A tensor convolution kernel acceleration chip and method

Abstract

This invention discloses a tensor convolution kernel acceleration chip, applicable to all deep learning networks involving convolution operations. This invention integrates a demultiplexer, modulator, delay weighting unit, coupling array, and detector—all components responsible for accelerating tensor convolution kernel operations—using photonic integration technology. Based on wavelength division multiplexing (WDM), the signal to be processed is loaded onto multiple optical carriers. Micro-loops and delay lines are used to achieve weighted and time-interleaved convolution kernel coefficients for different carrier signals. The weighted summation operation is then performed using the coupling array. This invention uses light as the information carrier and, based on multiple delay weighting units and the coupling array, accelerates tensor computation of signals, significantly improving the computational speed and energy efficiency of neural networks.

Description

Technical Field

[0001] This invention relates to the field of photonic integration technology, and in particular to a tensor convolution kernel acceleration chip and method. Background Technology

[0002] Stacking multidimensional data into tensors offers the opportunity to discover hidden structural features within the data that are invisible in bidirectional (matrix) data analysis. For example, multichannel representation of electroencephalogram (EEG) data is an efficient way to process neuroscience data, and tensors stacked across time, space, and spectrum are beneficial for detecting features in electromagnetic waveforms. Because tensors match the high-dimensional nature of the world, the concept of multichannel analysis has spawned a wide range of signal processing methods in fields such as life sciences, radar, data mining, and machine learning. In the basic operations of tensors, convolution can effectively extract structural features from data, filtering out target features as the convolution kernel traverses the tensor. Convolutional neural networks, as a microcosm of modern artificial intelligence (AI), are designed under the concept of multichannel tensor processing. Given that tensor processing, especially in the field of artificial intelligence, is consuming increasing amounts of computational resources, there is an urgent need for high-throughput and energy-efficient processors. Photonics has recently proven to be a promising candidate for building high-performance matrix processors. By designing photonic circuits as linear transformation functions, matrix multiplication can be performed as light passes through the circuit. The broadband spectrum of photonic circuits allows for clock frequencies to be increased to tens of gigahertz. Therefore, photonic circuits have proven to be superior GeMM processors with high throughput and high energy efficiency (see [Xu, Shaofu, et al. High-order tensor flow processing using integrated photonic circuits. arXiv preprint arXiv:2112.12322(2021)]). In fact, another advantage of photonics over electronics is the richness of available degrees of freedom in light. For example, linear transformations using wavelength, guiding mode, time, and space have been successfully studied. Summary of the Invention

[0003] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies by utilizing a delay-weighted unit containing a delay waveguide combined with a coupling array to achieve tensor computation of tensor signals in the optical domain. This addresses the limitations of computing power and high power consumption caused by the separation of storage and computation and data dimension conversion in traditional electronic technologies. Furthermore, the tensor kernel can be flexibly expanded, making it suitable for multi-dimensional data tensor computation. Except for the light source, all photonic components of the entire acceleration chip are integrated on a single chip, resulting in a compact, simple, small-sized, low-cost system with a flexibly expandable convolution kernel matrix.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This application discloses a tensor convolution kernel acceleration chip, which integrates a wavelength division multiplexer, a modulator, a delay weighting unit, a coupling array, and a detector.

[0006] The optical input terminal of the dewavelength division multiplexer is the same as the optical input terminal of the chip, used to receive multi-wavelength optical signals; the output terminal of the dewavelength division multiplexer is connected to the modulator, used to output sub-optical signals containing N wavelengths;

[0007] The optical input terminal of the modulator is connected to the optical output terminal of the dewavelength division multiplexer to receive sub-optical signals containing N wavelengths; the electrical input terminal of the modulator is used to receive the signal to be convolved; the optical output terminal of the modulator is connected to the delay weighting unit to output sub-modulated optical signals containing N wavelengths.

[0008] The optical input terminal of the delay weighting unit is connected to the optical output terminal of the modulator and is used to receive sub-modulated optical signals containing N wavelengths; the electrical input terminal of the delay weighting unit is used to receive convolution kernel matrix control signals; the optical output terminal of the delay weighting unit is connected to the coupling array and is used to output amplitude-weighted sub-modulated optical signals.

