One-dimensional convolution acceleration device and method for complex neural network

Abstract

The invention provides a one-dimensional convolution acceleration device of a complex neural network. The one-dimensional convolution acceleration device comprises an acceleration device module and anexternal storage module, the acceleration device module comprises a plurality of calculation units and a buffer calculation unit; the buffer calculation unit comprises a one-dimensional convolution calculation unit, a weight buffer and a feature map buffer; the plurality of calculation units are used for respectively calculating a real part and an imaginary part and outputting results to the output feature map storage area; and the one-dimensional convolution calculation unit is used for reading data from the weight buffer and the feature map buffer and performing calculation respectively, and a calculation result is output to the plurality of calculation units. The invention further provides a one-dimensional convolution acceleration method of the complex neural network based on the device. According to the method, the utilization rate of a computing unit can be improved, parallel real part and imaginary part computing is carried out for a complex number value storage format, the problem of channel intersection of complex convolution is solved, and one-dimensional convolution computing of the complex neural network is accelerated.

Description

technical field

[0001] The invention belongs to the field of hardware acceleration realization design of a neural network algorithm, in particular to a one-dimensional convolution acceleration device and method of a complex neural network. Background technique

[0002] A complex neural network refers to a neural network structure in which weight parameters and feature maps are represented by complex numbers, including network layers such as complex convolutions, complex activation functions, and complex batch normalization. Complex neural networks are mainly used in fields that require two parameters, amplitude and phase, to describe data. For example, in the radio frequency field, radio fingerprint identification transmission is carried out based on I / Q signals, and in the audio field, automatic music transcription, speech recognition, etc. The input data for this type of problem is a one-dimensional complex time series, and the processing involves the calculation of compl...

Examples

Embodiment 1

[0053] The first embodiment of the present invention provides a one-dimensional convolution calculation acceleration device for a complex neural network, and the structural diagram is as follows figure 1 shown. The acceleration device 100 is connected to the external storage 200; the external storage 200 stores a weight storage area 210 for input calculation, an input feature map storage area 220 and a calculation result output feature map storage area 230.

[0054] The acceleration device 100 includes a one-dimensional convolution calculation unit 110 , a weight buffer 120 , a feature map buffer 130 , and a complex number calculation unit 140 . Among them, the one-dimensional convolution calculation unit 110, the weight buffer 120, and the feature map buffer 130 are all four in number. Each weight buffer 120 is connected to the weight storage area 210 via a bus, and each feature map buffer 130 is connected to the input feature map storage area 220 via a bus. Each weight buf...

Embodiment 2

[0069] The second embodiment of the present invention provides a one-dimensional convolution calculation acceleration method for a complex neural network, the flow chart is as follows Figure 5 shown, including the following steps:

[0070] S100, the weight data and the input feature map data are respectively transmitted from the weight storage area 210 and the input feature map storage area 220 to the weight buffer 120 and the feature map buffer 130 .

[0071] All parameters of the neural network are stored in the weight storage area 210 . If it is a real number neural network, the input feature map storage area 220 stores 4 different input feature maps, and the input feature map channel is C i ; If it is a complex neural network, the input feature map storage area 220 stores one input feature map, and the input feature map channel is 2C i , where the former C i Channels are real data, after C i Channels are the imaginary part data.

[0072] S200, the one-dimensional con...

technical field

[0001] The invention belongs to the field of hardware acceleration realization design of a neural network algorithm, in particular to a one-dimensional convolution acceleration device and method of a complex neural network. Background technique

[0002] A complex neural network refers to a neural network structure in which weight parameters and feature maps are represented by complex numbers, including network layers such as complex convolutions, complex activation functions, and complex batch normalization. Complex neural networks are mainly used in fields that require two parameters, amplitude and phase, to describe data. For example, in the radio frequency field, radio fingerprint identification transmission is carried out based on I / Q signals, and in the audio field, automatic music transcription, speech recognition, etc. The input data for this type of problem is a one-dimensional complex time series, and the processing involves the calculation of compl...

Examples

Embodiment 1

[0053] The first embodiment of the present invention provides a one-dimensional convolution calculation acceleration device for a complex neural network, and the structural diagram is as follows figure 1 shown. The acceleration device 100 is connected to the external storage 200; the external storage 200 stores a weight storage area 210 for input calculation, an input feature map storage area 220 and a calculation result output feature map storage area 230.

[0054] The acceleration device 100 includes a one-dimensional convolution calculation unit 110 , a weight buffer 120 , a feature map buffer 130 , and a complex number calculation unit 140 . Among them, the one-dimensional convolution calculation unit 110, the weight buffer 120, and the feature map buffer 130 are all four in number. Each weight buffer 120 is connected to the weight storage area 210 via a bus, and each feature map buffer 130 is connected to the input feature map storage area 220 via a bus. Each weight buf...

Embodiment 2

[0069] The second embodiment of the present invention provides a one-dimensional convolution calculation acceleration method for a complex neural network, the flow chart is as follows Figure 5 shown, including the following steps:

[0070] S100, the weight data and the input feature map data are respectively transmitted from the weight storage area 210 and the input feature map storage area 220 to the weight buffer 120 and the feature map buffer 130 .

[0071] All parameters of the neural network are stored in the weight storage area 210 . If it is a real number neural network, the input feature map storage area 220 stores 4 different input feature maps, and the input feature map channel is C i ; If it is a complex neural network, the input feature map storage area 220 stores one input feature map, and the input feature map channel is 2C i , where the former C i Channels are real data, after C i Channels are the imaginary part data.

[0072] S200, the one-dimensional con...

Images