The invention belongs to the technical field of
deep learning accelerators, and particularly relates to a design method of a
deep learning accelerator vector
processing unit based on
SystemC. Comprising the following steps: constructing a uniform vector
processing unit by using a
SystemC transaction-level model to replace
pooling and normalization modules accelerated by a traditional
deep learning algorithm; the interior of the vector
processing unit comprises a reconfigurable dynamic configuration so as to realize various
data processing modes of a
pooling operator, a normalization operator, a pixel summation operator and an up-sampling operator; the enhanced multi-dimensional DMA engine is used for carrying out directional data
slicing and moving on the input three-dimensional data cube in different dimensions according to the configuration parameters; according to the method, the problems of hardware islands,
low resource utilization rate, inflexible data transfer and the like existing in the original discrete design are solved, and the universality, the energy efficiency ratio and the execution efficiency of the accelerator are remarkably improved.