Non-linear processing method and device based on all-optical computing architecture

Abstract

The embodiment of the invention provides a nonlinear processing method and device based on an all-optical computing architecture. According to the method, the problems of high hardware resource occupancy rate, low calculation efficiency and high power consumption caused by complex nonlinear function calculation in the deep learning reasoning and training process can be solved. Besides, an electronic logic unit is replaced by a micro-ring resonator, a Mach-Zehnder interferometer and phase modulation, wavelength division multiplexing and multi-ring resonance are utilized, all-optical parallel processing of a matrix block is supported, the photoelectric conversion frequency of nonlinear function processing in photoelectric hybrid calculation is reduced, delay is remarkably reduced, and the overall performance is improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of optical computing technology, and in particular, to a non-linear processing method and device based on an all-optical computing architecture. Background Art

[0002] With the rapid development of artificial intelligence technology, deep learning models have been widely used in many fields. However, in the process of model inference and training, a large number of complex non-linear functions and mathematical operations are involved, such as multiplication, division, exponential operations, and calculations of activation functions. These operations usually require a large amount of hardware resources and computing time. Especially in an optoelectronic hybrid computing architecture, since optical operations cannot directly complete non-linear function calculations, optoelectronic conversion needs to be performed frequently, increasing the computing delay and hardware complexity. Summary of the Invention

[0003] In view of this, the embodiments of the present application provide a non-linear processing method and device based on an all-optical computing architecture, which can reduce the number of optoelectronic conversions for non-linear function processing in optoelectronic hybrid computing, significantly reduce the delay, and improve the overall performance.

[0004] In a first aspect, a non-linear processing method based on an all-optical computing architecture is provided. The all-optical computing architecture includes a micro-ring resonator MRR and a Mach-Zehnder interferometer MZI. The non-linear processing method includes: inputting an optical signal matrix block X into the micro-ring resonator MRR to output a first optical signal, where the first optical signal is the maximum optical signal in the optical signal matrix block X; sequentially inputting the optical signal matrix block X and the first optical signal into a cascaded Mach-Zehnder interferometer MZI and a series of multiple micro-ring resonators MRR to obtain a second optical signal corresponding to any optical signal in the optical signal matrix block X, where the second optical signal is the optical signal corresponding to the exponential power t with base 2, where t = (x i -x max )·log 2 e,x i is the value of any optical signal, and x max is the value of the maximum optical signal; sequentially inputting the second optical signal into a low-wavelength channel and a high-wavelength channel to output a third optical signal corresponding to any optical signal, where the third optical signal is the optical signal corresponding to the local exponential value 2 t of any optical signal, where The low-wavelength channel includes a Mach-Zehnder interferometer MZI and a micro-ring resonator MRR, and is used to output 2 r corresponding optical signal, r is the fractional part of t, and the high-wavelength channel includes q cascaded Mach-Zehnder interferometers MZI, and is used to output 2 qThe corresponding optical signal, where q is the integer part of t; input the third optical signal corresponding to each optical signal in the optical signal matrix block X into a plurality of cascaded microring resonators MRR, and output a fourth optical signal, where the fourth optical signal is the sum of the third signals corresponding to all optical signals in the optical signal matrix block X, and the fourth optical signal is the optical signal corresponding to the local exponential value of the optical signal matrix block X; modulate the third signal corresponding to any optical signal into an optical signal corresponding to a first wavelength and modulate the fourth optical signal into an optical signal corresponding to a second wavelength, where the optical signal corresponding to the first wavelength is the corresponding optical signal, and the optical signal corresponding to the second wavelength is the corresponding optical signal, k 1 and v 1 are the parameters of the first wavelength, k 2 and v 2 are the parameters of the second wavelength, v 1 and v 2 are constants less than 1; based on the magnitudes of v 1 and v 2 , input the optical signal corresponding to the first wavelength and the optical signal corresponding to the second wavelength into a microring resonator MRR array and a cascaded Mach-Zehnder interferometer MZI in sequence, and output a fifth optical signal corresponding to any optical signal, where the fifth optical signal is the optical signal corresponding to the ratio of N 1 and N 2 , where, if v 1 -v 2 is greater than or equal to 0, the fifth optical signal is the corresponding optical signal; if v 1 -v 2 is less than 0, the fifth signal is the corresponding optical signal.

