Systolic array fault recovery circuit, chip and accelerator

By dividing the protection column and the redundant column in the pulsating array, and using the comparator and point multiplication calculation module for fault location and recovery, the problems of large hardware overhead and low fault coverage in the prior art are solved, and the failure of the pulsating array is realized Efficient positioning and recovery.

CN118519816BActive Publication Date: 2024-11-22PEKING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410971198.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-22
Estimated Expiration
2044-07-19

AI Technical Summary

Technical Problem

The existing pulsating array fault recovery method has high hardware overhead, low fault coverage, and cannot independently realize fault location and fault recovery of the processing unit.

Method used

The protection column and the redundant column are divided in the pulsating array, and the convolution calculation results of the redundant column and the protection column are compared through the comparator module when the redundant column signal is high, the target address information of the fault processing unit is generated, and the vector is used. The storage module obtains the target feature vector and weight vector, and restores the convolution calculation results through the point multiplication calculation module.

Benefits of technology

Reduce hardware overhead, improve fault coverage, and realize independent fault location and recovery of processing units.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118519816B_ABST
    Figure CN118519816B_ABST
Patent Text Reader

Abstract

The present application relates to the field of systolic array technology, and provides a systolic array fault recovery circuit, chip and accelerator. The circuit includes: a processing unit for performing convolution calculation on received feature vectors and weight vectors; a comparator module for comparing the redundant column convolution calculation results of this round with the corresponding protection column convolution calculation results, and if the comparison results are inconsistent, generating the target address information of the fault processing unit; a vector re-storage module for obtaining the target feature vector and target weight vector used in the convolution calculation of the fault processing unit in this round according to the target address information; a dot product calculation module for performing dot product calculation on the target feature vector and the target weight vector, so that the dot product calculation result is used to recover the convolution calculation result of the fault processing unit in this round. The present application can reduce hardware overhead, improve fault coverage, and independently and simultaneously realize fault location and fault recovery of the processing unit.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of systolic array technology, and in particular to a systolic array fault recovery circuit, chip and accelerator. Background Art

[0002] The systolic array is one of the most typical architectures in CNN (Convolutional Neural Network) accelerators. Its architecture is a network of processing units that rhythmically calculate and transmit data. The biggest advantage of using a systolic array is that it improves data reuse and saves IO bandwidth. Therefore, under the constraint of the "memory wall", that is, the performance bottleneck caused by the mismatch between the performance of the computing unit and the storage unit, the systolic array is one of the most common choices for increasing computing speed and saving power consumption.

[0003] CNNs are usually robust to a certain extent, but if the computing units of the systolic array fail, the computing accuracy of the CNN accelerator will be greatly lost.

[0004] The current pulsating array fault recovery method requires additional processing units, but when a processing unit fails, the additional processing units are used to replace the faulty processing units. On the one hand, the additional processing units added by this method increase the area of ​​the original pulsating array, resulting in a large hardware overhead; on the other hand, the number of additional processing units added by this method is limited. When there are many faulty processing units, it is impossible to recover all faulty processing units, resulting in a low fault coverage rate; on the third hand, this method needs to cooperate with other mechanisms to locate the faulty processing unit, so that the additional processing units are used to replace the faulty processing units at the corresponding positions, that is, this method cannot independently and simultaneously realize the fault location and fault recovery of the processing units. In summary, the current pulsating array fault recovery method has a large hardware overhead, a low fault coverage rate, and cannot independently and simultaneously realize the fault location and fault recovery of the processing units. Summary of the invention

[0005] Embodiments of the present application provide a systolic array fault recovery circuit, chip and accelerator to solve the technical problems of the current systolic array fault recovery method having large hardware overhead, low fault coverage and inability to independently and simultaneously realize fault location and fault recovery of processing units.

[0006] In a first aspect, an embodiment of the present application provides a systolic array fault recovery circuit, comprising: a systolic array, a comparator module, a vector re-storage module, and a dot product calculation module;

[0007] The systolic array is an array formed by a plurality of processing units, the systolic array includes a protection column and a corresponding redundant column, and the protection column is adjacent to the corresponding redundant column;

[0008] The processing unit is used to perform convolution calculation on the received feature vector and weight vector, and output the convolution calculation results to the comparator module by column; wherein the feature vector flows horizontally along the systolic array, the weight vector flows vertically along the systolic array, and the weight vector received by the protection column and its corresponding redundant column is the same;

[0009] The comparator module is used to compare the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level, and if the comparison results are inconsistent, generate the target address information of the fault processing unit to which the inconsistent comparison result belongs, and output the target address information to the vector re-storage module;

[0010] The vector re-storage module is used to obtain the target feature vector and target weight vector used by the fault processing unit in the current round of convolution calculation from the received and stored feature vectors and weight vectors according to the target address information, and output the target feature vector and the target weight vector to the dot product calculation module;

[0011] The dot product calculation module is used to perform dot product calculation on the target feature vector and the target weight vector, so that the dot product calculation result is used to restore the convolution calculation result of the fault processing unit in this round.

[0012] In one embodiment, it further includes: a vector input module;

[0013] The vector input module is used to:

[0014] The feature vector and the weight vector are stored, wherein the protected weight vector in the weight vector is redundantly stored according to a redundant protection configuration;

[0015] The feature vector and the weight vector are loaded into the systolic array; wherein the same protected weight vector is loaded into a protection column of the systolic array and a corresponding redundant column thereof, respectively.

[0016] In one embodiment, the comparator module includes a register unit, a comparator unit and a fault address unit;

[0017] The register unit is used to output the convolution calculation result of the protection column to the comparator unit;

[0018] The comparator unit is used to compare the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level, and if the comparison results are inconsistent, generate the target address information of the fault processing unit to which the inconsistent comparison result belongs, and output the target address information to the fault address unit;

[0019] The fault address unit is used to store the target address information and output the target address information to the vector re-storage module.

