Granted Patent B2

US 12,488,220 · App. 17/286,545 · Granted Dec 2, 2025

Inventors: Jinqing Yu (Shenzhen, CN); Minchao Liang (Shenzhen, CN); Shengnan Yan (Shenzhen, CN); Jie Liu (Shenzhen, CN)

Assignee: ZTE CORPORATION

G06N3/045G06F18/2137G06F18/217G06F18/24765G06N3/04G06N3/063G06N3/08

Abstract

Claims (89)

1 . A data processing method applied in a convolutional neural network, comprising performing convolution calculation on feature data and weights corresponding to respective feature data by the following operations performed by a data processing device comprising a calculation unit that comprises a plurality of classification processor (CP) calculation units and a multiply-add unit (MAU):

classifying, by the plurality of CP calculation units, the weights into at least one type of weights and acquiring, by the plurality of CP calculation units, the feature data corresponding to each type of the weights, wherein the weights are classified according to a set classification rule comprising one of:

weights, of which values are positive numbers and negative numbers having a same absolute value, are classified into one type, and then the weights of various types are further classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type;

weights of which values are positive numbers are classified into one type and weights of which values are negative numbers are classified into another type, and then the weights of various types are further classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type; or,

the weights are directly classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type;

acquiring, by the plurality of CP calculation units, the common factor corresponding to each type of the weights;

acquiring, by the plurality of CP calculation units according to the at least one type of weights and the common factor corresponding to each type of the weights, classification data corresponding to the feature data; and performing, by the plurality of CP calculation units, a left shift calculation on the respective feature data corresponding to each type of the weights according to the classification data corresponding to the respective feature data, and summing, by the plurality of CP calculation units, results of the left shift calculation to obtain a first processing result corresponding to each type of the weights, wherein the classification data is exponent in the power of 2; and performing, by the MAU comprising one multiplier and an accumulator,

multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain a second processing result,

wherein before performing, by the MAU comprising one multiplier and the accumulator, the multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain the second processing result, the method further comprises:

implementing a routing function from the plurality of CP calculation units to the multiplier through a multiplier queue.

2 . The data processing method according to claim 1 , wherein before classifying, by the plurality of CP calculation units, the weights into the at least one type of weights, the data processing method further comprises the following operations performed by the data processing device:

acquiring historical weights;

screening the historical weights according to a set screening condition to obtain screened weights; and

classifying the screened weights according to a set classification rule to obtain the at least one type of weights.

3 . The data processing method according to claim 2 , wherein performing, by the MAU comprising one multiplier and the accumulator, the multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain the second processing result comprises the following operations performed by the MAU:

sorting first processing results corresponding to respective types of weights according to a set sorting rule to obtain sorted first processing results;

according to a sorting order, sequentially selecting from the sorted first processing results a first processing result to be processed;

performing calculation according to the common factor and the first processing result to be processed to obtain a first calculation result;

performing calculation on the currently obtained first calculation result and a previously obtained first calculation result to obtain a second calculation result; and

determining, in a case where it is determined that the second calculation result is a negative number, the second calculation result as the second processing result.

4 . The data processing method according to claim 1 , wherein after performing, by the MAU comprising one multiplier and the accumulator, the multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain the second processing result, the data processing method further comprises the following operations performed by the calculation unit:

acquiring a first dimension of the feature data and a corresponding network dimension;

acquiring a second dimension of the second processing result;

adjusting, in a case where it is determined that the second dimension is lower than the first dimension, a current network dimension corresponding to the second processing result according to the acquired network dimension and a predetermined threshold range;

acquiring a weight corresponding to the second processing result; and

performing calculation on the second processing result based on the current network dimension and the weight corresponding to the second processing result to obtain a third processing result.

