NEURAL NETWORK ACCELERATOR, ACCELERATION METHOD, AND APPARATUS

    公开(公告)号:US20230236891A1

    公开(公告)日:2023-07-27

    申请号:US18191134

    申请日:2023-03-28

    CPC classification number: G06F9/5027 G06F17/144 G06F7/523 G06F7/78

    Abstract: A neural network accelerator is provided, including: a preprocessing module (301), configured to perform first forward winograd transform on a target matrix corresponding to an input feature map, to obtain a transformed target matrix, where the preprocessing module (301) is further configured to perform second forward winograd transform on a convolution kernel, to obtain a transformed convolution kernel; a matrix operation module (302), configured to perform a matrix multiplication operation on a first matrix and a second matrix, to obtain a multiplication result, where the first matrix is constructed based on the transformed target matrix, and the second matrix is constructed based on the transformed convolution kernel; and a vector operation module (303), configured to perform inverse winograd transform on the multiplication result, to obtain an output feature map.

    MULTIPLIER
    3.
    发明申请

    公开(公告)号:US20220206757A1

    公开(公告)日:2022-06-30

    申请号:US17698068

    申请日:2022-03-18

    Abstract: A multiplier (500) configured to simultaneously implement a plurality of low bit width multiplication operations is provided. The multiplier (500) includes a multiplicator input end (550) for receiving two low bit width multiplicators, a multiplicand input end (560) for receiving two low bit width multiplicands, a mask circuit (540) for masking each low bit width multiplicator, and a multiplication operation circuit (502) for multiplying a mask result and a multiplicand. When a sum of bit widths of the multiplicators is smaller than the multiplicator input end (550) and a sum of bit widths of the multiplicands is smaller than the multiplicand input end (560), masking is performed on each of the low bit width multiplicators, so that the multiplier (500) may respectively implement two low bit width multiplication operations, which resolves a waste of hardware resources that occurs because the multiplier can only process multiplication operations in one data format.

Patent Agency Ranking