FPGA accelerator of LSTM neural network and acceleration method of FPGA accelerator

Abstract

The invention provides an FPGA accelerator of an LSTM neural network and an acceleration method of the FPGA accelerator. The accelerator comprises a data distribution unit, an operation unit, a control unit and a storage unit; the operation unit comprises a sparse matrix vector multiplication module, a nonlinear activation function module and an element-by-element multiplication and addition calculation module; the control unit sends a control signal to the data distribution unit, and the data distribution unit reads an input excitation value and a neural network weight parameter from the storage unit and inputs the input excitation value and the neural network weight parameter to the operation unit for operation. The operation resources are uniformly distributed to each operation unit according to the number of the non-zero weight values, so that idling of operation resources is avoided, and the operation performance of the whole network is improved. Meanwhile, the pruned neural network is stored in the form of the sparse network, the weight value of each column is stored in the same address space, the neural network is coded according to the row index, and the operation performance and the data throughput rate are improved under the condition that the precision is guaranteed.

Description

technical field [0001] The invention relates to the field of computer hardware acceleration, in particular to an LSTM neural network-oriented FPGA accelerator and an acceleration method thereof. Background technique [0002] RNN neural network (Recurrent Neural Network) is usually used for processing sequence signal data. Due to its memory for historical data, RNN network can be used in machine translation, speech recognition, user behavior learning and other fields. The development of the variant LSTM network of the RNN network can solve the problem of gradient explosion and gradient disappearance in RNN network training. Because the LSTM network shows higher accuracy than the RNN network, the academic research on the accelerator of the LSTM network is blowout. [0003] Based on the characteristics of the LSTM network, its calculation process is a calculation-intensive operation. Using a general-purpose hardware platform such as a CPU to execute the LSTM algorithm does no...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the technical problem of unbalanced load in the operation of sparse neural network, the present invention proposes an FPGA accelerator of LSTM neural network and its acceleration method

Images