5 . The data processing method according to claim 4 , wherein adjusting, in a case where it is determined that the second dimension is lower than the first dimension, the current network dimension corresponding to the second processing result according to the acquired network dimension and the predetermined threshold range comprises the following operations:

performing, in a case where it is determined that the second dimension is lower than the first dimension, a modulus operation on the network dimension and a target network dimension for adjustment to obtain a modulus result;

calculating a ratio according to the modulus result and the target network dimension for adjustment; and

determining, in a case where the ratio is within a predetermined threshold range, the target network dimension for adjustment as the current network dimension corresponding to the second processing result.

6 . A data processing device applied in a convolutional neural network, comprising a calculation unit that comprises a plurality of classification processor (CP) calculation units and a multiply-add unit (MAU) and is configured to perform convolution calculation on feature data and weights corresponding to respective feature data, wherein:

the plurality of CP calculation units are configured to classify the weights into at least one type of weights and acquire the feature data corresponding to each type of the weights, wherein the weights are classified according to a set classification rule comprising one of:

weights, of which values are positive numbers and negative numbers having a same absolute value, are classified into one type, and then the weights of various types are further classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type;

weights of which values are positive numbers are classified into one type and weights of which values are negative numbers are classified into another type, and then the weights of various types are further classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type; or,

the weights are directly classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type;

the plurality of CP calculation units are configured to acquire the common factor corresponding to each type of the weights;

the plurality of CP calculation units are configured to acquire, according to the at least one type of weights and the common factor corresponding to each type of the weights, classification data corresponding to the feature data; and perform a left shift calculation on the respective feature data corresponding to each type of the weights according to the classification data corresponding to the respective feature data, and sum results of the left shift calculation to obtain a first processing result corresponding to each type of the weights, wherein the classification data is exponent in the power of 2;

a multiplier queue is provided to implement a routing function from the plurality of CP calculation units to the multiplier; and

the MAU comprising one multiplier and an accumulator is configured to perform multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain a second processing result.

7 . A non-transitory computer-readable storage medium applied in a convolutional neural network, on which a computer program is stored, wherein the computer program, when being executed by a data processing device comprising a calculation unit that comprises a plurality of classification processor (CP) calculation units and a multiply-add unit (MAU), causes the calculation unit to perform convolution calculation on feature data and weights corresponding to respective feature data by the following operations:

classifying, by the plurality of CP calculation units, the weights into at least one type of weights and acquiring, by the plurality of CP calculation units, the feature data corresponding to each type of the weights, wherein the weights are classified according to a set classification rule comprising one of:

weights, of which values are positive numbers and negative numbers having a same absolute value, are classified into one type, and then the weights of various types are further classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type;

weights of which values are positive numbers are classified into one type and weights of which values are negative numbers are classified into another type, and then the weights of various types are further classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type; or,

the weights are directly classified by converting the weights in a manner of multiplying a common factor by a power of 2 and classifying weights having the same common factor into one type;

acquiring, by the plurality of CP calculation units, the common factor corresponding to each type of the weights;

acquiring, by the plurality of CP calculation units according to the at least one type of weights and the common factor corresponding to each type of the weights, classification data corresponding to the feature data; and performing, by the plurality of CP calculation units, a left shift calculation on the respective feature data corresponding to each type of the weights according to the classification data corresponding to the respective feature data, and summing, by the plurality of CP calculation units, results of the left shift calculation to obtain a first processing result corresponding to each type of the weights, wherein the classification data is exponent in the power of 2; and

performing, by the MAU comprising one multiplier and an accumulator, multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain a second processing result,

wherein before performing, by the MAU comprising one multiplier and the accumulator, the multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain the second processing result, the data processing device is further caused to perform the following operations:

implementing a routing function from the plurality of CP calculation units to the multiplier through a multiplier queue.

8 . The data processing device according to claim 6 , wherein the data processing device is further configured to perform the following operations before classifying, by the plurality of CP calculation units, the weights into the at least one type of weights:

acquiring historical weights;

screening the historical weights according to a set screening condition to obtain screened weights; and

classifying the screened weights according to a set classification rule to obtain the at least one type of weights.