[0009] The optical input terminal of the coupling array is connected to the optical output terminal of the delay weighting unit for receiving amplitude-weighted sub-modulated optical signals; the electrical input terminal of the coupling array is used to receive coupling coefficient control signals; the output terminal of the coupling array is connected to a detector for outputting second-order amplitude-weighted sub-modulated optical signals containing N wavelengths.

[0010] The optical input terminal of the detector is connected to the optical output terminal of the coupling array and is used to receive a second-order amplitude-weighted sub-modulated optical signal containing N wavelengths; the optical output terminal of the detector is used to output the feature signal obtained after the signal to be convolved has completed the convolution operation.

[0011] Preferably, the delay weighting unit consists of a through waveguide, a coupling waveguide, and N micro-ring resonators; the output terminal of the coupling waveguide of the first micro-ring resonator is the optical output terminal of the delay weighting unit.

[0012] Preferably, the coupling array consists of several subarrays, each subarray consisting of several couplers and wavelength division multiplexers; the optical input terminal of the coupler in the first subarray is connected to the optical output terminal of the delay weighting unit as the optical input terminal of the coupling array; the optical output terminal of the coupler in each subarray is connected to the optical input terminal of the corresponding wavelength division multiplexer and the optical input terminal of the coupler in the next subarray; the optical output terminal of the wavelength division multiplexer is the optical output terminal of the coupling array.

[0013] Preferably, the N micro-ring resonators in the delay weighting unit have delay waveguides between each other at their through waveguide ends.

[0014] Preferably, the chip is integrated using a group III-V material integration process or a silicon-based integration process.

[0015] Preferably, the radii of the N microring resonators in the delay weighting unit increase sequentially, and the free spectral range corresponding to the microring resonator with the largest radius is greater than the spectral range occupied by the multi-wavelength optical signal.

[0016] This invention also discloses a method for accelerating tensor convolution kernels, comprising the following steps:

[0017] S1. The multi-wavelength optical signal is sent to the dewavelength division multiplexer, which divides the multi-wavelength optical signal into M sub-optical signals, each containing N wavelengths, and sends them to M modulators.

[0018] S2. The modulator receives M signals to be convolved, modulates the M sub-optical signals, each containing N wavelengths, to obtain M sub-modulated optical signals, each containing N wavelengths, and sends them to the corresponding M delay weighting units respectively.

[0019] S3. The delay weighting unit receives the convolution kernel matrix control signal and controls the coupling coefficients of the N micro-ring resonators of the delay weighting unit according to the convolution kernel matrix control signal. M sub-modulated optical signals, each containing N wavelengths, are sequentially coupled into the coupling waveguide according to different coupling coefficients to obtain M amplitude-weighted sub-modulated optical signals; and then sent into the coupling array.

[0020] S4. The coupling array has O sub-arrays. The coupling array receives coupling coefficient control signals and controls M amplitude-weighted sub-modulated optical signals to achieve second-level amplitude weighting, resulting in OM second-level amplitude-weighted sub-modulated optical signals containing N wavelengths. The OM second-level amplitude-weighted sub-modulated optical signals containing N wavelengths are then sent to the corresponding O wavelength division multiplexers. The wavelength division multiplexers combine the OM second-level amplitude-weighted sub-modulated optical signals containing N wavelengths into O weighted modulated multi-wavelength optical signals containing MN wavelengths. The O weighted modulated multi-wavelength optical signals containing MN wavelengths are then sent to the corresponding O detectors.

[0021] S5. The detector performs photoelectric conversion on O weighted modulated multi-wavelength optical signals containing MN wavelengths to obtain an electrical output signal, which is the feature signal obtained after the signal to be convolved completes the tensor convolution operation.

[0022] Preferably, in step S2, the signals to be convolved are obtained by tensor decomposition. The tensor is calculated by inputting the dimension of the signal to be processed and the number of signals to be processed. The signal to be processed is a one-dimensional signal or a one-dimensional signal obtained by flattening an actual two-dimensional signal.

[0023] Preferably, in step S3, the convolution kernel matrix control signal controls the coupling coefficients of the N micro-ring resonators in each delay weighting unit to achieve weighting of the convolution kernel matrix coefficients of the MN wavelength modulation signals. Specifically, the coupling coefficients of the micro-ring resonators are determined according to the magnitude of the corresponding convolution kernel matrix coefficients and the initial signal intensity of each wavelength in the multi-wavelength optical signal. The coupling coefficients of the micro-ring resonators are changed through thermo-optical effect or electro-optical effect. Each of the N micro-ring resonators in the delay weighting unit corresponds to a sub-optical signal.

