Search

CN-121997988-A - Parameter selection method and parameter selection system for real-time neural network operation architecture

CN121997988ACN 121997988 ACN121997988 ACN 121997988ACN-121997988-A

Abstract

The invention discloses a parameter selection method and a parameter selection system for a real-time neural network operation architecture. The parameter selection method comprises the steps of obtaining an access strategy combination for a target real-time neural network operation framework, wherein the access strategy combination comprises a plurality of access strategies. Each access policy defines a plurality of access parameters for use by the data access circuit in accessing the memory. The parameter selection method further comprises the steps of accessing the memory according to each access strategy by configuring a plurality of data access circuits, and simultaneously configuring a plurality of operation channels to execute a convolution operation process for each access strategy so as to obtain an optimized access strategy used for executing the convolution operation process.

Inventors

  • WU JIARONG

Assignees

  • 瑞昱半导体股份有限公司

Dates

Publication Date
20260508
Application Date
20241105

Claims (10)

  1. 1. A method for parameter selection for a real-time neural network computing architecture, comprising performing, by a computing device of a parameter selection system, the steps of: Obtaining an access policy combination for a target real-time neural network operation architecture, wherein the access policy combination comprises a plurality of access policies, and the target real-time neural network operation architecture comprises: A memory; A plurality of data access circuits connected to the memory via a bus, wherein each access policy defines a plurality of access parameters used when the plurality of data access circuits access the memory, and A plurality of operation channels respectively connected to the plurality of data access circuits and each including a plurality of processing element circuits; The memory is accessed according to each access strategy by configuring the plurality of data access circuits, and simultaneously, the convolution operation process is executed by configuring the plurality of operation channels aiming at each access strategy so as to acquire the optimized access strategy used for executing the convolution operation process.
  2. 2. The method of claim 1, wherein the memory comprises a plurality of memory blocks, and the plurality of access parameters in each access policy define an order in which the plurality of memory blocks are read by the plurality of data access circuits and a manner in which the plurality of memory blocks are read by the bus allocation, respectively.
  3. 3. The parameter selection method according to claim 2, wherein the convolution operation process includes: inputting the data to be operated obtained by reading the memory according to the corresponding access strategy into a plurality of channels of a convolutional neural network model, wherein each channel comprises a plurality of convolutional kernels, and And performing convolution operation according to the first direction step length and the second direction step length by using each convolution kernel to generate a plurality of output data, and recording corresponding data processing time.
  4. 4. A method of parameter selection according to claim 3, wherein each convolution kernel has a kernel size and the operation direction of the first and second direction steps is different.
  5. 5. The parameter selection method according to claim 4, wherein the step of obtaining an optimized access policy used for performing the convolution operation process includes obtaining, for each access policy, a data processing time spent performing the convolution operation process, and taking an access parameter having a shortest data processing time as the optimized access policy.
  6. 6. A parameter selection system for a real-time neural network operational architecture, comprising: computing device, and A target real-time neural network operational architecture, comprising: A memory; a plurality of data access circuits connected to the memory via a bus, and A plurality of operation channels respectively connected to the plurality of data access circuits and each including a plurality of processing element circuits; wherein the computing device is configured to perform the steps of: Obtaining an access policy combination for a target real-time neural network operation architecture, wherein the access policy combination comprises a plurality of access policies, each access policy defining a plurality of access parameters for use by the plurality of data access circuits in accessing the memory, and The memory is accessed according to each access strategy by configuring the plurality of data access circuits, and simultaneously, the convolution operation process is executed by configuring the plurality of operation channels aiming at each access strategy so as to acquire the optimized access strategy used for executing the convolution operation process.
  7. 7. The system of claim 6, wherein the memory comprises a plurality of memory blocks, and the plurality of access parameters in each access policy define an order in which the plurality of memory blocks are read by the plurality of data access circuits and a manner in which the plurality of memory blocks are read by the bus allocation, respectively.
  8. 8. The parameter selection system of claim 7, wherein the convolution operation process comprises: inputting the data to be operated obtained by reading the memory according to the corresponding access strategy into a plurality of channels of a convolutional neural network model, wherein each channel comprises a plurality of convolutional kernels, and And performing convolution operation according to the first direction step length and the second direction step length by using each convolution kernel to generate a plurality of output data, and recording corresponding data processing time.
  9. 9. The parameter selection system of claim 8, wherein each convolution kernel has a kernel size and the operation direction of the first direction step and the second direction step is different.
  10. 10. The parameter selection system of claim 9, wherein the step of obtaining an optimized access policy used to perform the convolution operation process comprises obtaining, for each access policy, a data processing time spent performing the convolution operation process, and taking as the optimized access policy an access parameter having a shortest data processing time.

