Neural core, neural processor, and data processing method thereof
A neural core, a neural processor, and a data processing method thereof are provided. The neural core includes a processing unit configured to generate output activation by performing calculation of input activation and a weight, a memory load/store unit configured to generate a memory calculation request requesting one-dimensional calculation of summing the output activation and target data, and a memory configured to receive the output activation and the memory calculation request, import the target data from a memory address in response to the memory calculation request, generate output data by performing the one-dimensional calculation of the target data and the output activation, and store the generated output data in the memory address.
1 . A neural processing device, comprising:
a memory comprising one-dimensional calculation circuitry and a plurality of memory cells; and
a processor operably coupled to the memory,
wherein the processor is configured to cause:
receiving input activation values and weight values for a neural network;
performing a two-dimensional calculation with the input activation values and the weight values to generate a first value for the neural network; and
requesting the memory to perform a one-dimensional calculation with the first value and a second value in a memory cell referenced by a memory address by using the one-dimensional calculation circuitry within the memory to generate a third value for the neural network and to store the third value in the memory cell referenced by the memory address.
2 . The processing device of claim 1 , further comprising:
load-store unit circuitry configured to load and store operations between the memory and the processor,
wherein requesting the memory further comprises:
sending the first value for the neural network to the load-store unit circuitry.
3 . The processing device of claim 2 ,
wherein the load-store unit circuitry is configured to cause:
loading the input activation values and the weight values from the memory to send the input activation values and the weight values to the processor;
in response to receiving the first value, executing an instruction to send a request requesting the memory to perform the one-dimensional calculation with the first value and the second value stored in the memory cell referenced by the memory address indicated by the instruction by using the one-dimensional calculation circuitry within the memory to generate the third value and to store the third value in the memory cell referenced by the memory address indicated by the instruction.
4 . The processing device of claim 3 ,
wherein the memory is configured to cause:
in response to receiving the request, reading the second value stored in the memory cell referenced by the memory address indicated by the request without using the load-store unit circuitry;
performing the one-dimensional calculation with the first value and the second value to generate the third value by using the one-dimensional calculation circuitry within the memory; and
storing the third value in the memory cell referenced by the memory address indicated by the request.
5 . A method performed by a processing device comprising a memory and a processor operably coupled to the memory, wherein the memory comprises one-dimensional calculation circuitry and a plurality of memory cells, the method comprising:
receiving, by the processor, input activation values and weight values for a neural network;
performing, by the processor, a two-dimensional calculation with the input activation values and the weight values to generate a first value for the neural network; and
requesting, by the processor, the memory to perform a one-dimensional calculation with the first value and a second value in a memory cell referenced by a memory address by using the one-dimensional calculation circuitry within the memory to generate a third value for the neural network and to store the third value for the neural network in the memory cell referenced by the memory address.
6 . The method of claim 5 , wherein the processing device further comprises load-store unit circuitry configured to load and store operations between the memory and the processor, the method further comprising:
sending, by the processor, the first value for the neural network to the load-store unit circuitry.
7 . The method of claim 6 , further comprising:
loading, by the load-store unit circuitry, the input activation values and the weight values from the memory to send the input activation values and the weight values to the processor; and
in response to receiving the first value, executing, by the load-store unit circuitry, an instruction to send a request requesting the memory to perform the one-dimensional calculation with the first value and the second value stored in the memory cell referenced by the memory address indicated by the instruction by using the one-dimensional calculation circuitry within the memory to generate the third value and to store the third value in the memory cell referenced by the memory address indicated by the instruction.
8 . The method of claim 7 , further comprising
in response to receiving the request, reading, by the one-dimensional calculation circuitry within the memory, the second value stored in the memory cell referenced by the memory address indicated by the request without using the load-store unit circuitry;
performing, by the one-dimensional calculation circuitry within the memory, the one-dimensional calculation with the first value and the second value to generate the third value; and
storing the third value in the memory cell referenced by the memory address indicated by the request.
9 . The method of claim 8 , wherein the one-dimensional calculation circuitry is associated with a set of memory cells belonging to a bank in the memory.
10 . The method of claim 9 ,
wherein the second value stored in the memory cell referenced by the memory address is read and the third value is stored via a port which is used by both the one-dimensional calculation circuitry and a controller controlling the bank.