CN118332980A - Three-level coverage point type function coverage rate collection method for multifunctional IP core - Google Patents

Abstract

The present invention belongs to the field of software verification technology, and specifically relates to a three-level coverage point function coverage collection method for a multifunctional IP core, the steps of which include: reading the configuration data of the IP core to be collected, collecting coverage function data, and judging whether the functions to be covered have been covered; building a single position for the incoming register data configuration and related use cases, which are regarded as the first-level coverage points and named as function-level coverage points; analyzing the function-level coverage points, establishing the second-level cross coverage points; establishing the third-level combination-level coverage points; performing use case regression and coverage collection; checking the collection of function coverage and analyzing the uncovered function combinations, writing directional use case regression to collect coverage, and realizing full coverage of IP core design functions. The present invention can realize the function coverage collection of convolutional neural network hardware accelerators, standardize the function coverage collection steps, and improve the completeness of the function combinations involved in the function coverage.

Description

A three-level coverage point functional coverage collection method for multifunctional IP cores

Technical Field

The invention belongs to the technical field of software verification, and in particular relates to a three-level coverage point type function coverage collection method for a multi-functional IP core.

Background technique

With the vigorous development of the chip industry, in order to meet the development of intelligence in today's society, various semiconductor companies have designed many IP cores of AI chips with complex functions. These AI chip cores often have many functions. In terms of verification, verification methodology has flourished. In the field of verification, a complete verification process from chip specification analysis, verification point decomposition to platform construction has been proposed to ensure the correctness of the verified chip functions. However, among the many verification methodologies, including the mainstream UVM verification methodology, there are few references to how to build a functional coverage file to more completely collect functional coverage, and how to solve the problem of functional coverage without omissions in single functions, between two functions, and multi-function combinations.

In the design of functional coverage, taking the convolutional neural network hardware accelerator as an example, the IP core functions include whether the convolution, pooling, activation, padding and other functions are executed or not; the output mode includes broadcast output and non-broadcast output, whether to output to the storage unit, whether to output to the computing unit, etc.; the two functions of convolution and pooling involve the size and step length of the convolution kernel area, the size and step length of the sliding window; in terms of input, it involves boundary value coverage, etc. Functional coverage not only collects whether a single function is covered or not, but also collects all possible combinations between multiple functions to ensure the completeness of the designed IP. However, there is currently a lack of a standardized and step-by-step collection method for the functional coverage of complex IPs. For example, if a function bin inside the coverpoint is counted as a combination, the convolutional neural network hardware accelerator IP will have nearly 50,000 functional combinations (which will vary depending on the coverage model built by the verification personnel), and it is unimaginable to exhaustively collect each combination.

Summary of the invention

In view of the above deficiencies in the prior art, the present invention provides a three-level coverage point functional coverage collection method for multi-functional IP cores, which can realize the functional coverage collection of convolutional neural network hardware accelerators, standardize the functional coverage collection steps, and improve the completeness of the functional combinations involved in the functional coverage.

To achieve the above objectives, the present invention provides a three-level coverage point function coverage collection method for a multifunctional IP core, comprising the following steps:

S1. Read the configuration data of the IP core to be collected, collect coverage function data according to the register data configuration and related use cases, determine whether the functions to be covered have been covered, and pass the register data to the functional coverage model (that is, the three-level coverage point model established after S2-S4 steps);

S2. A single position is built for the incoming register data configuration and related use cases, which are regarded as the first-level coverpoints and named as function-level coverpoints;

S3, analyzing the function-level coverage points, combining the related function-level coverage points, performing coverage point crossover, and randomly crossovering the remaining unrelated coverage points to form the second-level crossover coverage points, which are named as related-level coverage points;

S4. Establish the third-level combination-level coverage points, cross-combine all the association-level coverage points, and collect the coverage of all combinations of IP core functions;

S5. Perform use case regression and coverage collection;

S6. Check the collection of functional coverage and analyze the uncovered functional combinations. Use the characteristics of the uncovered functional combinations and the random number of associated level coverage points to write a directed use case regression to collect coverage and achieve full coverage of the IP core design functions.

