A Binary Neural Network Accelerator Based on RISC-V

Abstract

The present invention relates to the field of binary neural networks. Specifically, it relates to a binary neural network accelerator based on RISC-V, which, by carefully considering the requirements of the target network architecture and its current task requirements, screens the RISC-V instruction set adapted to the execution of the current task of the target network within the specified accelerator. During the driving process of the adapted RISC-V instruction set for the execution of the current task of the target network, it effectively evaluates the communication guarantee ability of the specified accelerator for the execution of the current task of the target network. After the execution of the current task of the target network is completed, it effectively evaluates the quality guarantee ability of the specified accelerator for the execution of the current task of the target network. By integrating the communication guarantee ability and the quality guarantee ability, it evaluates the driving effect of the specified accelerator for the execution of the current task of the target network and gives feedback, helping to deeply understand the task driving performance and processing ability of the binary neural network accelerator based on RISC-V, and further enhancing the use safety and decision support ability of the accelerator.

Description

Technical Field

[0001] The present invention relates to the field of binary neural networks, and more particularly, to a binary neural network accelerator based on RISC-V. Background Art

[0002] RISC-V, as an open-source and modular instruction set architecture, has set off a wave of innovation in the field of computer architecture in recent years. Its open-source nature has not only attracted a large number of developers but also promoted innovation in chip design. Especially in the field of neural network accelerators, the introduction of RISC-V has provided new opportunities for the efficient execution of binary neural networks.

[0003] The prior art has designed a series of binary neural network accelerators based on the RISC-V architecture by combining the flexibility of the RISC-V architecture and the computational efficiency of binary neural networks. However, the analysis of the driving performance of existing accelerators during the execution process of binary neural network tasks still has limitations. Specifically: 1. Existing accelerators lack a personalized matching mechanism for the RISC-V instruction set for binary neural network execution tasks. An overly low RISC-V instruction set configuration may limit the performance of the accelerator when processing binary neural network tasks and fail to meet the requirements of real-time or high efficiency. An overly high RISC-V instruction set configuration may increase the risk of resource waste.

[0004] 2. The communication performance evaluation of existing accelerators for binary neural network task execution is limited to the overall communication parameter performance level, ignoring the careful consideration of the adaptive communication configuration parameter regulation performance of existing accelerators, resulting in the lack of reliability and comprehensiveness of the communication performance evaluation results.

[0005] 3. The quality performance evaluation of existing accelerators for binary neural network task execution is limited to the output result accuracy and actual acceleration degree, ignoring the resource allocation efficiency of existing accelerators for binary neural network task execution, resulting in the lack of integrity of the quality performance evaluation results. Summary of the Invention

[0006] In view of this, to solve the problems raised in the above background art, a binary neural network accelerator based on RISC-V is proposed.

[0007] The technical solution adopted by the present invention to solve its technical problems is: The present invention provides a binary neural network accelerator based on RISC-V, including: an adapted instruction set screening module, a communication guarantee ability evaluation module, a quality guarantee ability evaluation module, a task driving effect feedback module, and a cloud database.

[0008] The adaptation instruction set screening module is connected to the communication guarantee capability evaluation module, the communication guarantee capability evaluation module is connected to the quality guarantee capability evaluation module, the quality guarantee capability evaluation module is connected to the task-driven effect feedback module, and the cloud database is respectively connected to the communication guarantee capability evaluation module, the quality guarantee capability evaluation module, and the task-driven effect feedback module.

[0009] The adaptation instruction set screening module is used to import the target binary neural network configuration data and its current task requirement data into the specified accelerator, record the target binary neural network as the target network, and screen the adapted RISC-V instruction set for the current task execution of the target network from various RISC-V instruction sets configured in the specified accelerator.

[0010] The communication guarantee capability evaluation module is used to monitor the communication link status during the execution drive of the adapted RISC-V instruction set for the current task of the target network in real time, and evaluate the communication guarantee capability of the specified accelerator for the current task execution of the target network.

[0011] The quality guarantee capability evaluation module is used to analyze the recognition accuracy, acceleration degree, and resource efficiency of the specified accelerator for the execution drive of the current task of the target network after the current task of the target network is completed, and evaluate the quality guarantee capability of the specified accelerator for the current task execution of the target network.

