with the preset related requirements.

Further, the theoretical performance calculation unit 303 is also used to: before the subfunctions of each module of the kernel: before passing the metadata and kernel configuration requirements of the group to the subfunctions, find at least one module of the kernel related to the metadata of the group subfunction of

The theoretical performance calculation unit is used for the sub-functions of each module of the kernel: passing this group of metadata and kernel configuration requirements to the sub-functions, and calling the sub-functions to calculate the performance theoretical value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata calculations, including:

For the found subfunctions of each module of the kernel: pass this group of metadata and kernel configuration requirements to the subfunction; call the subfunction to calculate the theoretical performance value of the corresponding kernel module based on the kernel configuration requirements and this group of metadata.

Further, the theoretical performance calculation unit 303 is used to find subfunctions of at least one module of the kernel related to this set of metadata, including:

Pre-establish the mapping relationship between keywords of multiple kernel modules and multiple sub-functions;

Match this group of metadata to the keyword of each kernel module one by one. When the keyword of a kernel module is successfully matched, it is determined that the metadata belongs to the stimulus of the kernel module, and the subfunction corresponding to the keyword of the kernel module is the metadata A related sub-function.

The device for calculating the theoretical value of kernel performance provided by this embodiment belongs to the same inventive concept as the foregoing method embodiments. For technical details not described in this embodiment, please refer to the relevant descriptions in the foregoing method embodiments, which will not be repeated here.

In addition, an embodiment of the present invention also provides a computer-readable storage medium, the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more central processing units to The method for calculating the theoretical value of the core performance described in the foregoing embodiments is realized.

It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is a relationship between these entities or operations. There is no such actual relationship or order between them. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

The term "and/or" in the embodiments of the present invention describes the association relationship of associated objects, indicating that there may be three relationships, for example, A and/or B, which may mean: A exists alone, A and B exist simultaneously, and B exists alone These three situations. The character "/" generally indicates that the contextual objects are an "or" relationship.

Each embodiment in this specification is described in a related manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments.

In particular, as for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for relevant parts, please refer to part of the description of the method embodiment.

For the convenience of description, the above devices are described by dividing their functions into various units/modules and describing them separately. Of course, when implementing the present invention, the functions of each unit/module can be implemented in one or more pieces of software and/or hardware.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented through computer programs to instruct related hardware, and the programs can be stored in a computer-readable storage medium. During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM) and the like.

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present invention. All should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (11)

1. A method for calculating a theoretical value of performance of a kernel, the method comprising:

acquiring a test case set, wherein the test case set is used for simulating different application scenes when kernel performance theoretical values are calculated;

analyzing the content of each test case in the test case set to obtain each group of metadata, wherein different groups of metadata are hardware operations and related data executed on the kernel under the application scene simulated by different test cases;

for each set of metadata: according to the kernel configuration requirement and the group of metadata, performing theoretical performance calculation on kernel performance to obtain a kernel performance theoretical value calculation result under the input of the group of metadata so as to finish calculation of theoretical values in the performance verification process;

the kernel is a core chip in a central processing unit or a general graphics processor, and the theoretical value calculation result comprises a theoretical value.

2. The method of claim 1, wherein, for each set of metadata: before performing theoretical performance calculation on the kernel performance according to the kernel configuration requirement and the group of metadata, the method further comprises:

and calculating corresponding items according to the kernel performance theory, and reading corresponding item description to obtain kernel configuration requirements.

3. The method of claim 1, wherein performing theoretical performance calculations on core performance based on core configuration requirements and the set of metadata comprises:

sub-functions for each module of the kernel: transferring the set of metadata and the kernel configuration requirements to the subfunctions; calling a sub-function to calculate a performance theoretical value of a corresponding kernel module based on the kernel configuration requirement and the group of metadata;

after all the subfunctions are called, the calculation results of the performance theoretical values of all the modules of the kernel under the input of the metadata of the group are summarized, and the performance bottleneck distribution and/or bandwidth occupation are predicted by combining the preset related requirements.

