Low-power-consumption floating-point multiply-accumulate operation method for neural network reasoning acceleration
The invention discloses a low-power-consumption floating-point multiply-accumulate operation method for neural network reasoning acceleration, and the method comprises the steps: carrying out the preprocessing of an input floating-point number, carrying out the rounding of a decimal part before calc...
Gespeichert in:
Hauptverfasser: | , , , , , |
---|---|
Format: | Patent |
Sprache: | chi ; eng |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | The invention discloses a low-power-consumption floating-point multiply-accumulate operation method for neural network reasoning acceleration, and the method comprises the steps: carrying out the preprocessing of an input floating-point number, carrying out the rounding of a decimal part before calculation, adjusting the exponent bit of the decimal part, carrying out the pre-rounding and regularization of data which may be rounded, and saving the calculation of unnecessary precision. Meanwhile, in order to enable the accumulation circuit to be quickly executed, the decimal part of the index expressed by the scientific and technical method is adopted to replace the original original code expression method, but adopts the complement expression method. According to the method, the characteristics of multiply-accumulate operation are utilized, and the precision close to floating-point operation is realized by using fixed-point operation resources which consume less resources, so that the problem of balance betwee |
---|