In the described S1, the configuration data is read in the comparison model or register model of the UVM verification platform. This is the data information preparation for the three-level coverage point type functional coverage model, which is the preliminary work preparation for the functional coverage model to build a warehouse.

In the above S2, building a warehouse is to establish an enabling warehouse in the coverage point, and the enabling warehouse is used to collect functional coverage. For a coverage point with a single function, only one enabling warehouse needs to be established for the entire coverage point. For a coverage point with multiple functions, multiple enabling warehouses are correspondingly established in the coverage point.

In S2, the function-level coverage points include:

Input eigenvalue coverage points, including boundary value bins, inner boundary value bins, and intermediate value bins; output eigenvalue coverage points, including boundary value bins, inner boundary value bins, and intermediate value bins; convolution enable coverage points; convolution mode coverage points; convolution kernel area coverage points, including 1*1 bins, 2*2 bins, 3*3 bins, and 5*5 bins; convolution kernel step coverage points, including step 1 bin, step 2 bins, step 3 bins, and step 5 bins; activation function enable coverage points; activation function mode coverage points; pooling enable coverage points; pooling mode coverage points; sliding window area coverage points; sliding window step coverage points; padding enable coverage points; padding size coverage points; output mode coverage points, including non-broadcast non-output storage bins and broadcast output storage bins; computing unit coverage points; storage unit coverage points.

Functional-level coverage points can be further expanded on the above basis according to the specific IP core design functions.

In the S3 described above, the association-level coverage points are formed by analyzing the functions of the function-level coverage points, and the mutually related function-level coverage points are cross-combined. After all the mutually related function-level coverage points are analyzed and cross-combined, the remaining unrelated function-level coverage points are formed into a group, and the function-level coverage points therein are randomly cross-combined.

When analyzing the mutual correlation between function-level cover points, if the functions of one or several function-level cover points can only be realized based on another function-level cover point, then these function-level cover points are determined to be mutually correlated.

In S3 and S4, weights are set for function-level coverage points and association-level coverage points. If the weight is set to 0, only bins are built without collecting coverage. For example, the weights of most "function-level" coverage points can be set to 0, because the second-level association-level coverage points will collect the bins built by the original function-level coverage points again, so there is no need to collect coverage at the bottom layer.

The analysis and calculation involved in the present invention can be performed by an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the above algorithm is implemented by executing the program by the processor.

The beneficial effects of the present invention are:

The present invention can realize the function coverage collection of convolutional neural network hardware accelerators or other large IPs, establishes a standardized function coverage collection step, makes the combination between functions more organized, improves the completeness of the function combination involved in the function coverage, checks the coverage collection status after the random use case runs for a period of time, and uses separate directional use cases to collect function combinations with a small number of regressions based on the collection status of the function coverage and the number of coverages, thereby achieving full coverage of the designed IP functions, solving the dilemma of being difficult to collect a certain function combination in completely random or blind randomness, and reducing the time consumption of function regression.

In terms of shortening the time, the random number of coverage points at the "association level" of the present invention can greatly reflect the common characteristics of the combinations that are not randomly selected, write targeted use case coverage, save aimless random time, and take into account the use case regression test time, shortening the time for collecting 100% functional coverage in the entire verification process.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of configuration data reading of the present invention;

Fig. 2 is a schematic diagram of the process of the present invention;

FIG. 3 is a hierarchical relationship diagram of an embodiment of the present invention.

Detailed ways

The embodiment of the present invention is further described below in conjunction with the accompanying drawings: As shown in FIG2 , a three-level coverage point function coverage collection method for a multifunctional IP core includes the following steps:

S1. Read the configuration data of the IP core to be collected, collect coverage function data according to the register data configuration and related use cases, determine whether the functions to be covered have been covered, and pass the register data to the functional coverage model;

S2. A single position is built for the incoming register data configuration and related use cases, which are regarded as the first-level coverpoints and named as function-level coverpoints;

S3, analyzing the function-level coverage points, combining the related function-level coverage points, performing coverage point crossover, and randomly crossovering the remaining unrelated coverage points to form the second-level crossover coverage points, which are named as related-level coverage points;

S4. Establish the third-level combination-level coverage points, cross-combine all the association-level coverage points, and collect the coverage of all combinations of IP core functions;

S5. Perform use case regression and coverage collection;

S6. Check the collection of functional coverage and analyze the uncovered functional combinations. Use the characteristics of the uncovered functional combinations and the random number of associated-level coverage points to write a directed use case regression to collect coverage until the functional coverage reaches 100%, thus achieving full coverage of the IP core design functions.