[0020] In one embodiment, the vector re-storage module includes a vector transformation unit, a feature vector storage unit and a weight vector storage unit;

[0021] The vector transformation unit is used to perform position transformation on the received feature vector and weight vector so that the feature vector and weight vector after position transformation match the calculation logic of the dot product calculation module;

[0022] The feature vector storage unit is used to store and output the feature vector after position transformation;

[0023] The weight vector storage unit is used to store and output the weight vector after position transformation.

[0024] In one embodiment, the processing unit includes a first portion and a register and a second portion and a register;

[0025] The plurality of processing units of the systolic array are arranged alternately in columns as first-type processing units and second-type processing units;

[0026] The first type processing unit is used to:

[0027] Receive a first feature vector to be processed output by a previous processing unit in the same row and a weight vector to be processed output by a previous processing unit in the same column;

[0028] Performing a convolution calculation on the first feature vector to be processed and the weight vector to be processed to obtain a first convolution calculation result;

[0029] The first convolution calculation result is saved to the first part and register of the first part and the second part and register of the second part;

[0030] Outputting the first part of itself and the first convolution calculation result in the register to the comparator module;

[0031] The second type processing unit is used for:

[0032] Receive a second feature vector to be processed output by a previous processing unit in the same row and a weight vector to be processed output by a previous processing unit in the same column;

[0033] Performing a convolution calculation on the second feature vector to be processed and the weight vector to be processed to obtain a second convolution calculation result;

[0034] The second convolution calculation result is saved to the first part and register of the second convolution operation and the second part and register of the second convolution operation;

[0035] Outputting the second part of itself and the second convolution calculation result in the register to the comparator module;

[0036] The first feature vector to be processed and the second feature vector to be processed are the same feature vector at an initial moment.

[0037] In one embodiment, the processing unit further includes a first feature vector register, a second feature vector register, a weight vector register, and a convolution calculator;

[0038] The first feature vector register in the first type processing unit is used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row, and output the first feature vector to be processed to the first feature vector register in the next processing unit in the same row;

[0039] The second feature vector register in the first type processing unit is used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row, and output the second feature vector to be processed to the convolution calculator and the second feature vector register in the next processing unit in the same row;

[0040] The weight vector register in the first type of processing unit is used to receive the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and output the weight vector to be processed to the convolution calculator and weight vector register in the next processing unit in the same column;

[0041] The convolution calculator in the first type of processing unit is used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row, and the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and perform convolution calculation on the first feature vector to be processed and the weight vector to be processed to obtain a first convolution calculation result;

[0042] The first part and register and the second part and register in the first type processing unit are used to receive the first convolution calculation result, wherein the first part and register are also used to output the first convolution calculation result to the comparator module.

[0043] In one embodiment, the processing unit further includes a first feature vector register, a second feature vector register, a weight vector register, and a convolution calculator;

[0044] The first feature vector register in the second type processing unit is used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row, and output the first feature vector to be processed to the convolution calculator and the first feature vector register in the next processing unit in the same row;

[0045] The second feature vector register in the second type processing unit is used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row, and output the second feature vector to be processed to the second feature vector register in the next processing unit in the same row;

[0046] The weight vector register in the second type processing unit is used to receive the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and output the weight vector to be processed to the convolution calculator and weight vector register in the next processing unit in the same column;

[0047] The convolution calculator in the second type processing unit is used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row, and the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and perform convolution calculation on the second feature vector to be processed and the weight vector to be processed to obtain a second convolution calculation result;

[0048] The first part and register and the second part and register in the second type processing unit are used to receive the second convolution calculation result, wherein the second part and register are also used to output the second convolution calculation result to the comparator module.

[0049] In one embodiment, it further includes: a calculation result storage module;

[0050] The calculation result storage module is used to store the convolution calculation results of each processing unit of the systolic array, and use the dot product calculation result of the dot product calculation module to restore the convolution calculation result of the fault processing unit.

[0051] In a second aspect, an embodiment of the present application provides a systolic array fault recovery chip, comprising the systolic array fault recovery circuit described in the first aspect.

[0052] In a third aspect, an embodiment of the present application provides an accelerator, comprising the systolic array fault recovery chip described in the second aspect.

[0053] The systolic array fault recovery circuit, chip and accelerator provided in the present application include a systolic array, a comparator module, a vector re-storage module and a dot product calculation module. The systolic array is an array formed by multiple processing units. The systolic array includes a protection column and its corresponding redundant column, and the protection column is adjacent to the corresponding redundant column. The processing unit is used to perform convolution calculation on the received feature vector and weight vector, and output the convolution calculation result to the comparator module by column. The comparator module is used to compare the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level. If the comparison result is not If they are consistent, the target address information of the fault processing unit to which the inconsistent comparison result belongs is generated, and the target address information is output to the vector recovery module. The vector recovery module is used to obtain the target feature vector and target weight vector used in the current round of convolution calculation of the fault processing unit from the received and stored feature vectors and weight vectors according to the target address information, and output the target feature vector and target weight vector to the dot product calculation module. The dot product calculation module is used to perform dot product calculation on the target feature vector and the target weight vector, so that the dot product calculation result is used to restore the convolution calculation result of the fault processing unit in this round. The present application divides a protection column and a redundant column in a systolic array and receives the same weight vector. Since the feature vector flows horizontally along the systolic array and the weight vector flows vertically along the systolic array, and each processing unit completes a convolution calculation in row and column order in each calculation round, the convolution calculation of the redundant column processing unit and the corresponding protection column processing unit in the same row in each calculation round is based on the same feature vector and the same weight vector. If the processing unit of the systolic array has not failed, the convolution calculation results of the redundant column processing unit and the corresponding protection column processing unit should be consistent. When the convolution calculation results are inconsistent, an address is generated for the faulty processing unit, and the faulty processing unit is located according to the address, and its corresponding target feature vector and target weight vector are obtained, and the convolution calculation is re-performed on the two, and the convolution calculation result of the faulty processing unit is restored using the recalculated convolution calculation result. On the first aspect, the present application divides the protection column and the redundant column on the original structure of the pulsating array without adding any additional processing units, thus reducing the hardware overhead; on the second aspect, the present application sets a corresponding redundant column for any protection column, realizing a one-to-one correspondence between the protection column and the redundant column, and can use the redundant column to cover all faults of the protection column, thereby improving the fault coverage rate; on the third aspect, the present application can use the comparator module to generate the address of the fault processing unit, without the need for other mechanisms to cooperate, that is, it can independently and simultaneously realize the fault location and fault recovery of the processing unit. In summary, the present application can reduce hardware overhead, improve fault coverage rate, and independently and simultaneously realize the fault location and fault recovery of the processing unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the present application or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 is a schematic diagram of a systolic array fault recovery circuit provided in an embodiment of the present application;