[0024] The beneficial effects of this invention are:

[0025] 1) Based on the parallel processing capability of photons, this invention combines wavelength, time, and space in the optical domain to realize tensor calculation of tensor data, which can effectively avoid the increased computational complexity and power consumption caused by the separation of storage and computation in the electrical domain and the conversion of multidimensional data.

[0026] 2) The monolithic photonic integration of all functional components in this invention, compared with the use of general discrete optoelectronic devices, does not require additional optoelectronic functional devices, thereby simplifying the system, improving system stability, and greatly expanding the chip size.

[0027] 3) This invention achieves accelerated calculation of data convolution kernel convolution based on multiple delay weighting units embedded in cascaded delay waveguides. The scheme is simple and efficient. Compared with fiber dispersion technology, the system is compact, has strong anti-interference, and does not require dispersion calibration compensation.

[0028] 4) This invention controls the coefficients of the convolution kernel matrix based on multiple delay weighting units, which enables rapid updating of the coefficients of the convolution kernel matrix in tensor computation and is suitable for real-time data processing applications.

[0029] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the tensor convolution kernel acceleration chip structure of the present invention.

[0031] Figure 2 This is a schematic diagram of a specific embodiment of the tensor convolution kernel acceleration chip of the present invention.

[0032] Figure 3 This is a spectral distribution diagram of the multi-wavelength optical signal in the embodiment;

[0033] Figure 4 This is the spectral distribution diagram of the output signal of the first delay-weighted unit in the embodiment;

[0034] Figure 5 This is the spectral distribution diagram of the output signal of the second delay-weighted unit in the embodiment;

[0035] Figure 6 This is the spectral distribution diagram of the output signal of the Mth delay-weighted micro-ring unit in the embodiment;

[0036] Figure 7 This is a time series diagram showing the relationship between the output optical signal of the first wavelength division multiplexer in the embodiment and the wavelength.

[0037] Figure 8 This is a time series diagram showing the relationship between the output optical signal of the second wavelength division multiplexer in the embodiment and the wavelength.

[0038] Figure 9 This is a time series diagram showing the relationship between the output optical signal of the Oth wavelength division multiplexer in the embodiment and the wavelength. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0040] A tensor convolution kernel acceleration chip of the present invention is specifically as follows: Figure 1 As shown, the integrated photonic components include: a dewavelength division multiplexer (DWDM), a modulator, a delay weighting unit, a coupling array, and a detector (PD); the photonic components are connected by optical waveguides; the dewavelength division multiplexer has one optical input terminal and M optical output terminals, the optical input terminal is the optical input terminal of the entire chip, used to receive external multi-wavelength optical signals, and the M optical output terminals are connected to the optical input terminals of the corresponding modulators;

[0041] There are M modulators in total, each modulator having one electrical input terminal, one optical input terminal, and one optical output terminal. The optical output terminal is connected to the optical input terminal of the corresponding delay weighting unit, and the electrical input terminal is used to receive the external signal to be convolved. There are M delay weighting units in total, each delay weighting unit consisting of one through waveguide, one coupling waveguide, and N microring resonators. The output terminal of the coupling waveguide of the first microring resonator of each delay weighting unit is the optical output terminal of each delay weighting unit. The coupling array has O groups of OM couplers and O wavelength division multiplexing (WDM) arrays. Each coupler has one optical input and two optical outputs. The optical input of the first set of couplers is connected to the optical output of the delay-weighting unit, serving as the optical input of the coupling array. The two optical outputs of each set of couplers are respectively connected to the optical input of the next set of couplers and the optical input of the corresponding wavelength division multiplexer. Each wavelength division multiplexer has M optical inputs and one optical output. There are O detectors in total, each with one optical input connected to the optical output of the corresponding wavelength division multiplexer. The chip's operation is as follows: First, multi-wavelength optical signals are sent in... The demultiplexer divides a multi-wavelength optical signal into M sub-optical signals, each containing N wavelengths, and sends them to M modulators. The M signals to be convolved are then loaded onto their respective sub-optical signals through their modulators to obtain M sub-modulated optical signals. These M sub-modulated optical signals are then fed into M corresponding delay-weighting units. A control signal, based on the coefficients of the M sets of convolution kernel matrices, controls the coupling coefficients of the MN micro-rings within the M delay-weighting units to achieve amplitude weighting of the MN wavelength signals. The amplitude-weighted M sub-modulated optical signals are then fed into a coupling array, where the coupling coefficients control the signal... The M sub-modulated optical signals after amplitude weighting are subjected to second-level amplitude weighting to obtain OM second-level amplitude-weighted sub-modulated optical signals containing N wavelengths. The OM second-level amplitude-weighted sub-modulated optical signals are respectively sent to the corresponding O wavelength division multiplexers. The wavelength division multiplexers combine the OM second-level amplitude-weighted sub-modulated optical signals into O weighted modulated multi-wavelength optical signals containing MN wavelengths. The O weighted modulated multi-wavelength optical signals are respectively sent to the corresponding O detectors for photoelectric conversion to obtain O electrical signals. These signals are the feature signals obtained after the signal to be convolved completes the tensor convolution operation.