[0012] The task-driven effect feedback module is used to comprehensively evaluate the communication guarantee capability and the quality guarantee capability, evaluate the drive effect of the specified accelerator for the current task execution of the target network, and give feedback.

[0013] The cloud database is used to store the reference communication guarantee capability evaluation index and the reference quality guarantee capability evaluation index planned for the task-driven effect evaluation standard, store the reasonable communication bandwidth utilization rate preset for the communication link during the execution of the binary neural network task, and store the content of the real label of the current task.

[0014] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) By carefully considering the requirements of the target network architecture and its current task requirements, and screening the adapted RISC-V instruction set for the current task execution of the target network from various RISC-V instruction sets configured in the specified accelerator, the present invention not only helps to achieve the optimal energy efficiency ratio of the accelerator RISC-V instruction set, but also ensures the high efficiency and guarantee of task execution.

[0015] (2) By carefully considering the overall communication quality coefficient and the overall communication maintenance coefficient of the specified accelerator for the execution drive of the current task of the target network, the present invention accurately evaluates the communication guarantee capability of the specified accelerator for the current task execution of the target network, which is beneficial to the continuous optimization of the communication quality during the accelerator task drive process, and further enhances the use safety and decision support ability of the accelerator.

[0016] (3) The present invention accurately evaluates the quality assurance ability of a specified accelerator for the current task execution of a target network by analyzing the recognition accuracy, acceleration degree, and resource efficiency of the specified accelerator for driving the execution of the current task of the target network, directly measures the driving performance and processing ability of the accelerator for the binary neural network task, and provides a basis for system optimization and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described with reference to the accompanying drawings. However, the embodiments in the drawings do not constitute any limitation to the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the following drawings.

[0018] Figure 1 It is a schematic diagram of the module connection of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] Referring to Figure 1 As shown, the present invention provides a binary neural network accelerator based on RISC-V, including: an adaptation instruction set screening module, a communication guarantee ability evaluation module, a quality guarantee ability evaluation module, a task driving effect feedback module, and a cloud database.

[0021] The adaptation instruction set screening module is connected to the communication guarantee ability evaluation module, the communication guarantee ability evaluation module is connected to the quality guarantee ability evaluation module, the quality guarantee ability evaluation module is connected to the task driving effect feedback module, and the cloud database is respectively connected to the communication guarantee ability evaluation module, the quality guarantee ability evaluation module, and the task driving effect feedback module.

[0022] The adaptation instruction set screening module is used to import the configuration data of the target binary neural network and its current task requirement data into the specified accelerator, record the target binary neural network as the target network, and screen the adaptation RISC-V instruction set for the current task execution of the target network from various RISC-V instruction sets configured in the specified accelerator.

[0023] Specifically, the specific analysis process of the adaptation instruction set screening module includes: extracting each configuration level and its corresponding various logical operation requirements and various data processing method requirements from the target network configuration data, and retrieving whether there are supported operation instructions for various logical operation requirements corresponding to each configuration level of the target network and supported operation instructions for various data processing method requirements in various RISC-V instruction sets of the specified accelerator configuration, so as to analyze the instruction coverage IC of various RISC-V instruction sets of the specified accelerator configuration for the target network architecture j , where j is the number of various RISC-V instruction sets of the specified accelerator configuration, and j = 1, 2,..., b.

[0024] Retrieve the execution log content of each task within the historical preset period of various RISC-V instruction sets of the specified accelerator, including the task difficulty level, accuracy time series change curve, speed time series change curve, and power consumption time series change curve, and screen the minimum accuracy value, minimum speed value, and maximum power consumption value per unit time among them, so as to statistically obtain the average accuracy limit y of each difficulty level task execution of various RISC-V instruction sets of the specified accelerator within the historical preset period jh , average speed limit v jh and average power consumption limit p jh , where h is the number of tasks at each difficulty level, h = 1, 2,..., l, extract the current task requirement data of the target network, including power consumption limit, accuracy limit, and speed limit, and compare and analyze the requirement matching degree RM of various RISC-V instruction sets of the specified accelerator configuration for the current task of the target network j .