[0056] Figure 2 It is a mapping diagram of outputting a fixed data stream in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0057] Figure 3 It is a structural diagram of a dot product calculation module in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0058] Figure 4 It is a schematic diagram of a systolic array protection configuration ratio in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0059] Figure 5 is a schematic diagram of the structure of a comparator module in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0060] Figure 6 It is a transformation logic diagram of a vector transformation unit in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0061] Figure 7 is a schematic diagram of the structure of a first type of processing unit in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0062] Figure 8 is a schematic diagram of the structure of a second type of processing unit in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0063] Fig. 9 It is a schematic diagram of the arrangement of the processing units in the systolic array of the systolic array fault recovery circuit provided in an embodiment of the present application;

[0064] Fig.10 It is one of the timing flow charts of the convolution calculation result loading phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0065] Fig.11 This is the second timing flow chart of the convolution calculation result loading phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0066] Fig.12This is the third timing flow chart of the convolution calculation result loading phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0067] Fig.13 It is one of the timing flow charts of the rewrite phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0068] Fig.14 This is the second timing flow chart of the rewrite phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0069] Fig.15 This is the third timing flow chart of the rewrite phase of the systolic array fault recovery circuit provided in the embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0071] It should be noted that in the description of the embodiments of the present application, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The orientation or position relationship indicated by the terms "upper", "lower" and the like is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. Unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or it can be a connection between the two elements. For those skilled in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0072] The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0073] Figure 1 is a schematic diagram of a systolic array fault recovery circuit provided in an embodiment of the present application;

[0074] Figure 2 It is a mapping diagram of outputting a fixed data stream in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0075] Figure 3 It is a structural diagram of a dot product calculation module in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0076] Figure 4 It is a schematic diagram of a systolic array protection configuration ratio in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0077] Reference Figure 1 , an embodiment of the present application provides a systolic array fault recovery circuit, which may include: a systolic array, a comparator module, a vector re-storage module and a dot product calculation module;

[0078] The systolic array is an array formed by multiple processing units, and the systolic array includes a protection column and its corresponding redundant column, and the protection column is adjacent to the corresponding redundant column;

[0079] A processing unit, configured to perform convolution calculation on the received feature vector and weight vector, and output the convolution calculation results to the comparator module by column; wherein the feature vector flows horizontally along the systolic array, the weight vector flows vertically along the systolic array, and the weight vectors received by the protection column and its corresponding redundant column are the same;

[0080] A comparator module is used to compare the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level. If the comparison results are inconsistent, the target address information of the fault processing unit to which the inconsistent comparison result belongs is generated, and the target address information is output to the vector re-storage module;

[0081] A vector re-storage module is used to obtain the target feature vector and target weight vector used by the fault processing unit in the current round of convolution calculation from the received and stored feature vectors and weight vectors according to the target address information, and output the target feature vector and target weight vector to the dot product calculation module;

[0082] The dot product calculation module is used to perform dot product calculation on the target feature vector and the target weight vector, so that the dot product calculation result is used to restore the convolution calculation result of the fault processing unit in this round.

[0083] Reference Figure 2 , the systolic array includes The output fixed data flow mode is adopted for each processing unit, that is, the feature vector flows horizontally along the pulsation array, and the weight vector flows vertically along the pulsation array. The convolution calculation results of each processing unit are stored in itself. Compared with other data flow modes, the output fixed data flow mode has lower energy consumption because the partial sum is fixed during the calculation process and the bit width of the partial sum is generally larger. In addition, since this mode does not require pre-stored data in each processing unit, the data flow runs faster. During the data flow calculation process, the feature vector is propagated horizontally to realize the reuse of the feature vector, and the weight vector is propagated vertically to realize the reuse of the weight vector. The calculations involving adjacent feature vectors or adjacent weight vectors differ by one cycle in time. Among them, each feature vector corresponds to the data in a convolution window in the convolution operation, and each weight vector corresponds to the data in a complete convolution kernel. For example Figure 2 In the Corresponding to the data in convolution window 1 in the input feature map, the feature vector Corresponding to the data in convolution window 2 in the input feature map, and so on, the feature vector Corresponding to the convolution window in the input feature map The data in , the feature vector Corresponding to the convolution window in the input feature map The data in; weight vector Corresponding to the data in convolution kernel 1, the weight vector Corresponding to the data in convolution kernel 2, and so on, the weight vector Corresponding convolution kernel When all processing units have completed the convolution calculation, the convolution calculation results are passed to the first column of the systolic array by column and output.

[0084] Reference Figure 1The comparator module is connected to the first column of the systolic array to receive the convolution calculation results output by the systolic array in columns after all processing units of the systolic array complete the convolution calculation, and compare the currently received redundant column convolution calculation results with the corresponding protection column convolution calculation results each time a high-level redundant column signal is received, until all redundant column convolution calculation results of this round are compared with the corresponding protection column convolution calculation results; the dot product calculation module is connected to the vector re-storage module to receive the target feature vector and the target weight vector output by the vector re-storage module, and After cycles, the dot product calculation result of the target feature vector and the target weight vector is obtained, where Indicates rounding up.