[0005] In a possible implementation, the non-linear processing method further includes: storing the value of the first optical signal corresponding to each optical signal matrix block X in a plurality of the optical signal matrix blocks X in an optical buffer ring OBR, and outputting a sixth signal, where the sixth signal is the optical signal corresponding to the maximum value x' max among the values of the first optical signals corresponding to the plurality of optical signal matrix blocks X; input the sixth signal and the first signal corresponding to any one of the optical signal matrix blocks X into a cascaded Mach-Zehnder interferometer MZI and a plurality of cascaded microring resonators MRR in sequence, to obtain a seventh signal corresponding to any one of the optical signal matrix blocks X, where the seventh signal is the optical signal corresponding to the exponential power t' with base 2, where t' = x max -x' max; Input the seventh signal into the low-wavelength channel and the high-wavelength channel in sequence, and output the eighth optical signal corresponding to any optical signal matrix block X, where the eighth optical signal is 2 t′ The corresponding optical signal, where, The low-wavelength channel is used to output 2 r′ The corresponding optical signal, r' is the fractional part of t', and the high-wavelength channel includes q' cascaded Mach-Zehnder interferometers MZI for outputting 2 q′ The corresponding optical signal, q' is the integer part of t'; Modulate the fifth optical signal corresponding to any optical signal in the optical signal matrix block X into an optical signal corresponding to the third wavelength and modulate the eighth optical signal corresponding to the optical signal matrix block X into an optical signal corresponding to the fourth wavelength. The optical signal corresponding to the third wavelength is The corresponding optical signal, and the optical signal corresponding to the fourth wavelength is The corresponding optical signal, k 3 and v 3 Are parameters of the third wavelength, k 4 and v 4 Are parameters of the fourth wavelength, v 3 and v 4 Are constants less than 1; Based on the magnitudes of v 3 and v 4 , Input the optical signal corresponding to the third wavelength and the optical signal corresponding to the fourth wavelength into the micro-ring resonator MRR array and the cascaded Mach-Zehnder interferometer MZI in sequence, and output the ninth optical signal corresponding to any optical signal in the optical signal matrix block X. The ninth optical signal is the optical signal corresponding to the product of N 3 and N 4 , And the ninth optical signal is the optical signal corresponding to the global exponential value of any optical signal in the optical signal matrix block X. Among them, if v 3 +v 4 Is less than 1, the ninth optical signal is The corresponding optical signal; If v 3 +v 4 Is greater than or equal to 1, the ninth optical signal is The corresponding optical signal.

[0006] Second aspect, a non-linear processing device based on an all-optical computing architecture is provided. The all-optical computing architecture includes a microring resonator MRR and a Mach-Zehnder interferometer MZI. The non-linear processing device includes: a maximum light selection module for inputting an optical signal matrix block X into the microring resonator MRR and outputting a first optical signal, where the first optical signal is the maximum optical signal in the optical signal matrix block X; an optical phase shift module for sequentially inputting the optical signal matrix block X and the first optical signal into a cascaded Mach-Zehnder interferometer MZI and a series of multiple microring resonators MRR to obtain a second optical signal corresponding to any optical signal in the optical signal matrix block X, where the second optical signal is the optical signal corresponding to the exponential power t with base 2, and t=(x i -x max )·log 2 e,x i is the value of the any optical signal, x max is the value of the maximum optical signal; an optical approximate exponent module for sequentially inputting the second optical signal into a low-wavelength channel and a high-wavelength channel and outputting a third optical signal corresponding to the any optical signal, where the third optical signal is the optical signal corresponding to the local exponential value 2 t of the any optical signal, and where the low-wavelength channel includes a Mach-Zehnder interferometer MZI and a microring resonator MRR for outputting the optical signal corresponding to 2 r , r is the fractional part of t, the high-wavelength channel includes q cascaded Mach-Zehnder interferometers MZI for outputting the optical signal corresponding to 2 q , q is the integer part of t; an optical wavelength division multiplexing accumulation module for inputting the third optical signal corresponding to each optical signal in the optical signal matrix block X into a series of multiple microring resonators MRR and outputting a fourth optical signal, where the fourth optical signal is the sum of the third signals corresponding to all optical signals in the optical signal matrix block X, and the fourth optical signal is the optical signal corresponding to the local exponential value of the optical signal matrix block X; an optical division shift module for modulating the third signal corresponding to the any optical signal into an optical signal corresponding to a first wavelength and modulating the fourth optical signal into an optical signal corresponding to a second wavelength, where the optical signal corresponding to the first wavelength is corresponding optical signal, the optical signal corresponding to the second wavelength is corresponding optical signal, k 1 and v 1 are parameters of the first wavelength, k 2 and v 2 are parameters of the second wavelength, v 1 and v 2 are constants less than 1; and based on v 1 and v2 According to the size, the optical signals corresponding to the first wavelength and the optical signals corresponding to the second wavelength are sequentially input into the microring resonator MRR array and the cascaded Mach-Zehnder interferometer MZI, and the fifth optical signal corresponding to any of the optical signals is output. The fifth optical signal is an optical signal corresponding to the ratio of N 1 and N 2 . If v 1 -v 2 is greater than or equal to 0, the fifth optical signal is the optical signal corresponding to . If v 1 -v 2 is less than 0, the fifth signal is the optical signal corresponding to .