As shown in FIG. 1 , in S1 , configuration data is read in a comparison model or a register model of the UVM verification platform. In FIG. 1 , the functional coverage model is referred to as a coverage model, and the interface is a data interface for data transmission or communication.

In S2, building a warehouse is to establish an enabling warehouse within the coverage point. The enabling warehouse is used to collect functional coverage. For a coverage point with a single function, only one enabling warehouse needs to be established for the entire coverage point. For a coverage point with multiple functions, multiple enabling warehouses are established within the coverage point.

In S2, function-level coverage points include:

Input eigenvalue coverage points, including boundary value bins, inner boundary value bins, and intermediate value bins; output eigenvalue coverage points, including boundary value bins, inner boundary value bins, and intermediate value bins; convolution enable coverage points; convolution mode coverage points; convolution kernel area coverage points, including 1*1 bins, 2*2 bins, 3*3 bins, and 5*5 bins; convolution kernel step coverage points, including step 1 bin, step 2 bins, step 3 bins, and step 5 bins; activation function enable coverage points; activation function mode coverage points; pooling enable coverage points; pooling mode coverage points; sliding window area coverage points; sliding window step coverage points; padding enable coverage points; padding size coverage points; output mode coverage points, including non-broadcast non-output storage bins and broadcast output storage bins; computing unit coverage points; storage unit coverage points.

In S3, association-level coverage points are formed by analyzing the functions of function-level coverage points, and the mutually related function-level coverage points are cross-combined. After analyzing and cross-combining all mutually related function-level coverage points, the remaining unrelated function-level coverage points are grouped together, and the function-level coverage points therein are randomly cross-combined.

When analyzing the mutual correlation between function-level cover points, if the functions of one or several function-level cover points can only be realized based on another function-level cover point, then these function-level cover points are determined to be mutually correlated.

At the same time, during the analysis process, correlation can be further divided into strong correlation, weak correlation and no correlation.

In S3 and S4, weights are set for function-level cover points and association-level cover points.

As shown in Figure 3, in a simplified coverage collection process, function-level coverage points include:

Convolution enable coverage point, convolution mode coverage point, convolution kernel area coverage point, convolution kernel step size coverage point;

Pooling enabled coverage points, pooling mode coverage points, sliding window area coverage points, sliding window step size coverage points;

Output mode coverage points, calculation unit coverage points, storage unit coverage points;

Activation function enabled (activation enabled) coverpoint, activation function mode (activation mode) coverpoint, fill enabled coverpoint (fill 1 and fill 2 in Figure 3, that is, two fill enabled coverpoints of the same type, distinguished by 1 and 2).

Among them, the convolution kernel area coverage points include 1*1 bin, 2*2 bin, 3*3 bin, and 5*5 bin (the bin in Figure 3 is the bin).

When analyzing the function-level coverage points, for the convolution enable coverage points, convolution mode coverage points, convolution kernel area coverage points, and convolution kernel step coverage points, only when the enable bin conditions in the convolution enable coverage points are valid, the enable bins in the convolution mode coverage points, convolution kernel area coverage points, and convolution kernel step coverage points will collect the coverage of related functions. Therefore, these four coverage points are cross-combined to form association-level coverage points, and only when the enable bin of the convolution enable coverage point is valid, the function bin cross coverage in the other three coverage points is collected. If the convolution enable coverage point does not have an enable bin valid, the ignore_bins instruction is used to ignore the enable bin combination in the other three coverage points.

Similarly, the pooling-enabled coverage points, pooling-mode coverage points, sliding window area coverage points, and sliding window step coverage points can only collect coverage when the enabling bin conditions inside the pooling-enabled coverage points are met. Thus, these four function-level coverage points are cross-combined to form association-level coverage points.

Output mode coverage points, computation unit coverage points, and storage unit coverage points are cross-combined to form association-level coverage points.