[0025] Accumulate the requirement matching degree and the instruction coverage, and screen the RISC-V instruction set corresponding to the maximum accumulated value as the adapted RISC-V instruction set for the current task execution of the target network.

[0026] Specifically, the specific analysis process of the IC j includes: integrating the total number of types of logical operation requirements and the total number of types of data processing method requirements of the target network configuration level, calculating the supported instruction coverage rate for the logical operation requirements of the target network configuration level and the supported instruction coverage rate for the data processing method requirements in various RISC-V instruction sets of the specified accelerator configuration, and accumulating the two to obtain the instruction coverage of various RISC-V instruction sets of the specified accelerator configuration for the target network architecture.

[0027] It should be noted that the support instruction coverage rate of various RISC-V instruction sets for the logical operation requirements at the target network configuration level in the above-specified accelerator configuration is obtained by analyzing the ratio of the number of support instructions for the logical operation requirements at the target network configuration level in various RISC-V instruction sets of the specified accelerator configuration to the total number of types of logical operation requirements at the target network configuration level. Similarly, the support instruction coverage rate of various RISC-V instruction sets for the data processing method requirements at the target network configuration level in the specified accelerator configuration is obtained by analyzing the ratio of the number of support instructions for the data processing method requirements at the target network configuration level in various RISC-V instruction sets of the specified accelerator configuration to the total number of types of data processing method requirements at the target network configuration level.

[0028] Specifically, the RM j The specific analysis process includes: From the formula Analyze the quality scores of tasks at each difficulty level executed by various RISC-V instruction sets of the specified accelerator in the historical preset period, where y 0 、v 0 、p 0 Are respectively the preset reference accuracy value, reference speed value and reference power consumption value. Similarly, calculate the required quality score for the current task execution of the target network

[0029] According to the difficulty level corresponding to the current task of the target network, screen the quality scores Q jh′ Of the execution of each relatively high-difficulty level task in various RISC-V instruction sets of the specified accelerator in the historical preset period, the quality scores Q jh″ Of the execution of each relatively low-difficulty level task and the quality scores of the execution of tasks at the same difficulty level Where h′ is the number of each relatively high-difficulty level task, h′ = 1, 2,..., l′, h″ is the number of each relatively low-difficulty level task, h″ = 1, 2,..., l″, and thus analyze the reference quality score of various RISC-V instruction sets of the specified accelerator for the current task execution of the target network

[0030] It should be noted that the specific calculation formula for the reference quality score of various RISC-V instruction sets of the above-specified accelerator for the current task execution of the target network is Where l′ and l″ are the numbers of relatively high-difficulty level tasks and relatively low-difficulty level tasks respectively, and δ 1 、δ 2 、δ 3 Are respectively the preset reference value weights corresponding to the relatively high-difficulty level, relatively low-difficulty level and the same difficulty level.

[0031] Exemplarily, the above δ1 、 δ 2 、 δ 3 can specifically take values of 0.3, 0.3, and 0.4.

[0032] Obtain the matching degree of various RISC-V instruction sets of the specified accelerator configuration for the current task of the target network from the calculation model For the current task of the target network.

[0033] In the embodiments of the present invention, by carefully considering the requirements of the target network architecture and its current task requirements, among various RISC-V instruction sets of the specified accelerator configuration, the adapted RISC-V instruction sets for the execution of the current task of the target network are screened, which not only helps to achieve the optimal energy efficiency ratio of the accelerator RISC-V instruction set, but also ensures the efficiency and guarantee of task execution.

[0034] The communication guarantee ability evaluation module is used to monitor the communication link status in real time during the execution drive of the adapted RISC-V instruction set for the current task of the target network, and evaluate the communication guarantee ability of the specified accelerator for the execution of the current task of the target network.

[0035] Specifically, the specific analysis process of the communication guarantee ability evaluation module includes: respectively recording the communication bandwidth utilization rate change curve, communication delay change curve, packet loss rate change curve, and bit error rate change curve of the communication link during the execution drive of the adapted RISC-V instruction set for the current task of the target network, and analyzing the overall communication quality coefficient CQ of the specified accelerator for the execution drive of the current task of the target network.