[0042] The convolution kernel matrix control signal consists of MN convolution kernel matrix control signals generated based on M sets of convolution kernel matrix coefficients. These control signals alter the coupling coefficients of the corresponding microring resonators through thermo-optic or electro-optic effects. Each microring resonator coupling coefficient is determined based on the magnitude of the convolution kernel matrix coefficients and the initial signal intensity of each wavelength in the multi-wavelength optical signal. The amplitudes of the MN wavelengths in the multi-wavelength optical signal may be equal or unequal. M and N are positive integers, representing the number of signals to be processed and the number of coefficients in each set of convolution kernel matrices, respectively. Preferably, these represent the maximum number of signals to be processed and the maximum number of convolution kernel matrix coefficients that can be supported. Preferably, the radii of the MN microring resonators in the delay-weighted microring array increase sequentially, each corresponding to a resonant wavelength, and the microring resonator with the largest radius corresponds to a free spectral range Δf. FSR It should be greater than the spectral range MNΔf occupied by the multi-wavelength optical signal.

[0043] The coupling coefficient control signal is used to control the coupling coefficients of OM couplers, thereby realizing the amplitude adjustment of N sub-modulated optical signals in the M-path weighted modulated multi-wavelength optical signal and realizing the second-level amplitude weighting.

[0044] The channel spacing of the dewavelength division multiplexer is NΔf, where Δf is the frequency spacing between two wavelengths in the multi-wavelength optical signal, thereby correspondingly dividing the multi-wavelength optical signal into M sub-optical signals, each containing N wavelengths.

[0045] The photonic components such as the demultiplexer, modulator, delay weighting unit, coupling array, and detector, as well as the optical waveguide, can all be fabricated and integrated using group III-V materials or silicon. The chip can be integrated using mature processes such as group III-V material integration technology or silicon-based integration technology.

[0046] To facilitate public understanding, the technical solution of the present invention will be further described in detail below through a specific embodiment. It should be noted that the multi-wavelength light source is generated by a multi-wavelength laser, laser array, mode-locked laser, femtosecond laser, optical frequency comb generator, or optical soliton optical frequency comb generator. In this embodiment, a mode-locked laser is preferred, and the signal is amplified by an optical amplifier. The convolution kernel matrix control signal and coupling coefficient control signal are emitted by a tensor kernel control signal. The characteristic signal is acquired and processed by a signal acquisition and processing unit. Furthermore, the amplitudes of the MN wavelength signals in the output signal of the multi-wavelength light source may be equal or unequal; in this embodiment, they are preferably equal.

[0047] like Figure 2As shown, the tensor convolution kernel acceleration chip in this embodiment includes: one mode-locked laser, one optical amplifier, one wavelength division multiplexer, M modulators, one signal to be convolved, M delay weighting units, one tensor kernel matrix control single signal, one coupling array (composed of OM couplers and O wavelength division multiplexers), O detectors, and one signal acquisition and processing unit, etc.