[0085] Reference Figure 3 ,The point multiplication calculation module can include a multiplier, an addition tree, and a partial sum register, Represents the length of the input vector, that is, the length of the target feature vector or the target weight vector. The length of the target feature vector is equal to the length of the target weight vector. Indicates the number of multipliers, that is, the maximum vector length that can be calculated by the dot product calculation module in a single cycle. The partial sum calculation result of the addition tree will be output to the partial sum register, and then sent back to the addition tree for calculation until the final dot product calculation result is obtained.

[0086] Reference Figure 4 , the protection column is located in the previous column of the corresponding redundant column, and the protection configuration ratio of the systolic array is 0%, 33.33%, 66.66% and 100% respectively, where the protection configuration ratio is the ratio of the sum of the number of protection columns and redundant columns to the total number of columns of the systolic array. The simulation results show that when the protection configuration ratio is 100%, the fault coverage reaches 99.96%, which converges to the ASILD (Automotive Safety Integrity Level D) standard of ISO26262; when the protection configuration ratio is greater than 87.5%, the fault coverage is greater than 97%, which converges to the ASILC (Automotive Safety Integrity Level C) standard of ISO26262; when the protection configuration ratio is greater than 37.5%, the fault coverage is greater than 90%, which converges to the ASILB (Automotive Safety Integrity Level B) standard of ISO26262.

[0087] The systolic array fault recovery circuit provided in the present embodiment includes a systolic array, a comparator module, a vector re-storage module and a dot product calculation module. The systolic array is an array formed by multiple processing units. The systolic array includes a protection column and its corresponding redundant column, and the protection column is adjacent to the corresponding redundant column. The processing unit is used to perform convolution calculation on the received feature vector and weight vector, and output the convolution calculation result to the comparator module by column. The comparator module is used to compare the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level. If the comparison result is inconsistent, the target address information of the fault processing unit to which the inconsistent comparison result belongs is generated, and the target address information is output to the vector re-storage module. The vector re-storage module is used to obtain the target feature vector and target weight vector used for the convolution calculation of the fault processing unit in this round from the received and stored feature vector and weight vector according to the target address information, and output the target feature vector and target weight vector to the dot product calculation module. The dot product calculation module is used to perform dot product calculation on the target feature vector and the target weight vector, so that the dot product calculation result is used to recover the convolution calculation result of the fault processing unit in this round. In this embodiment, a protection column and a redundant column are divided in a systolic array, and the same weight vector is received. Since the feature vector flows horizontally along the systolic array, the weight vector flows vertically along the systolic array, and each processing unit completes a convolution calculation in row and column order in each calculation round, therefore, in each calculation round, the convolution calculation of the redundant column processing unit and the corresponding protection column processing unit located in the same row is performed based on the same feature vector and the same weight vector. If the processing unit of the systolic array does not fail, the convolution calculation results of the redundant column processing unit and the corresponding protection column processing unit should be consistent. When the convolution calculation results are inconsistent, an address is generated for the faulty processing unit, and the faulty processing unit is located according to the address, and its corresponding target feature vector and target weight vector are obtained, and the convolution calculation is re-performed on the two, and the convolution calculation result of the faulty processing unit is restored using the recalculated convolution calculation result. First, this embodiment divides the protection column and the redundant column on the original structure of the systolic array without adding any additional processing units, thus reducing hardware overhead; second, this embodiment sets a corresponding redundant column for any protection column, realizing a one-to-one correspondence between the protection column and the redundant column, and can use the redundant column to cover all faults of the protection column, thereby improving the fault coverage rate; third, this embodiment can use the comparator module to generate the address of the fault processing unit, without the need for other mechanisms to cooperate, that is, it can independently and simultaneously realize the fault location and fault recovery of the processing unit. In summary, this embodiment can reduce hardware overhead, improve fault coverage rate, and independently and simultaneously realize the fault location and fault recovery of the processing unit.

[0088] Reference Figure 1,In one embodiment, the systolic array fault recovery circuit may further include: a vector input module;

[0089] Vector input module for:

[0090] The feature vector and the weight vector are stored, wherein the protected weight vector in the weight vector is redundantly stored according to the redundant protection configuration;

[0091] The feature vector and the weight vector are loaded into the processing unit of the systolic array; wherein the same protected weight vector is loaded into the protection column of the systolic array and its corresponding redundant column respectively.

[0092] The vector input module can be divided into a feature vector storage array and a weight vector storage array, wherein the feature vector storage array is used to store feature vectors, and the weight vector storage array is used to store weight vectors. The feature vector storage array is connected to the first column of the systolic array, and the weight vector storage array is connected to the first row of the systolic array to load feature vectors and weight vectors into the processing units of the systolic array according to the output fixed data stream mode.

[0093] In addition, the feature vector storage array and the weight vector storage array are also respectively connected to the vector re-storage module to synchronously output the feature vectors and weight vectors loaded into the pulsation array to the vector re-storage module for re-storage, so that the vector re-storage module can subsequently acquire the target feature vector and target weight vector.

[0094] When the vector input module stores the weight vector, it will decide whether to redundantly store the weight vector being stored according to the received configuration signal. If so, it means that the weight vector being stored is a protected weight vector, and the protected weight vector is redundantly stored, and the redundantly stored protected weight vector is arranged adjacent to the same protected weight vector stored in the weight vector storage array. When loading the weight vector to the systolic array, the protected weight vector stored in the weight vector storage array is loaded into the protection column of the systolic array, and the redundantly stored same protected weight vector is loaded into the redundant column processing unit corresponding to the protection column.