[0036] Retrieve the time stamps of each regulation of the communication link configuration parameters during the execution drive of the adapted RISC-V instruction set for the current task of the target network recorded in the communication management log of the specified accelerator and detect their corresponding communication improvement effects, so as to analyze the overall communication maintenance coefficient CM of the specified accelerator for the execution drive of the current task of the target network.

[0037] Take the cumulative value of the overall communication quality coefficient and the overall communication maintenance coefficient as the evaluation index of the communication guarantee ability of the specified accelerator for the execution of the current task of the target network.

[0038] Specifically, the specific analysis process of the CQ includes: importing the communication bandwidth utilization rate change curve of the communication link during the execution drive of the adapted RISC-V instruction set for the current task of the target network into the matlab software to obtain its corresponding best fitting function, denoted as F(t), and collecting the start time point t 1 and the end time point t 2 of the execution of the current task of the target network. From the formula Analyze the communication bandwidth compliance coefficient of the communication link during the execution drive of the adapted RISC-V instruction set for the current task of the target network, where η0 The reasonable communication bandwidth utilization rate preset for the communication link in the execution process of the binarized neural network task stored in the cloud database.

[0039] Analyze and adapt the compliance coefficient χ of the communication delay of the communication link during the current task execution driving process of the target network for the RISC-V instruction set 2 , the compliance coefficient χ of the packet loss rate 3 and the compliance coefficient χ of the bit error rate 4 .

[0040] Similarly, where G(t), R(t), and Z(t) are the best fitting functions corresponding to the communication delay change curve, packet loss rate change curve, and bit error rate change curve of the communication link during the current task execution driving process of the target network for the adapted RISC-V instruction set, respectively, and σ 0 , ψ 0 , are the permitted communication delay threshold, permitted packet loss rate threshold, and permitted bit error rate threshold preset for the communication link in the execution process of the binarized neural network task stored in the cloud database, respectively.

[0041] From the formula CQ = χ l +χ 2 +χ 3 +χ 4 Obtain the overall communication quality coefficient of the specified accelerator for the current task execution driving of the target network.

[0042] Specifically, the specific analysis process of the CM includes: according to the communication bandwidth utilization rate, communication delay, packet loss rate, and bit error rate of each regulation timestamp of the communication link configuration parameters during the current task execution driving process of the target network for the adapted RISC-V instruction set, compare them with the reasonable communication bandwidth utilization rate, permitted communication delay threshold, permitted packet loss rate threshold, and permitted bit error rate threshold preset for the communication link in the execution process of the binarized neural network task, calculate the deviation ratios corresponding to the communication bandwidth utilization rate, communication delay, packet loss rate, and bit error rate of each regulation timestamp of the communication link configuration parameters during the current task execution driving process of the target network for the adapted RISC-V instruction set and accumulate them, take the accumulated value as the communication anomaly risk coefficient of each regulation timestamp of the communication link configuration parameters during the current task execution driving process of the target network for the adapted RISC-V instruction set, and further perform a ratio analysis with the warning threshold of the communication anomaly risk coefficient corresponding to the preset communication link configuration parameter regulation trigger to retrieve the regulation trigger compliance degree of each regulation timestamp.

[0043] It should be noted that the deviation ratio of the communication bandwidth utilization rate of each adjustment timestamp of the communication link configuration parameters during the execution drive of the current task of the target network by the above-mentioned RISC-V instruction set adaptation is obtained by calculating the absolute difference between the communication bandwidth utilization rate of each adjustment timestamp of the communication link configuration parameters during the execution drive of the current task of the target network by the RISC-V instruction set adaptation and the reasonable communication bandwidth utilization rate preset for the communication link during the execution of the binary neural network task, and further analyzing the ratio of the calculated absolute difference to the reasonable communication bandwidth utilization rate preset for the communication link during the execution of the binary neural network task.

[0044] The deviation ratio of the communication delay of each adjustment timestamp of the communication link configuration parameters during the execution drive of the current task of the target network by the above-mentioned RISC-V instruction set adaptation is obtained by calculating the difference between the communication delay of each adjustment timestamp of the communication link configuration parameters during the execution drive of the current task of the target network by the RISC-V instruction set adaptation and the permitted communication delay threshold preset for the communication link during the execution of the binary neural network task. If the calculated difference is non-positive, the communication delay deviation ratio is set to 0. If the calculated difference is positive, the ratio of the calculated difference to the permitted communication delay threshold preset for the communication link during the execution of the binary neural network task is further analyzed to obtain the deviation ratio.