[0048] First, the mode-locked laser outputs multi-wavelength optical signals with equal intensity at each wavelength, which are then amplified by an optical amplifier. The intensity of each wavelength in the amplified multi-wavelength optical signal can be represented by a matrix as A = [A, A, A, ..., A]. T MN Its spectral distribution is as follows Figure 3 As shown, M and N are positive integers, representing the maximum number of signals that can be processed and the maximum number of coefficients in each convolutional kernel matrix, respectively, and A is the single-wavelength signal intensity. The amplified, equal-amplitude multi-wavelength optical signal is fed into a demultiplexer and divided into M sub-multi-wavelength optical signals, each containing N wavelengths. These sub-multi-wavelength optical signals can be represented as A. m =[A,A,A,…,A] T N A m Let x represent the m-th sub-multiwavelength optical signal, where m = 1, 2, ..., M. M modulators correspond one-to-one with the M output ports of the demultiplexer. The M sub-multiwavelength optical signals are respectively fed into the M modulators. The M signals to be processed output from the signal source are respectively used to modulate the intensity of the sub-multiwavelength optical signals through their corresponding modulators. By loading the M signals to be processed onto the corresponding sub-multiwavelength optical signals, M sub-multiwavelength modulated optical signals are obtained. The sequence of signals to be processed can be represented as x. m (n)=[x m (1),x m (2),x m (3),…,x m (P)], where n represents the discretization time index, x m (n) represents the m-th signal sequence to be processed, where P is the length of the signal to be processed. The signal to be processed is a one-dimensional signal or a one-dimensional signal obtained by flattening an actual two-dimensional signal. Flattening specifically involves converting a two-dimensional matrix into a one-dimensional matrix. The M dimensions of the signals to be processed output by the signal source can be represented by tensors as [D data ,S in ], where D data To determine the dimension of the input signal to be processed, for a one-dimensional signal, D data D represents the number of signal data; for a two-dimensional signal, D data = [W, H], where W and H represent the number of data points in the width and height of the two-dimensional signal, respectively, and S in S represents the number of signals to be processed.in =M. Sub-multi-wavelength modulated optical signal S Mod_m This can be represented using a matrix as follows:

[0049]

[0050] M sub-multi-wavelength modulated optical signals are sequentially coupled into M delay-weighted units using fiber waveguide coupling technology. A schematic diagram of the delay-weighted micro-ring unit is shown below. Figure 3 As shown, each delay-weighted microring unit consists of one through waveguide, one coupling waveguide, and N microring resonators. The microring resonators are separated from each other at the through waveguide ends by a length of Δl = cΔt / n. w Delayed waveguide, where n w The effective refractive index of the waveguide delay line is Δt = 1 / S M S is the duration of a single symbol of the signal to be processed. M The symbol rate of the signal to be processed is given. The resonant characteristics of the N microrings in each delay-weighting unit correspond sequentially to a wavelength. The convolution kernel matrix control signal output by the tensor kernel control signal first controls the resonant characteristics of the first microring resonator, causing the corresponding wavelength sub-intensity modulated optical signal transmitted in the through-pass waveguide to be coupled to the coupled waveguide with a specific coupling coefficient. The coupling coefficient is set according to the magnitude of the convolution kernel matrix coefficients, thus achieving weighting of the convolution kernel matrix coefficients. The sub-multi-wavelength modulated optical signal in the through-pass waveguide enters the delay waveguide of the through-pass waveguide after passing through the first microring resonator, achieving a delay of Δt. The delayed multi-wavelength intensity modulated optical signal is then weighted by the corresponding wavelength signal through the second microring resonator, and the weighting of all wavelength signals is completed sequentially after the delay. M sub-multi-wavelength weighted modulated optical signals are obtained at the output of the coupled waveguide.

[0051] Let the coefficients of the m-th convolution kernel matrix be W. mN =[w m1 ,w m2 ,w m3 ,…,w mN ] T Let the dimension of the convolution kernel be D. k This means that when the convolution kernel is one-dimensional, D k Let N be the number of coefficients in a one-dimensional convolution kernel, and D be the number of coefficients in a two-dimensional convolution kernel. k = [C,L], where C and L are the number of coefficients in the rows and columns of the two-dimensional convolution kernel, respectively, and C·L = N. The sub-multi-wavelength weighted modulated optical signal S output from the time-delayed weighted micro-ring unit coupled waveguide. Modcon_m It can be represented as:

[0052]