The activation function enabled coverage points, activation function pattern coverage points, and padding enabled coverage points are weakly related items (i.e., the weakly associated items in Figure 3), and are also cross-combined to form association-level coverage points.

After the association-level coverage points are determined, all association-level coverage points are cross-combined to form combination-level coverage points, thereby collecting coverage for all combinations of IP core functions.

Claims (7)

1. The three-level coverage point type function coverage rate collection method for the multifunctional IP core is characterized by comprising the following steps of:

S1, reading configuration data of IP cores to be collected, collecting coverage rate function data according to register data configuration and related use cases, judging whether functions to be covered are covered or not, and transmitting the register data to a function coverage rate model;

s2, carrying out single bin building on the input register data configuration and related use cases, and regarding the single bin building as a first-stage coverage point, and naming the first-stage coverage point as a functional-stage coverage point;

S3, analyzing the functional level coverage points, combining the correlated functional level coverage points, intersecting the coverage points, randomly intersecting the remaining unassociated coverage points, and jointly forming a second-level intersecting coverage point which is named as a correlated level coverage point;

s4, establishing third-level combination level coverage points, and carrying out cross combination on all the associated level coverage points to realize coverage rate collection of all combinations of IP core functions;

s5, carrying out use case regression and coverage rate collection;

S6, checking the collection condition of the function coverage rate, analyzing the uncovered function combinations, and compiling the directional use case regression collection coverage rate by utilizing the characteristics of the uncovered function combinations and the random times of the association-level coverage points so as to realize the full coverage of the IP core design function.

2. The three-level coverage point type function coverage rate collection method for a multifunctional IP core according to claim 1, wherein: in the step S1, the configuration data reading is performed in a comparison model or a register model of the UVM verification platform.

3. The three-level coverage point type function coverage rate collection method for a multifunctional IP core according to claim 1, wherein: in the step S2, the bin is built in order to build an enabling bin in the coverage point, the enabling bin is used for collecting the coverage rate of the function, for the coverage point with single function, the whole coverage point only needs to build one enabling bin, and for the coverage point with multiple functions, a plurality of enabling bins are correspondingly built in the coverage point.

4. The three-level coverage point type function coverage rate collection method for a multifunctional IP core according to claim 3, wherein: in the step S2, the function level coverage point includes:

Inputting a characteristic value coverage point, wherein the characteristic value coverage point comprises a boundary value bin, a boundary inner side value bin and an intermediate value bin; outputting a characteristic value coverage point, wherein the characteristic value coverage point comprises a boundary value bin, a boundary inner side value bin and an intermediate value bin; convolution enables coverage points; convolutionally pattern coverage points; the convolution kernel area coverage points comprise 1*1 bins, 2 x 2 bins, 3*3 bins and 5*5 bins; the step coverage point of the convolution kernel comprises a step 1 bin, a step 2 bin, a step 3 bin and a step 5 bin; activating a function-enabled overlay point; activating function mode coverage points; pooling enabling overlay points; pooling pattern coverage points; sliding window area coverage points; sliding window step coverage points; filling the enabling overlay points; filling size coverage points; the output mode coverage point comprises a non-broadcast non-output storage bin and a broadcast output storage bin; calculating a unit coverage point; the memory cells cover the dots.

5. The three-level coverage point type function coverage rate collection method for a multifunctional IP core according to claim 1, wherein: in the step S3, the relevant level coverage points are formed by analyzing the functions of the functional level coverage points, the relevant functional level coverage points are combined in a crossing way, after all the relevant functional level coverage points are analyzed and combined in a crossing way, the rest functional level coverage points which are not relevant to each other form a group, and the functional level coverage points are combined in a random crossing way.

6. The three-level coverage point type function coverage rate collection method for a multifunctional IP core according to claim 5, wherein: in analyzing the interrelation between the function level coverage points, if the function of one or several function level coverage points is realized based on another function level coverage point, it is determined that the function level coverage points are interrelated.

7. The three-level coverage point type function coverage rate collection method for a multifunctional IP core according to claim 1, wherein: in the S3 and S4, weights are set for the function-level coverage points and the association-level coverage points.