Description

Parameter selection method and parameter selection system for real-time neural network operation architecture Technical Field The present invention relates to a method and a system, and more particularly, to a parameter selection method and a parameter selection system for a real-time neural network operation architecture. Background In recent years, artificial intelligence has rapidly developed, and neural network models are widely applied in the fields of life and science and technology. According to different application fields, the neural network model is divided into a non-real-time operation architecture and a real-time operation architecture. In a non-real-time computing architecture, all data needs to be loaded into memory and then computed. In the real-time neural network architecture, for example, when applied to the sound noise reduction function, the real-time operation needs to be performed simultaneously when data is input. However, the neural network architecture has various model parameters, and different combinations of model parameters can cause the neural network architecture to change in various ways. In addition, depending on the arithmetic circuit used, the setting used may also affect the result of the neural network execution. However, these parameters cannot be determined in advance how to adjust to obtain the best performance. Disclosure of Invention The invention aims to solve the technical problem of providing a parameter selection method and a parameter selection system for a real-time neural network operation architecture aiming at the defects of the prior art, and the method and the system can gradually find out better parameter combinations and improve operation efficiency in the process of processing data. In order to solve the above-mentioned problems, one of the technical solutions adopted in the present invention is to provide a parameter selection method for a real-time neural network operation architecture, which includes the steps of obtaining an access policy combination for a target real-time neural network operation architecture by a computing device of a parameter selection system, wherein the access policy combination includes a plurality of access policies, and the target real-time neural network operation architecture includes a memory, a plurality of data access circuits and a plurality of operation channels. The plurality of data access circuits are connected to the memory through a bus. Each access policy defines a plurality of access parameters for use by the plurality of data access circuits in accessing the memory. The operation channels are respectively connected with the data access circuits and each include a plurality of processing unit circuits. The parameter selection method further comprises accessing the memory according to each access strategy by configuring the plurality of data access circuits, and simultaneously configuring the plurality of operation channels to execute a convolution operation process for each access strategy so as to acquire an optimized access strategy used for executing the convolution operation process. In order to solve the above-mentioned technical problems, another technical solution adopted by the present invention is to provide a parameter selection system for a real-time neural network operation architecture, which includes a computing device and a target real-time neural network operation architecture. The target real-time neural network operation structure comprises a memory, a plurality of data access circuits and a plurality of operation channels. The plurality of data access circuits are connected to the memory through a bus. The operation channels are respectively connected with the data access circuits and each comprise a plurality of processing element circuits. The computing device is configured to acquire an access policy combination for a target real-time neural network operation architecture, wherein the access policy combination comprises a plurality of access policies, each access policy defines a plurality of access parameters used when the plurality of data access circuits access the memory, and the computing device is configured to access the memory according to each access policy by configuring the plurality of data access circuits, and simultaneously configure the plurality of operation channels to execute a convolution operation process for each access policy to acquire an optimized access policy used for executing the convolution operation process. The parameter selection method and the parameter selection system for the real-time neural network operation architecture have the advantages that the optimal parameter combination can be found out step by step in the process of processing data, and the operation efficiency can be effectively improved due to the fact that the gradual adjustment mechanism can be optimized immediately. For a further understanding of the nature and the technical aspects of