[0095] In this embodiment, the vector input module stores the feature vector and the weight vector, and loads the feature vector into the first column of the systolic array and the weight vector into the first row of the systolic array, ensuring that the feature vector and the weight vector flow in the systolic array according to the output fixed data flow mode. Furthermore, the vector input module can load the protected weight vector into the protection column and the redundant column of the systolic array respectively after redundantly storing the protected weight vector, so that the systolic array can use the redundant column to achieve comprehensive protection of the protection column.

[0096] Figure 5 is a schematic diagram of the structure of the comparator module in the systolic array fault recovery circuit provided in the embodiment of the present application; Figure 5 ,In one embodiment, the comparator module may include a register unit, a comparator unit and a fault address unit;

[0097] A register unit, used for outputting the convolution calculation result of the protection column to the comparator unit;

[0098] A comparator unit, used for comparing the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level, and if the comparison results are inconsistent, generating target address information of the fault processing unit to which the inconsistent comparison result belongs, and outputting the target address information to the fault address unit;

[0099] The fault address unit is used to store target address information and output the target address information to the vector re-storage module.

[0100] Reference Figure 1 and Figure 5 The register unit is connected to the first column of the systolic array and the comparator unit. The comparator unit is also connected to the first column of the systolic array and the fault address unit. The fault address unit is also connected to the vector re-storage module.

[0101] The register unit receives and stores the convolution calculation results of the protection column output by the systolic array, and outputs the convolution calculation results of the protection column to the comparator unit. The comparator unit receives the convolution calculation results of the redundant column output by the systolic array, and compares the currently received redundant column convolution calculation result with the convolution calculation result of the corresponding protection column each time a high-level redundant column signal is received, until all redundant column convolution calculation results of this round are compared with the corresponding protection column convolution calculation results. During the comparison process, if the comparison results are inconsistent, the target address information of the fault processing unit to which the inconsistent comparison results belong is generated, and the target address information and the error signal are output to the fault address unit, the error signal enables the fault address unit to store the target address information, and subsequently outputs the target address information to the vector re-storage module.

[0102] Among them, the redundant column signal is stored separately outside the systolic array. It is automatically generated when the vector input module receives the vector according to the configuration and redundantly stores the protected weight vector. It is used to characterize whether the current output of the systolic array is the convolution calculation result of the redundant column. If it is a high level, it means that the current output is the convolution calculation result of the redundant column. When it is a low level, it means that the current output is not the convolution calculation result of the redundant column. The target address information includes the feature vector address and weight vector address used by the fault processing unit in this round of convolution calculation. Based on this, the fault processing unit can be located. At the same time, it provides a basis for the vector re-storage module to obtain the target feature vector and target weight vector.

[0103] Furthermore, the comparator unit can adjust the tolerance according to the comparison threshold signal, that is, the difference between the redundant column convolution calculation result of this round and the corresponding protection column convolution calculation result is considered to be consistent within a certain range, so as to adapt to the high robustness of the convolutional neural network.

[0104] In this embodiment, the comparator module receives the convolution calculation results of the protection column and the redundant column of the systolic array respectively through the register unit and the comparator unit, and compares the convolution calculation results of the redundant column with the convolution calculation results of the corresponding protection column after receiving the convolution calculation results of the redundant column through the comparator unit each time, so as to timely find inconsistent comparison results and improve the fault detection efficiency. At the same time, the comparator unit generates the corresponding feature vector address and weight vector address and outputs them to the fault address unit, and then outputs the feature vector address and weight vector address to the vector re-storage module through the fault address unit, so that the vector re-storage module can quickly locate the fault processing unit and quickly obtain the target feature vector and target weight vector.

[0105] Figure 6 is a transformation logic diagram of a vector transformation unit in a systolic array fault recovery circuit provided in an embodiment of the present application; Figure 6 ,In one embodiment, the vector re-storage module includes a vector transformation unit, a feature vector storage unit, and a weight vector storage unit;

[0106] A vector transformation unit, used for performing position transformation on the received feature vector and weight vector so that the feature vector and weight vector after position transformation match the calculation logic of the dot product calculation module;

[0107] A feature vector storage unit, used for storing and outputting the feature vector after position transformation;

[0108] The weight vector storage unit is used to store and output the weight vector after position transformation.

[0109] Reference Figure 1 and Figure 6The vector transformation unit is connected to the vector input module and the storage unit, and the storage unit is connected to the comparator module and the dot product calculation module, wherein the storage unit includes a feature vector storage unit and a weight vector storage unit.

[0110] Reference Figure 6 For feature vectors, each row corresponds to a convolution window of the input feature map before the transformation, and each row corresponds to the pixel bit of each convolution window after the transformation; for weight vectors, each row corresponds to a convolution kernel before the transformation, and each row corresponds to a data bit of each convolution kernel after the transformation. After the transformation, no matter any feature vector or any weight vector, only one read of the storage unit is needed to obtain the complete vector, which improves the storage efficiency of the vector storage module. Figure 6 Included vector, vector, vector, vector, Vector Sum vector, where:

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] by For example, before the transformation, the storage unit needs to be read 6 times to get the complete vector, and after the transformation, due to The order of the elements of the vector in each row is consistent, and the storage unit only needs to be read once to get the complete The same is true for other vectors, which will not be described here.

[0118] This embodiment uses a vector transformation unit to transform the positions of received feature vectors and weight vectors, so that each feature vector and each weight vector can be read once by the storage unit to obtain a complete vector, thereby improving the storage efficiency of the vector storage module. At the same time, by separately storing and outputting the feature vectors and weight vectors after the position transformation, mutual interference between the two types of vectors can be avoided, thereby improving the data accuracy during the vector circulation process.

[0119] Figure 7 is a schematic diagram of the structure of a first type of processing unit in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0120] Figure 8 is a schematic diagram of the structure of a second type of processing unit in a systolic array fault recovery circuit provided in an embodiment of the present application;

[0121] Fig. 9 It is a schematic diagram of the arrangement of the processing units in the systolic array of the systolic array fault recovery circuit provided in an embodiment of the present application.

