NEURAL PROCESSOR WITH ACTIVATION COMPRESSION
A neural processing device is provided. The neural processing device comprises: an activation buffer in which first and second input activations are stored, an activation compressor configured to generate a first compressed input activation by using the first and second input activations, and a tensor unit configured to perform two-dimensional calculations using the first compressed input activation, wherein the first compressed input activation comprises first input row data comprising at least a portion of the first input activation and at least a portion of the second input activation, and first metadata corresponding to the first input row data.
1 . A data processing method, performed by a neural core included in a neural processor, the method comprising:
storing first input activation and second input activation;
generating a first compressed input activation by using the first input activation and the second input activation; and
performing data calculation using the first compressed input activation,
wherein the first compressed input activation comprises:
a first input row data comprising first data element and second data element,
a first metadata comprising first source element and second source element,
the first source element comprises information indicating that the first data element has originated from the first input activation, and
the second source element comprises information indicating that the second data element has originated from the second input activation.
2 . The data processing method of claim 1 , wherein
the first input activation comprises a first input element that is an effective element and a second input element that is an ineffective element, and
the second input activation comprises a third input element that is an effective element, the method further comprising generating the first input row data by pushing the third input element to the second input element.
3 . The data processing method of claim 2 , wherein
the first data element corresponds to the first input element, and
the second input element corresponds to the third input element.
4 . The data processing method of claim 2 , wherein
a position of the second input element and a position of the third input element correspond to each other.
5 . The data processing method of claim 2 , wherein
a number of ineffective elements of the first input activation is greater than or equal to a number of effective elements of the second input activation.
6 . The data processing method of claim 2 , wherein
a position of the second input element and a position of the third input element do not correspond to each other.
7 . The data processing method of claim 2 , wherein
the pushing the third input element to the second input element further comprises sequentially pushing the third input element to the second input element.
8 . The data processing method of claim 1 , wherein
the first source element and the second source element have different values each other.
9 . The data processing method of claim 1 , wherein
a position of the first source element corresponds to a position of the first data element, and a position of the second source element corresponds to a position of the second data element.
10 . The data processing method of claim 1 , wherein
the first metadata further comprises a first operation index comprising information for a weight that is calculated with the first input row data.
11 . The data processing method of claim 10 , wherein the first operation index comprises first and second operation elements, and
the first operation element is associated with the first data element, and the second operation element is associated with the second data element.
12 . The data processing method of claim 10 , wherein the first operation element comprises information for a first weight element that is calculated with the first data element, and the second operation element comprises information for a second weight element that is calculated with the second data element.
13 . The data processing method of claim 1 , wherein the method further comprises:
generating a first preliminary input row data by using the first and second input activations,
generating a second input row data and a second preliminary input row data by using third and fourth input activations, and
generating a third input row data by using the first and second preliminary input row data.
14 . The data processing method of claim 13 , wherein the first and second preliminary input row data is temporarily stored until the third input row data is generated.
15 . The data processing method of claim 1 , wherein the method further comprises:
generating a first compressed weight matrix by using first and second weight matrices,
wherein the first compressed weight matrix comprises at least a portion of the first weight matrix and at least a portion of the second weight matrix.
16 . A data processing method performed by a neural core included in a neural processor, the method comprising:
storing first input activation and second input activation;
generating a first compressed input activation by using the first input activation and the second input activation; and
performing 2-dimensional calculation using the first compressed input activation,
wherein the first compressed input activation comprises a first input row data and a first metadata,
the first input row data comprises at least a portion of the first input activation and at least a portion of the second input activation,
the first metadata comprises a first source index, and
the first source index comprises a first index indicating that a first data element has originated from the first input activation and a second index indicating that a second data element has originated from the second input activation.
17 . The data processing method of claim 16 , wherein the first source index comprises information that elements included in the first input row data have originated from one of the first and second input activations.
18 . The data processing method of claim 17 , wherein the first input row data comprises first and second data elements, and the first source index comprises first and second source elements,
the first data element corresponds to the first source element and the second data element corresponds to the second source element, and
the first source element has a value of the first index and the second source element has a value of the second index.
19 . The data processing method of claim 16 , wherein the method further comprises:
generating a first preliminary input activation by using remaining data after generating the first compressed input activation,
storing third and fourth input activation,
generating a second compressed input activation and a second preliminary input activation by using the third and fourth input activation, and
generating a third compressed input activation by using the first and second preliminary input activations.
20 . The data processing method of claim 16 , wherein the first input activation comprises a first input element that is an effective element and a second input element that is an ineffective element, and the second input activation comprises a third input element that is an effective element,
wherein the method further comprises generating the first input row data by pushing the third input element to the second input element.