[0045] The calculation processes of the bit error rate deviation ratio and the packet loss rate deviation ratio of each adjustment timestamp of the communication link configuration parameters during the execution drive of the current task of the target network by the RISC-V instruction set adaptation are the same as the calculation process of the communication delay deviation ratio.

[0046] For the communication bandwidth utilization rate change curve, communication delay change curve, packet loss rate change curve, and bit error rate change curve of the communication link during the execution drive of the current task of the target network by the RISC-V instruction set adaptation, mark the adjustment timestamps of the configuration parameters respectively, plan the combination of each inspection period after each adjustment timestamp, detect the relative improvement degrees of the communication bandwidth utilization rate, communication delay, packet loss rate, and bit error rate of the communication link for each inspection period combination after each adjustment timestamp, and analyze the relative communication improvement degree of each adjustment timestamp of the communication link configuration parameters during the execution drive of the current task of the target network by the RISC-V instruction set adaptation according to the preset influence weights corresponding to each inspection period combination.

[0047] It should be noted that the detection process of the relative improvement degrees of the communication bandwidth utilization rate, communication delay, packet loss rate, and bit error rate of each inspection period combination after the above-mentioned each regulation timestamp for the communication link is as follows: Obtain the average communication bandwidth utilization rate, average communication delay, average packet loss rate, and average bit error rate of the communication link of each inspection period combination corresponding to a certain regulation timestamp, compare them with the communication bandwidth utilization rate, communication delay, packet loss rate, and bit error rate of the communication link at this regulation timestamp, and take the communication delay reduction rate, packet loss rate reduction rate, and average bit error rate reduction rate of the communication link of each inspection period combination corresponding to this regulation timestamp after this regulation timestamp as the relative improvement degrees corresponding to the communication delay, packet loss rate, and bit error rate respectively, and take the deviation ratio reduction rate of the communication bandwidth utilization rate of the communication link of each inspection period combination corresponding to this regulation timestamp after this regulation timestamp and the reasonable communication bandwidth utilization rate preset for the communication link during the execution process of the binary neural network task as the relative improvement degree of the communication bandwidth utilization rate, so as to obtain the relative improvement degrees of the communication bandwidth utilization rate, communication delay, packet loss rate, and bit error rate of each inspection period combination after each regulation timestamp for the communication link.

[0048] It should also be noted that the relative communication improvement degree of each regulation timestamp of the communication link configuration parameters during the current task execution driving process of the target network for the above-mentioned adaptation of the RISC-V instruction set is obtained by accumulating the relative improvement degrees of the communication bandwidth utilization rate, communication delay, packet loss rate, and bit error rate of each inspection period combination after each regulation timestamp for the communication link, so as to obtain the overall relative improvement degree of each inspection period combination for the communication link after each regulation timestamp, and then accumulating the product of the overall relative improvement degree of each inspection period combination for the communication link and its corresponding preset influence weight, so as to obtain the relative communication improvement degree of each regulation timestamp of the communication link configuration parameters during the current task execution driving process of the target network for the adaptation of the RISC-V instruction set.

[0049] Take the product of the regulation trigger compliance degree and the relative communication improvement degree as the detection index of the communication improvement effect, collect the detection indexes of the communication improvement effect of each regulation timestamp of the communication link configuration parameters during the current task execution driving process of the target network for the adaptation of the RISC-V instruction set and calculate their average value, so as to obtain the overall communication maintenance coefficient of the specified accelerator for the current task execution driving of the target network.

[0050] In the embodiment of the present invention, by carefully considering the overall communication quality coefficient and the overall communication maintenance coefficient of the specified accelerator for the current task execution driving of the target network, accurately evaluating the communication guarantee ability of the specified accelerator for the current task execution of the target network is beneficial to the continuous optimization of the communication quality during the accelerator task driving process, and further enhances the use safety and decision support ability of the accelerator.