[0122] Reference Figures 7 to 9 , in one embodiment, the processing unit includes a first portion and a register and a second portion and a register;

[0123] The plurality of processing units of the systolic array are arranged alternately in columns as first-type processing units and second-type processing units;

[0124] Reference Figure 7 , the first type of processing unit , used for:

[0125] Receive the first feature vector to be processed output by the previous processing unit in the same row and the weight vector to be processed output by the previous processing unit in the same column ;

[0126] For the first feature vector to be processed and the weight vector to be processed Perform convolution calculation to get the first convolution calculation result ;

[0127] The first convolution calculation result Save to the first part of itself and register and the second part of itself and the register ;

[0128] The first part of itself and the register The first convolution calculation result in Output to the comparator module, and at the same time, the second part of itself and the register The first convolution calculation result in Output, when there is no fault in the part and register, and equal.

[0129] Reference Figure 8 , the second type of processing unit , used for:

[0130] Receive the second feature vector to be processed output by the previous processing unit in the same row and the weight vector to be processed output by the previous processing unit in the same column ;

[0131] For the second feature vector to be processed and the weight vector to be processed Perform convolution calculation to obtain the second convolution calculation result ;

[0132] The second convolution calculation result Save to the first part of itself and register and the second part of itself and the register ;

[0133] The second part of itself and the register The second convolution calculation result in Output to the comparator module, and at the same time, the first part of itself and the register The second convolution calculation result in Output, when there is no fault in the part and register, and equal.

[0134] The first feature vector to be processed and the second feature vector to be processed At the initial moment it is the same eigenvector.

[0135] Furthermore, the processing unit also includes a first feature vector register, a second feature vector register, a weight vector register and a convolution calculator.

[0136] Reference Figure 7 and Fig. 9 , the first type of processing unit The first eigenvector register in , used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row , and the first feature vector to be processed Output to the next processing unit in the same row, that is, the second type of processing unit The first eigenvector register in ;when In When there is no fault, of and equal;

[0137] First type processing unit The second eigenvector register in , used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row , and the second feature vector to be processed Output to the next processing unit in the same row, that is, the second type of processing unit The convolution calculator and the second eigenvector register in ;when In When there is no fault, of and equal;

[0138] First type processing unit The weight vector registers in , used to receive the weight vector to be processed output by the weight vector register in the previous processing unit in the same column , and the weight vector to be processed Output to the convolution calculator and weight vector register in the next processing unit in the same column ;when In When there is no fault, of and equal;

[0139] First type processing unit The convolution calculator in the embodiment includes a multiplier and an adder for receiving the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row. , and the weight vector to be processed output by the weight vector register of the previous processing unit in the same column , and the first feature vector to be processed and the weight vector to be processed Perform convolution calculation to get the first convolution calculation result ;

[0140] First type processing unit The first part and register and the second part and register , used to receive the first convolution calculation result and , where the first part and register It is also used to convert the first convolution calculation result Output to comparator module, second section and register It is also used to convert the first convolution calculation result Output, when there is no fault in the part and register, and equal.

[0141] Reference Figure 8 and Fig. 9 , the second type of processing unit The first eigenvector register in , used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row , and the first feature vector to be processed Output to the next processing unit in the same row, i.e. the first type of processing unit The convolution calculator and first eigenvector register in ;when In When there is no fault, of and equal;

[0142] Second type processing unit The second eigenvector register in , used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row , and the second feature vector to be processed Output to the next processing unit in the same row, i.e. the first type of processing unit The second eigenvector register in ;when In When there is no fault, of and equal;

[0143] Second type processing unit The weight vector registers in , used to receive the weight vector to be processed output by the weight vector register in the previous processing unit in the same column , and the weight vector to be processed Output to the convolution calculator and weight vector register in the next processing unit in the same column ;when In When there is no fault, of and equal;

[0144] Second type processing unit The convolution calculator in the embodiment includes a multiplier and an adder for receiving the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row. , and the weight vector to be processed output by the weight vector register of the previous processing unit in the same column , and for the second feature vector to be processed and the weight vector to be processed Perform convolution calculation to obtain the second convolution calculation result ;

[0145] Second type processing unit The first part and register and the second part and register , used to receive the second convolution calculation result and , where the second part and register It is also used to convert the second convolution calculation result Output to comparator module, first section and register It is also used to convert the second convolution calculation result Output, when there is no fault in the part and register, and equal.

[0146] According to the above, when the systolic array performs convolution calculations, since the weight vectors of the protection column and the corresponding redundant column in the same round are the same, when the processing unit is not faulty, the convolution calculation results of the processing units in the same row of the protection column and the corresponding redundant column in the same round are the same, otherwise it means that one of them is faulty. Since each processing unit of the systolic array not only needs to perform its own convolution calculations, but also needs to transfer feature vectors by column and weight vectors by row, if the processing unit through which the vector flows fails, it may pollute the vector, resulting in the subsequent redundant columns in the same row being unable to form redundant protection for the protection column. Specifically, it can be divided into the following three situations:

[0147] 1. If a register storing a feature vector in a processing unit fails, the polluted feature vector output by the register is the feature vector input of the next processing unit in the same row as the processing unit. By analogy, the feature vectors of the subsequent processing units in the same row as the processing unit will all be wrong, causing the redundant column processing units and the protection column processing units in the subsequent processing units in the same row to produce the same wrong results, and the difference in convolution calculation between the redundant column processing units and the protection column processing units cannot be reflected, so the faulty processing unit cannot be identified, resulting in the failure of redundant protection;

[0148] 2. If a register storing a weight vector in a processing unit fails, since the weight vector is transmitted row by row, the polluted weight vector output by the register will not affect the subsequent processing units in the same row. The subsequent redundant column processing units and protection column processing units can still reflect the convolution calculation difference when a faulty processing unit occurs, and the redundant protection is effective.