9 . The data processing device according to claim 8 , wherein the MAU, when performing the multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain the second processing result, is configured to perform the following operations:

sorting first processing results corresponding to respective types of weights according to a set sorting rule to obtain sorted first processing results;

according to a sorting order, sequentially selecting from the sorted first processing results a first processing result to be processed;

performing calculation according to the common factor and the first processing result to be processed to obtain a first calculation result;

performing calculation on the currently obtained first calculation result and a previously obtained first calculation result to obtain a second calculation result; and

determining, in a case where it is determined that the second calculation result is a negative number, the second calculation result as the second processing result.

10 . The data processing device according to claim 6 , wherein the calculation unit is further configured to, after performing, by the MAU comprising one multiplier and the accumulator, the multiply-add calculation according to the common factor and the first processing result corresponding to each type of the weights to obtain the second processing result, execute the following operations:

acquiring a first dimension of the feature data and a corresponding network dimension;

acquiring a second dimension of the second processing result;

adjusting, in a case where it is determined that the second dimension is lower than the first dimension, a current network dimension corresponding to the second processing result according to the acquired network dimension and a predetermined threshold range;

acquiring a weight corresponding to the second processing result; and

performing calculation on the second processing result based on the current network dimension and the weight corresponding to the second processing result to obtain a third processing result.

11 . The data processing device according to claim 10 , wherein adjusting, in a case where it is determined that the second dimension is lower than the first dimension, the current network dimension corresponding to the second processing result according to the acquired network dimension and the predetermined threshold range comprises the following operations:

performing, in a case where it is determined that the second dimension is lower than the first dimension, a modulus operation on the network dimension and a target network dimension for adjustment to obtain a modulus result;

calculating a ratio according to the modulus result and the target network dimension for adjustment; and

determining, in a case where the ratio is within a predetermined threshold range, the target network dimension for adjustment as the current network dimension corresponding to the second processing result.

12 . The non-transitory computer-readable storage medium according to claim 7 , wherein the computer program, when being executed by the data processing device, further causes the data processing device to:

acquire historical weights;

screen the historical weights according to a set screening condition to obtain screened weights; and

classify the screened weights according to a set classification rule to obtain the at least one type of weights.

13 . The non-transitory computer-readable storage medium according to claim 12 , wherein the computer program, when being executed by the MAU, causes the MAU to:

sort first processing results corresponding to respective types of weights according to a set sorting rule to obtain sorted first processing results;

according to a sorting order, sequentially select from the sorted first processing results a first processing result to be processed;

perform calculation according to the common factor and the first processing result to be processed to obtain a first calculation result;

perform calculation on the currently obtained first calculation result and a previously obtained first calculation result to obtain a second calculation result; and

determine, in a case where it is determined that the second calculation result is a negative number, the second calculation result as the second processing result.

14 . The non-transitory computer-readable storage medium according to claim 7 , wherein the computer program, when being executed by the calculation unit, further causes the calculation unit to:

acquire a first dimension of the feature data and a corresponding network dimension;

acquire a second dimension of the second processing result;

adjust, in a case where it is determined that the second dimension is lower than the first dimension, a current network dimension corresponding to the second processing result according to the acquired network dimension and a predetermined threshold range;

acquire a weight corresponding to the second processing result; and

perform calculation on the second processing result based on the current network dimension and the weight corresponding to the second processing result to obtain a third processing result.

15 . The non-transitory computer-readable storage medium according to claim 14 , wherein the computer program, when being executed by the calculation unit, causes the calculation unit to:

perform, in a case where it is determined that the second dimension is lower than the first dimension, a modulus operation on the network dimension and a target network dimension for adjustment to obtain a modulus result;

calculate a ratio according to the modulus result and the target network dimension for adjustment; and

determine, in a case where the ratio is within a predetermined threshold range, the target network dimension for adjustment as the current network dimension corresponding to the second processing result.

Assignments (2)

Priority Claims (1)

Continuity (1)

References Cited (21)