[0051] The quality assurance ability evaluation module is used to analyze the recognition accuracy, acceleration degree, and resource efficiency of the specified accelerator for driving the execution of the current task of the target network after the current task of the target network is completed, and evaluate the quality assurance ability of the specified accelerator for driving the execution of the current task of the target network.

[0052] Specifically, the specific analysis process of the quality assurance ability evaluation module includes: comparing the output label content driven by the specified accelerator for the execution of the current task of the target network with the true label content of the current task stored in the cloud database, and analyzing the precision ζ and recall φ of the specified accelerator for driving the execution of the current task of the target network, so as to calculate the recognition accuracy of the specified accelerator for driving the execution of the current task of the target network.

[0053] Simulate and simulate the execution of the current task of the target network without using the specified accelerator and record the task completion duration, and perform ratio analysis with the completion duration of the specified accelerator for driving the execution of the current task of the target network to obtain the acceleration degree of the specified accelerator for driving the execution of the current task of the target network.

[0054] Sort out the computing resource utilization rate ε 1 、storage resource occupancy rate ε 2 and the cumulative hardware power consumption μ of the specified accelerator for driving the execution of the current task of the target network, and analyze the resource efficiency of the specified accelerator for driving the execution of the current task of the target network from the formula where ε′ 1 、ε′ 2 、μ 0 are respectively the reasonable reference computing resource utilization rate, reasonable reference storage resource occupancy rate, and reasonable reference cumulative hardware power consumption preset by the specified accelerator for driving the binary neural network task.

[0055] Accumulate the recognition accuracy, acceleration degree, and resource efficiency of the specified accelerator for driving the execution of the current task of the target network to obtain the quality assurance ability evaluation index of the specified accelerator for driving the execution of the current task of the target network.

[0056] Through the embodiments of the present invention, by analyzing the recognition accuracy, acceleration degree, and resource efficiency of the specified accelerator for driving the execution of the current task of the target network, the quality assurance ability of the specified accelerator for driving the execution of the current task of the target network is accurately evaluated, directly measuring the driving performance and processing ability of the accelerator for the binary neural network task, and providing a basis for system optimization and decision-making.

[0057] Specifically, the calculation formula for the recognition accuracy of the specified accelerator for driving the execution of the current task of the target network is:

[0058] The task-driven effect feedback module is used to comprehensively evaluate the communication guarantee ability and quality guarantee ability, evaluate the driving effect of a specified accelerator for the current task execution of the target network, and provide feedback.

[0059] Specifically, the specific analysis process of the task-driven effect feedback module includes: extracting the communication guarantee ability evaluation index TX and the quality guarantee ability evaluation index ZL of the specified accelerator for the current task execution of the target network, and using the formula to evaluate the driving effect of the specified accelerator for the current task execution of the target network, where TX 0 and ZL 0 are respectively the reference communication guarantee ability evaluation index and the reference quality guarantee ability evaluation index planned for the task-driven effect evaluation standard of the cloud database storage task.

[0060] The cloud database is used to store the reference communication guarantee ability evaluation index and the reference quality guarantee ability evaluation index planned for the task-driven effect evaluation standard, store the reasonable communication bandwidth utilization rate, permitted communication delay threshold, permitted packet loss rate threshold, and permitted error code rate threshold preset for the communication link during the execution of the binary neural network task, and store the content of the real label of the current task.

[0061] The data sources in the cloud database of this embodiment are shown in Table 1 below.

[0062] Table 1 Detailed Explanation of Data Sources in the Cloud Database

[0063]

[0064]

[0065] The above content is only an example and explanation of the structure of the present invention. Those skilled in the art of this technology can make various modifications, supplements, or use similar methods to replace the specific embodiments described, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should fall within the protection scope of the present invention.