[0149] 3. If a register storing the convolution calculation result in a processing unit fails, although the convolution calculation result is fixed in each processing unit, the loading process after the calculation is completed is transferred from the last column of the systolic array to the first column. Therefore, the redundant column processing unit and the protection column processing unit in the same forward processing unit will produce the same erroneous result, which cannot reflect the difference in convolution calculation between the redundant column processing unit and the protection column processing unit, and thus the faulty processing unit cannot be identified, resulting in the failure of redundant protection.

[0150] In this embodiment, the feature vectors sequentially input into the processing units in the same row are the same feature vectors at the initial moment. The feature vector has two paths in the horizontal transmission process. The two paths correspond to a feature vector register in each processing unit respectively. If any feature vector register fails and causes the feature vector of the corresponding path to be contaminated during the transmission process, the contaminated feature vector will be different from the feature vector of the other path. Since the two types of processing units are alternately arranged, the feature vectors input into the subsequent protection column processing unit and the corresponding redundant column processing unit always come from different paths (see Fig. 9 The two different dotted line paths in the figure make the subsequent protection column processing unit and the corresponding redundant column processing unit always perform convolution calculation based on different feature vectors, thereby generating a convolution calculation difference, and the difference is maintained during the feature vector transmission process, that is, the subsequent redundant column processing unit and the protection processing unit will not produce the same erroneous result, which can reflect the convolution calculation difference between the redundant column processing unit and the protection column processing unit, thereby identifying the faulty processing unit and making the redundant protection effective.

[0151] Furthermore, since the transmission path of the weight vector does not change, the conclusion that the redundant protection is effective in the above point 2 will not change.

[0152] Furthermore, since the convolution calculation results in different parts and registers of two adjacent processing units in the same row are forwarded, if any convolution calculation result is contaminated, it will cause a difference between it and the other convolution calculation result, and since the two types of processing units are alternately arranged, the difference is maintained during the forward transmission process, so that the forward redundant column processing unit and the protection processing unit will not produce the same erroneous results, and the convolution calculation difference between the redundant column processing unit and the protection column processing unit can be reflected, thereby identifying the faulty processing unit and the redundant protection takes effect.

[0153] Reference Figure 1 ,In one embodiment, the systolic array fault recovery circuit may further include: a calculation result storage module;

[0154] The calculation result storage module is used to store the convolution calculation results of each processing unit of the systolic array, and use the point product calculation results of the point product calculation module to restore the convolution calculation results of the fault processing unit.

[0155] The calculation result storage module is connected to the first column of the systolic array and the dot product calculation module to receive and store the convolution calculation results output by each processing unit of the systolic array and the dot product calculation result output by the dot product calculation module, so as to use the dot product calculation result of the dot product calculation module to restore the convolution calculation result of the fault processing unit.

[0156] In this embodiment, the calculation result storage module can store the convolution calculation results of each processing unit of the systolic array and the dot product calculation results of the dot product calculation module respectively, so that the dot product calculation results can be used to cover the erroneous convolution calculation results to restore them.

[0157] Fig.10 It is one of the timing flow charts of the convolution calculation result loading phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0158] Fig.11 This is the second timing flow chart of the convolution calculation result loading phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0159] Fig.12 This is the third timing flow chart of the convolution calculation result loading phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0160] Fig.13 It is one of the timing flow charts of the rewrite phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0161] Fig.14 This is the second timing flow chart of the rewrite phase of the systolic array fault recovery circuit provided in the embodiment of the present application;

[0162] Fig.15 This is the third timing flow chart of the rewrite phase of the systolic array fault recovery circuit provided in the embodiment of the present application.

[0163] To include Taking a systolic array with 3 processing units and 3 faulty processing units as an example, the timing process of a complete convolution calculation is introduced. Assuming the convolution kernel size is , the number of channels is , then the convolution length for , and each convolution kernel is weighted, so the systolic array takes time to complete a convolution calculation cycle.

[0164] Reference Figures 10 to 12 , it takes 32 cycles from the completion of a convolution calculation of the systolic array to the stage of loading the calculation results, that is, to In the cycle, the completed convolution calculation results are loaded into the calculation result storage module. Due to the existence of the fault processing unit, there are 3 erroneous results in the calculation result storage module. At the same time, the comparator module compares the convolution calculation results of the protection column and the redundant column of the systolic array, generates the target address information of the fault processing unit to which the 3 erroneous results belong, realizes fault location, and prepares for the dot product calculation of the dot product calculation module.

[0165] Reference Figures 13 to 15 , after loading is completed, the rewriting phase begins, which is the phase of overwriting and restoring the erroneous convolution calculation results. to cycle, the vector re-storage module receives the target address information of the comparator module, thereby obtaining the target feature vector and target weight vector corresponding to the fault processing unit, and outputs the target feature vector and target weight vector to the dot product calculation module through the feature vector storage unit and the weight vector storage unit in the vector re-storage module, and performs dot product calculation through the dot product calculation module to cover the error result and complete fault recovery.

[0166] An embodiment of the present application provides a systolic array fault recovery circuit chip, including the above-mentioned systolic array fault recovery circuit.

[0167] Therefore, the systolic array fault recovery circuit chip of this embodiment has the same technical advantages as the aforementioned systolic array fault recovery circuit, which will not be described in detail here.

[0168] An embodiment of the present application provides an accelerator, comprising the aforementioned systolic array fault recovery chip.

