Neural network operation apparatus and method
AssigneesSamsung Electronics Co., Ltd., Seoul National University R&Db Foundation
InventorsHanwoong Jung, Soonhoi Ha, Donghyun Kang, Duseok Kang
Life of this patent
Estimating the term…
Abstract
A neural network operation apparatus and method is provided. The neural network operation apparatus includes a memory configured to store data for a neural network operation, and a processor configured to validate the data based on a determination that the neural network operation should be performed on the data, obtain a real memory address to perform the neural network operation based on a result of the validating and a virtual tensor address of the data, and perform the neural network operation based on the real memory address.
Claims12
A neural network operation apparatus, comprising:
- a memory, configured to store data for a neural network operation;
- and
- one or more processors, configured to:
- validate the data based on a determination that the neural network operation should be performed on the data and based on a modulo operation between a virtual tensor address of the data and a stride of the neural network operation, the virtual tensor address of the data corresponding to a position of the data on a virtual tensor;
- when a result of the validating is valid, obtain a real memory address to perform the neural network operation based on a value an output of a floor function of a value obtained by dividing the virtual tensor address of the data by the stride of the neural network operation;
- provide input data corresponding to the real memory address to an operator to perform the neural network operation;
- transmit an output of the operator to a corresponding real memory address;
- and
- perform inference operation associated with a neural network application by accelerating the neural network application based on the output of the operator transmitted to the corresponding real memory address.
A neural network operation apparatus, comprising:
- a memory, configured to store data for a neural network operation;
- and
- one or more processors configured to:
- obtain a real memory address, depending on a validity of the data, to store a result of the neural network operation based on an output of a floor function using a virtual tensor address of the data, a channel index of input data, a parameter of the neural network operation, and a number of channels of the input data, wherein the data is verified for the validity of the data based on verification that neural network operations should be performed on the data and based on a modulo operation between the virtual tensor address of the data and a stride of the neural network operation, the virtual tensor address of the data corresponding to a position of the data on a virtual tensor,
- provide the input data to an operator to perform the neural network operation,
- transmit an output of the operator to the real memory address, and
- perform inference operation associated with a neural network application by accelerating the neural network application based on the output of the operator transmitted to the corresponding real memory address.
A neural network operation method, comprising:
- receiving data for a neural network operation;
- validating the data based on a determination that the neural network operation should be performed on the data and based on a modulo operation between a virtual tensor address of the data and a stride of the neural network operation, the virtual tensor address of the data corresponding to a position of the data on a virtual tensor;
- when a result of the validating is valid, obtaining a real memory address to perform the neural network operation based on an output of a floor function of a value obtained by dividing the virtual tensor address of the data by the stride of the neural network operation;
- providing input data corresponding to the real memory address to an operator to perform the neural network operation based on the real memory address;
- transmit an output of the operator to a corresponding real memory address;
- and
- perform inference operation associated with a neural network application by accelerating the neural network application based on the output of the operator transmitted to the corresponding real memory address.
Dependencies are inferred from the claim text.