Claims

1. A RISC-V-based binary neural network accelerator, characterized in that: include: Adapt the instruction set screening module, import the target binary neural network configuration data and its current task requirement data to the specified accelerator, record the target binary neural network as the target network, extract each configuration level and its corresponding various logical operation requirements and various data processing method requirements from the target network configuration data, and search for various RISC-V instruction sets configured by the specified accelerator to see whether the target network configuration levels correspond to various logical operation requirements and various data processing method requirements Support operation instructions, so as to analyze the instruction coverage of various RISC-V instruction sets configured by the specified accelerator for the target network architecture. , The numbers of various RISC-V instruction sets configured for the specified accelerator. ; Retrieve the execution log content of various tasks in the historical preset period of various RISC-V instruction sets of the specified accelerator, including the task difficulty level, precision timing change curve, speed timing change curve and power consumption timing change curve, and filter the minimum precision value, minimum speed value and maximum power consumption value per unit time, so as to calculate the average precision limit of the execution of tasks of various difficulty levels in the historical preset period of various RISC-V instruction sets of the specified accelerator , average speed limit and average power consumption limit , is the number of tasks of each difficulty level, , extract the current task demand data of the target network, including power consumption limit, accuracy limit and speed limit, and compare and analyze the matching degree of various RISC-V instruction sets configured by the specified accelerator with the current task demand of the target network ; Will and Accumulate and select the RISC-V instruction set corresponding to the maximum accumulated value as the adapted RISC-V instruction set for the current task of the target network; The communication assurance capability evaluation module monitors the communication link status of the target network during the execution of the current task driven by the RISC-V instruction set in real time, and evaluates the communication assurance capability of the specified accelerator for the current task of the target network; The quality assurance capability evaluation module, after the current task of the target network is completed, analyzes the recognition accuracy, acceleration degree and resource efficiency of the designated accelerator for the current task of the target network, and evaluates the quality assurance capability of the designated accelerator for the current task of the target network; The task driving effect feedback module integrates communication assurance capabilities and quality assurance capabilities to evaluate the driving effect of the specified accelerator on the current task of the target network and provide feedback; The cloud database stores the reference communication assurance capability evaluation indicators and reference quality assurance capability evaluation indicators planned by the task-driven effect evaluation standard, stores the reasonable communication bandwidth utilization rate preset by the communication link during the execution of the binary neural network task, and stores the real label content of the current task.

2. A RISC-V based binary neural network accelerator according to claim 1, characterized in that: Said The specific analysis process includes: integrating the total number of logic operation requirement types and the total number of data processing method requirement types of the target network configuration level, calculating the support instruction coverage of various RISC-V instruction sets of the specified accelerator configuration for the target network configuration level logic operation requirements and the support instruction coverage of the data processing method requirements, and accumulating the two to obtain the instruction coverage of various RISC-V instruction sets of the specified accelerator configuration for the target network architecture.

3. The RISC-V based binary neural network accelerator according to claim 1, characterized in that: Said The specific analysis process includes: Analyze the quality scores of tasks of various difficulty levels within the historical preset cycle of various RISC-V instruction sets of the specified accelerator, including are the preset reference accuracy value, reference speed value and reference power consumption value respectively. Similarly, the required quality score of the target network’s current task execution is calculated. ; According to the difficulty level corresponding to the current task of the target network, the quality scores of the execution of tasks with relatively high difficulty levels, the quality scores of the execution of tasks with relatively low difficulty levels, and the quality scores of the execution of tasks with the same difficulty levels of various RISC-V instruction sets of the specified accelerator within the historical preset period are screened, so as to analyze the reference quality scores of the execution of various RISC-V instruction sets of the specified accelerator for the current task of the target network. ; By computational model Get the matching degree of various RISC-V instruction sets configured by the specified accelerator to the requirements of the current task of the target network.

4. The RISC-V based binary neural network accelerator according to claim 1, characterized in that: The specific analysis process of the communication assurance capability evaluation module includes: recording the communication bandwidth utilization change curve, communication delay change curve, packet loss rate change curve and bit error rate change curve of the communication link in the process of driving the current task of the target network by adapting the RISC-V instruction set, and analyzing the overall communication quality coefficient of the specified accelerator for the current task execution drive of the target network. ; Retrieve the adapted RISC-V instruction set recorded in the communication management log of the specified accelerator to adjust the timestamps of the communication link configuration parameters during the current task execution drive of the target network and detect the corresponding communication improvement effect, so as to analyze the overall communication maintenance coefficient of the specified accelerator for the current task execution drive of the target network ; The accumulated value of the overall communication quality coefficient and the overall communication maintenance coefficient is used as an evaluation indicator of the communication assurance capability of the designated accelerator for the current task of the target network.