[0053] The spectrum of the multi-wavelength weighted modulation optical signal is as follows: Figure 4 , Figure 5and Figure 6 As shown, it should be noted that, for the public's ease of understanding, Figure 4 , Figure 5 and Figure 6 These correspond to the sub-multi-wavelength weighted modulated optical signals output from the 1st, 2nd, and Mth delay-weighted micro-ring units, respectively. Based on waveguide fiber coupling technology, the M sub-weighted intensity-modulated optical signals are coupled into an optical fiber and then input into a coupling array. The coupling array contains OM couplers and O wavelength division multiplexers. The coupling coefficients of the couplers are controlled by a coupling coefficient control signal output from the tensor kernel matrix signal, thereby achieving amplitude adjustment of the M sub-modulated optical signals in each weighted modulated multi-wavelength optical signal, realizing two-stage amplitude weighting. The coupling coefficients M of the OM couplers in the coupling array... con This can be represented as a matrix:

[0054]

[0055] v is the coupling coefficient for each coupler. O wavelength division multiplexers combine OM second-order amplitude-weighted sub-modulated optical signals into O weighted modulated multi-wavelength optical signals containing MN wavelengths. Let the dimension of the coupling array be dimension D. m =[S Min ,S Mout [S represents the number of effective input ports of the coupling array and the number of output ports of the wavelength division multiplexer.] Min =M, where S Mout =O. The relationship between the time series and wavelength of weighted modulated multi-wavelength optical signals is as follows: Figure 7 , Figure 8 , Figure 9 As shown. It should be noted that, Figure 7 The value in the figure corresponds to the time series and wavelength relationship of the weighted modulated multi-wavelength optical signal output from the first wavelength division multiplexer. Figure 8 The value in the figure corresponds to the time series and wavelength relationship of the weighted modulated multi-wavelength optical signal output from the second wavelength division multiplexer. Figure 9 The value in the figure corresponds to the time series and wavelength relationship of the weighted modulated multi-wavelength optical signal output from the Oth wavelength division multiplexer. After the O weighted modulated multi-wavelength optical signals are fed into O detectors for photoelectric conversion, the electrical output signal is obtained. The signal within the effective time sequence of the electrical output signal can be expressed as:

[0056]

[0057] Among them, S ca_o (r) represents the result of the r-th tensor calculation for the o-th detector. After acquiring the o electrical output signals, the acquisition and processing unit processes the valid timing signals to obtain the o tensor calculation results. That is, the data dimension after tensor acceleration can be represented by a tensor as [D]. data ,Sout ], where D data S represents the output signal dimension, which is the same as the input signal dimension. out S represents the number of signals to be processed. out =O. In summary, the dimension of the tensor convolution kernel can be represented as [D]. k D m ,S in ,S out The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A tensor convolution kernel acceleration chip, characterized in that: The chip is integrated by an optical demultiplexer, a modulator, a delay weighting unit, a coupling array and a detector; The optical input end of the optical demultiplexer is the optical input end of the chip, and is used for receiving a multi-wavelength optical signal; the output end of the optical demultiplexer is connected with the modulator, and is used for outputting a sub optical signal containing N wavelengths; The optical input end of the modulator is connected with the optical output end of the optical demultiplexer, and is used for receiving the sub optical signal containing N wavelengths; the electrical input end of the modulator is used for receiving a convolution signal; and the optical output end of the modulator is connected with the delay weighting unit, and is used for outputting a sub modulated optical signal containing N wavelengths; The optical input end of the delay weighting unit is connected with the optical output end of the modulator, and is used for receiving the sub modulated optical signal containing N wavelengths; the electrical input end of the delay weighting unit is used for receiving a convolution kernel matrix control signal; and the optical output end of the delay weighting unit is connected with the coupling array, and is used for outputting an amplitude weighted sub modulated optical signal; The optical input end of the coupling array is connected with the optical output end of the delay weighting unit, and is used for receiving the amplitude weighted sub modulated optical signal; the electrical input end of the coupling array is used for receiving a coupling coefficient control signal; and the output end of the coupling array is connected with the detector, and is used for outputting a second amplitude weighted sub modulated optical signal containing N wavelengths; The optical input end of the detector is connected with the optical output end of the coupling array, and is used for receiving the second amplitude weighted sub modulated optical signal containing N wavelengths; and the optical output end of the detector is used for outputting a feature signal obtained after the convolution operation of the convolution signal is completed; The coupling array is composed of a plurality of groups of sub arrays, and each sub array is composed of a plurality of couplers and a wavelength division multiplexer; the optical input end of the coupler of the first group of sub arrays is connected with the optical output end of the delay weighting unit as the optical input end of the coupling array; the optical output end of the coupler of each group of sub arrays is connected with the optical input end of the corresponding wavelength division multiplexer and the coupler of the next group of sub arrays; and the optical output end of the wavelength division multiplexer is the optical output end of the coupling array.

