Pop count-based deep learning neural network computation method, multiply accumulator and device

A multiply accumulator for a deep learning neural network operation, the multiply accumulator comprising:

• one or more processors configured to:
• classify result values of first multiplications of weights (W) and input values (A) according to a result value repetition probability, each first multiplication result value comprising a type of a plurality of types which represent constant values;
• select between a first operation method and a second operation method based on the classified multiplication result values;
• the first operation method comprising:
• generating, using an encoder, one-hot encoding codes according to each type of first multiplication result values from the first multiplications;
• determining, using a pop counter, pop-counts for each generated code of the one-hot encoding codes;
• and
• performing, with a multiply-accumulate operation circuit, a second multiplication between each type of the first multiplication result values and each count value of the pop-counts, wherein each type of the first multiplication results values is a different constant value;
• the second operation method comprising:
• performing the first multiplications by applying the weights and input values to at least one multiplier circuit of a multiply-accumulate hardware.

An electronic device for performing an inference operation of a deep learning neural network, the electronic device comprising:

• a memory configured to store weights (W) of the deep learning neural network and input values (A) that are input to an input layer of the deep learning neural network;
• and
• a processor configured to:
• classify result values of first multiplications of weights (W) and input values (A) according to a result value repetition probability, each first multiplication result value comprising a type of a plurality of types which represent constant values;
• select between a first operation method and a second operation method based on the classified multiplication result values;
• the first operation method comprising:
• determining an accumulated result value for the first multiplications between the weights and the input values that are stored in the memory;
• generating, using an encoder, one-hot encoding codes according to each type of first multiplication result values for the first multiplications, respectively;
• determining, using a pop counter, pop-counts for each generated code of the one-hot encoding codes;
• and
• performing, with a multiply-accumulate operation circuit, a second multiplication between each type of the first multiplication result values and each count value of the pop-counts, wherein each type of the first multiplication results values is a different constant value;
• the second operation method comprising:
• performing the first multiplications by applying the weights and input values to at least one multiplier circuit of a multiply-accumulate hardware.