5. The RISC-V-based binary neural network accelerator according to claim 4, characterized in that: Said The specific analysis process includes: importing the communication bandwidth utilization change curve of the communication link during the current task execution driving process of the target network adapted to the RISC-V instruction set into the matlab software to obtain its corresponding best fitting function, which is recorded as , collect the starting time point of the current task execution of the target network and end time point , according to the formula Analyze the communication bandwidth compliance coefficient of the communication link in the process of driving the current task execution of the target network by adapting the RISC-V instruction set, where A reasonable communication bandwidth utilization ratio preset for the communication link of the binary neural network task execution process stored in the cloud database; Analyze the communication delay compliance coefficient of the communication link during the current task execution drive process of the target network adapted to the RISC-V instruction set , Packet loss rate compliance coefficient and bit error rate compliance factor ; By formula Get the overall communication quality coefficient of the specified accelerator driven by the current task execution of the target network.

6. The RISC-V based binary neural network accelerator according to claim 4, characterized in that: Said The specific analysis process includes: according to the adapted RISC-V instruction set, the communication bandwidth utilization, communication delay, packet loss rate and bit error rate of each control timestamp of the communication link configuration parameters in the current task execution driving process of the target network are retrieved, and the control trigger compliance of each control timestamp is retrieved; The communication bandwidth utilization rate change curve, communication delay change curve, packet loss rate change curve and bit error rate change curve of the communication link in the process of driving the current task of the target network are respectively marked with the timestamps of each adjustment and control of the configuration parameters by the adapted RISC-V instruction set, and each inspection period combination after each adjustment and control timestamp is planned, and the relative improvement of the communication bandwidth utilization rate, communication delay, packet loss rate and bit error rate of the communication link in each inspection period combination after each adjustment and control timestamp is detected, and the relative communication improvement of each adjustment and control timestamp of the communication link configuration parameters in the process of driving the current task of the target network by the adapted RISC-V instruction set is analyzed according to the preset influence weights corresponding to each inspection period combination; The product of the control trigger compliance and the relative communication improvement is used as the communication improvement effect detection index. The communication improvement effect detection index of each control timestamp of the communication link configuration parameters in the process of driving the current task execution of the target network adapted to the RISC-V instruction set is collected and the average is calculated to obtain the overall communication maintenance coefficient of the specified accelerator for the current task execution drive of the target network.

7. The RISC-V based binary neural network accelerator according to claim 4, characterized in that: The specific analysis process of the quality assurance capability evaluation module includes: comparing the output label content of the specified accelerator under the current task execution drive of the target network with the real label content of the current task stored in the cloud database, and analyzing the accuracy of the specified accelerator under the current task execution drive of the target network. and recall , in order to calculate the recognition accuracy of the specified accelerator for the current task execution driver of the target network; The target network is simulated to execute the current task without the help of the specified accelerator and the task completion time is recorded. The ratio analysis is performed with the completion time of the specified accelerator for the target network's current task execution drive to obtain the acceleration degree of the specified accelerator for the target network's current task execution drive; The computing resource utilization, storage resource occupancy, and hardware cumulative power consumption of the specified accelerator driven by the current task execution of the target network are sorted out to analyze the resource efficiency of the specified accelerator driven by the current task execution of the target network; The recognition accuracy, acceleration degree, and resource efficiency of the designated accelerator driven by the current task execution of the target network are accumulated to obtain the quality assurance capability evaluation index of the designated accelerator for the current task execution of the target network.

8. The RISC-V based binary neural network accelerator according to claim 7, characterized in that: The calculation of the recognition accuracy of the designated accelerator for the current task execution drive of the target network is shown as follows: .

9. The RISC-V based binary neural network accelerator according to claim 7, characterized in that: The specific analysis process of the task-driven effect feedback module includes: extracting the communication guarantee capability evaluation index of the specified accelerator for the current task of the target network and quality assurance capability assessment indicators , according to the formula Evaluate the driving effect of the specified accelerator on the current task of the target network, where They are respectively the reference communication assurance capability evaluation indicators and the reference quality assurance capability evaluation indicators planned for the cloud database storage task-driven effect evaluation standard.

Images