4. A method according to claim 3, wherein, in the case of sub-functions for each module of the kernel: before passing the set of metadata and kernel configuration requirements to the sub-function, the method further comprises: searching a sub-function of at least one module of the kernel related to the group of metadata;

sub-functions for each module of the kernel: transferring the set of metadata and the kernel configuration requirements to the subfunctions; calling the subfunction to calculate the performance theoretical value of the corresponding kernel module based on the kernel configuration requirement and the group of metadata comprises the following steps:

sub-functions for each module of the found kernel: transferring the set of metadata and the kernel configuration requirements to the subfunctions; and calling the subfunction to calculate the performance theoretical value of the corresponding kernel module based on the kernel configuration requirement and the group of metadata.

5. The method of claim 4, wherein finding a sub-function of at least one module of the kernel associated with the set of metadata comprises:

pre-establishing mapping relations between keywords of a plurality of kernel modules and a plurality of sub-functions;

and matching the group of metadata with the keywords of each kernel module one by one, and determining that the metadata belongs to the excitation of one kernel module when the key of the kernel module is successfully matched with the keywords of the kernel module, wherein the sub-function corresponding to the key of the kernel module is a sub-function related to the metadata.

6. A core performance theory value calculation apparatus, characterized in that the apparatus comprises:

the test case set acquisition unit is used for acquiring a test case set, wherein the test case set is used for simulating different application scenes when the kernel performance theoretical value is calculated;

the metadata generation unit is used for analyzing the content of each test case in the test case set to obtain each group of metadata, wherein different groups of metadata are hardware operations and related data executed on the kernel under the application scene simulated by different test cases;

theoretical performance calculation unit for, for each set of metadata: according to the kernel configuration requirement and the group of metadata, performing theoretical performance calculation on kernel performance to obtain a kernel performance theoretical value calculation result under the input of the group of metadata so as to finish calculation of theoretical values in the performance verification process;

the kernel is a core chip in a central processing unit or a general graphics processor, and the theoretical value calculation result comprises a theoretical value.

7. The apparatus according to claim 6, further comprising a kernel configuration requirement acquisition unit configured to:

for each set of metadata at the theoretical performance calculation unit: and before theoretical performance calculation is carried out on the kernel performance according to the kernel configuration requirements and the group of metadata, corresponding items are calculated according to the kernel performance theory, and corresponding item descriptions are read to obtain the kernel configuration requirements.

8. The apparatus of claim 6, wherein the theoretical performance calculation unit is configured to perform theoretical performance calculation on the kernel performance according to the kernel configuration requirement and the set of metadata, and comprises:

sub-functions for each module of the kernel: transferring the set of metadata and the kernel configuration requirements to the subfunctions; calling a sub-function to calculate a performance theoretical value of a corresponding kernel module based on the kernel configuration requirement and the group of metadata;

after all the subfunctions are called, the calculation results of the performance theoretical values of all the modules of the kernel under the input of the metadata of the group are summarized, and the performance bottleneck distribution and/or bandwidth occupation are predicted by combining the preset related requirements.

9. The apparatus of claim 8, wherein the theoretical performance calculation unit is further configured to: in the sub-functions for each module of the kernel: before the group of metadata and the kernel configuration requirement are transferred to the subfunction, searching the subfunction of at least one module of the kernel related to the group of metadata;

the theoretical performance calculation unit is used for carrying out the subfunction of each module of the kernel: transferring the set of metadata and the kernel configuration requirement to a subfunction, and calling the subfunction to calculate a corresponding kernel module performance theoretical value based on the kernel configuration requirement and the set of metadata, wherein the method comprises the following steps:

sub-functions for each module of the found kernel: transferring the set of metadata and the kernel configuration requirements to the subfunctions; and calling the subfunction to calculate the performance theoretical value of the corresponding kernel module based on the kernel configuration requirement and the group of metadata.

10. The apparatus of claim 9, wherein the theoretical performance calculation unit is configured to find a sub-function of at least one module of the kernel associated with the set of metadata, comprising:

pre-establishing mapping relations between keywords of a plurality of kernel modules and a plurality of sub-functions;

and matching the group of metadata with the keywords of each kernel module one by one, and determining that the metadata belongs to the excitation of one kernel module when the key of the kernel module is successfully matched with the keywords of the kernel module, wherein the sub-function corresponding to the key of the kernel module is a sub-function related to the metadata.

11. A computer readable storage medium storing one or more programs executable by one or more central processing units to implement the method of any of claims 1-5.

Images