2. The tensor convolution kernel acceleration chip of claim 1, wherein: The delay weighting unit is composed of a straight-through waveguide, a coupling waveguide and N micro ring resonators; and the coupling waveguide output end of the first micro ring resonator is the optical output end of the delay weighting unit.

3. The tensor convolution kernel acceleration chip of claim 2, wherein: The N micro ring resonators in the delay weighting unit are provided with delay waveguides between each other at the straight-through waveguide end.

4. The tensor convolution kernel acceleration chip of claim 1, wherein: The chip is integrated based on a III-V material integration process or a silicon-based integration process.

5. The tensor convolution kernel acceleration chip of claim 2, wherein: The radii of the N micro ring resonators in the delay weighting unit increase in turn, and the micro ring resonator with the largest radius corresponds to a free spectral range greater than the spectral range occupied by the multi-wavelength optical signal.

6. A method for accelerating a tensor convolution kernel, the method comprising: The method comprises the following steps: S1, sending a multi-wavelength optical signal into an optical demultiplexer, and dividing the multi-wavelength optical signal into M sub optical signals each containing N wavelengths by the optical demultiplexer and sending the M sub optical signals into M modulators; S2, receiving M convolution signals by the modulators, modulating the M sub optical signals each containing N wavelengths to obtain M sub modulated optical signals each containing N wavelengths, and sending the M sub modulated optical signals into corresponding M delay weighting units; S3, the delay weighting unit receives the convolution kernel matrix control signal, controls the N micro-ring resonator coupling coefficients of the delay weighting unit according to the convolution kernel matrix control signal, sequentially couples M sub-modulated optical signals each containing N wavelengths into the coupling waveguide according to different coupling coefficients, and obtains M amplitude weighted sub-modulated optical signals; and sends them into the coupling array; S4, the coupling array has O groups of sub-arrays; The coupling array receives the coupling coefficient control signal, controls the M amplitude weighted sub-modulated optical signals to realize two-stage amplitude weighting through the coupling coefficient control signal, obtains OM two-stage amplitude weighted sub-modulated optical signals containing N wavelengths, sends the OM two-stage amplitude weighted sub-modulated optical signals containing N wavelengths into corresponding O wavelength division multiplexers respectively; the wavelength division multiplexer combines the OM two-stage amplitude weighted sub-modulated optical signals containing N wavelengths into O weighted modulated multi-wavelength optical signals containing MN wavelengths; and sends the O weighted modulated multi-wavelength optical signals containing MN wavelengths into corresponding O detectors respectively; S5, the detector performs photoelectric conversion on the O weighted modulated multi-wavelength optical signals containing MN wavelengths, obtains an electrical output signal, which is a feature signal obtained after the tensor convolution operation of the convolution signal.

7. The method of claim 6, wherein: The convolution signal in step S2 is obtained by tensor decomposition, and the tensor is calculated by the input signal dimension and the number of input signals; the input signal is a one-dimensional signal or a one-dimensional signal obtained by flattening an actual two-dimensional signal.

8. The method of claim 6, wherein: In step S3, the convolution kernel matrix control signal realizes the weighting of the MN wavelength modulation signal convolution kernel matrix coefficients by controlling the coupling coefficients of the N micro-ring resonators in each delay weighting unit. Specifically, the coupling coefficients of the micro-ring resonators are determined according to the size of the corresponding convolution kernel matrix coefficients and the initial signal intensity of each wavelength in the multi-wavelength optical signal, and the coupling coefficients of the micro-ring resonators are changed through the thermo-optic effect or the electro-optic effect. The N micro-ring resonators in each delay weighting unit correspond to a sub-optical signal.

Images