[0169] Therefore, the accelerator of this embodiment has the same technical advantages as the aforementioned systolic array fault recovery chip, which will not be described in detail here.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A systolic array fault recovery circuit, characterized in that: include: Systolic array, comparator module, vector re-storage module and dot product calculation module; The systolic array is an array formed by a plurality of processing units, the systolic array includes a protection column and a corresponding redundant column, and the protection column is adjacent to the corresponding redundant column; The processing unit is used to perform convolution calculation on the received feature vector and weight vector, and output the convolution calculation results to the comparator module by column; wherein the feature vector flows horizontally along the systolic array, the weight vector flows vertically along the systolic array, and the weight vector received by the protection column and its corresponding redundant column is the same; The comparator module is used to compare the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level, and if the comparison results are inconsistent, generate the target address information of the fault processing unit to which the inconsistent comparison result belongs, and output the target address information to the vector re-storage module; The vector re-storage module is used to obtain the target feature vector and target weight vector used by the fault processing unit in the current round of convolution calculation from the received and stored feature vectors and weight vectors according to the target address information, and output the target feature vector and the target weight vector to the dot product calculation module; The dot product calculation module is used to perform dot product calculation on the target feature vector and the target weight vector, so that the dot product calculation result is used to restore the convolution calculation result of the fault processing unit in this round.

2. The systolic array fault recovery circuit according to claim 1, characterized in that: Also includes: Vector input module; The vector input module is used to: The feature vector and the weight vector are stored, wherein the protected weight vector in the weight vector is redundantly stored according to a redundant protection configuration; The feature vector and the weight vector are loaded into the systolic array; wherein the same protected weight vector is loaded into a protection column of the systolic array and a corresponding redundant column thereof, respectively.

3. The systolic array fault recovery circuit according to claim 1, characterized in that: The comparator module includes a register unit, a comparator unit and a fault address unit; The register unit is used to output the convolution calculation result of the protection column to the comparator unit; The comparator unit is used to compare the redundant column convolution calculation result of this round with the corresponding protection column convolution calculation result when the redundant column signal is at a high level, and if the comparison results are inconsistent, generate the target address information of the fault processing unit to which the inconsistent comparison result belongs, and output the target address information to the fault address unit; The fault address unit is used to store the target address information and output the target address information to the vector re-storage module.

4. The systolic array fault recovery circuit according to claim 1, characterized in that: The vector re-storage module includes a vector conversion unit, a feature vector storage unit and a weight vector storage unit; The vector transformation unit is used to perform position transformation on the received feature vector and weight vector so that the feature vector and weight vector after position transformation match the calculation logic of the dot product calculation module; The feature vector storage unit is used to store and output the feature vector after position transformation; The weight vector storage unit is used to store and output the weight vector after position transformation.

5. The systolic array fault recovery circuit according to claim 1, characterized in that: The processing unit includes a first part and a register and a second part and a register; The plurality of processing units of the systolic array are arranged alternately in columns as first-type processing units and second-type processing units; The first type processing unit is used to: Receive a first feature vector to be processed output by a previous processing unit in the same row and a weight vector to be processed output by a previous processing unit in the same column; Performing a convolution calculation on the first feature vector to be processed and the weight vector to be processed to obtain a first convolution calculation result; The first convolution calculation result is saved to the first part and register of the first part and the second part and register of the second part; Outputting the first part of itself and the first convolution calculation result in the register to the comparator module; The second type processing unit is used for: Receive a second feature vector to be processed output by a previous processing unit in the same row and a weight vector to be processed output by a previous processing unit in the same column; Performing a convolution calculation on the second feature vector to be processed and the weight vector to be processed to obtain a second convolution calculation result; The second convolution calculation result is saved to the first part and register of the second convolution operation and the second part and register of the second convolution operation; Outputting the second part of itself and the second convolution calculation result in the register to the comparator module; The first feature vector to be processed and the second feature vector to be processed are the same feature vector at an initial moment.

6. The systolic array fault recovery circuit according to claim 5, characterized in that: The processing unit also includes a first feature vector register, a second feature vector register, a weight vector register and a convolution calculator; The first feature vector register in the first type processing unit is used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row, and output the first feature vector to be processed to the first feature vector register in the next processing unit in the same row; The second feature vector register in the first type processing unit is used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row, and output the second feature vector to be processed to the convolution calculator and the second feature vector register in the next processing unit in the same row; The weight vector register in the first type of processing unit is used to receive the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and output the weight vector to be processed to the convolution calculator and weight vector register in the next processing unit in the same column; The convolution calculator in the first type of processing unit is used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row, and the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and perform convolution calculation on the first feature vector to be processed and the weight vector to be processed to obtain a first convolution calculation result; The first part and register and the second part and register in the first type processing unit are used to receive the first convolution calculation result, wherein the first part and register are also used to output the first convolution calculation result to the comparator module.

7. The systolic array fault recovery circuit according to claim 5, characterized in that: The processing unit also includes a first feature vector register, a second feature vector register, a weight vector register and a convolution calculator; The first feature vector register in the second type processing unit is used to receive the first feature vector to be processed output by the first feature vector register in the previous processing unit in the same row, and output the first feature vector to be processed to the convolution calculator and the first feature vector register in the next processing unit in the same row; The second feature vector register in the second type processing unit is used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row, and output the second feature vector to be processed to the second feature vector register in the next processing unit in the same row; The weight vector register in the second type processing unit is used to receive the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and output the weight vector to be processed to the convolution calculator and weight vector register in the next processing unit in the same column; The convolution calculator in the second type processing unit is used to receive the second feature vector to be processed output by the second feature vector register in the previous processing unit in the same row, and the weight vector to be processed output by the weight vector register in the previous processing unit in the same column, and perform convolution calculation on the second feature vector to be processed and the weight vector to be processed to obtain a second convolution calculation result; The first part and register and the second part and register in the second type processing unit are used to receive the second convolution calculation result, wherein the second part and register are also used to output the second convolution calculation result to the comparator module.

8. The systolic array fault recovery circuit according to claim 1, characterized in that: Also includes: Calculation result storage module; The calculation result storage module is used to store the convolution calculation results of each processing unit of the systolic array, and use the dot product calculation result of the dot product calculation module to restore the convolution calculation result of the fault processing unit.

9. A systolic array fault recovery chip, characterized in that: A systolic array fault recovery circuit comprising any one of claims 1 to 8.

10. An accelerator, characterized in that: Including the systolic array fault recovery chip as described in claim 9.