[0007] In a possible implementation, the maximum optical selection module is further configured to: store the values of the first optical signals corresponding to each optical signal matrix block X in the multiple optical signal matrix blocks X in the optical buffer ring OBR, and output a sixth signal. The sixth signal is the maximum value x' among the values of the first optical signals corresponding to the multiple optical signal matrix blocks X max corresponding optical signal; the optical phase shift module is further configured to: sequentially input the sixth signal and the first signal corresponding to any one of the optical signal matrix blocks X into the cascaded Mach-Zehnder interferometer MZI and multiple serially connected microring resonators MRR to obtain the seventh signal corresponding to any one of the optical signal matrix blocks X. The seventh signal is an optical signal corresponding to the exponential power t' with base 2, where t' = x max -x' max ; the optical approximate exponent module is further configured to: input the seventh signal into the low-wavelength channel and the high-wavelength channel in sequence, and output the eighth optical signal corresponding to any one of the optical signal matrix blocks X. The eighth optical signal is an optical signal corresponding to 2 t′ . Among them, the low-wavelength channel is used to output an optical signal corresponding to 2 r′ , r' is the fractional part of t', and the high-wavelength channel includes q' cascaded Mach-Zehnder interferometers MZI for outputting an optical signal corresponding to 2 q′ , q' is the integer part of t'; the nonlinear processing device further includes: an optical multiplication and shift module, configured to: modulate the fifth optical signal corresponding to any optical signal in the optical signal matrix block X into an optical signal corresponding to a third wavelength and modulate the eighth optical signal corresponding to the optical signal matrix block X into an optical signal corresponding to a fourth wavelength. The optical signal corresponding to the third wavelength is corresponding optical signal, the optical signal corresponding to the fourth wavelength is corresponding optical signal, k3 and v 3 is a parameter of the third wavelength, k 4 and v 4 is a parameter of the fourth wavelength, v 3 and v 4 is a constant less than 1; and based on v 3 and v 4 the magnitude of, the optical signal corresponding to the third wavelength and the optical signal corresponding to the fourth wavelength are sequentially input into a microring resonator MRR array and a cascaded Mach-Zehnder interferometer MZI, and the ninth optical signal corresponding to any optical signal in the optical signal matrix block X is output, and the ninth optical signal is the optical signal corresponding to the product of N 3 and N 4 and the ninth optical signal is the optical signal corresponding to the global exponential value of any optical signal in the optical signal matrix block X, where if v 3 + v 4 is less than 1, the ninth optical signal is the corresponding optical signal; if v 3 + v 4 is greater than or equal to 1, the ninth optical signal is the corresponding optical signal.

[0008] In a third aspect, a non-linear processing device based on an all-optical computing architecture is provided, including: a processor, and a memory communicatively connected to the processor, where the memory stores program instructions executable by the processor, and the processor can execute the non-linear processing method based on the all-optical computing architecture in the first aspect and any possible implementation manner in the first aspect by invoking the program instructions.

[0009] In a fourth aspect, a computer-readable storage medium is provided, and the computer-readable storage medium includes a stored program, where when the program runs, it controls the device where the computer-readable storage medium is located to execute the non-linear processing method based on the all-optical computing architecture in the first aspect and any possible implementation manner in the first aspect.

[0010] Based on the above technical solutions, it is possible to solve the problems of high hardware resource occupancy rate, low computing efficiency, and high power consumption caused by complex non-linear function calculations (such as exponential operations, activation function calculations, etc.) during the deep learning inference and training processes. In addition, by using microring resonators MRR, Mach-Zehnder interferometers MZI and phase modulation to replace electronic logic units, and using wavelength division multiplexing and multi-ring resonance, it supports all-optical parallel processing of matrix blocks, reduces the number of optoelectronic conversions for non-linear function processing in optoelectronic hybrid computing, significantly reduces latency, and improves overall performance. Description of the Drawings

[0011] To more clearly illustrate the technical solutions of the embodiments of the present application, the accompanying drawings required for use in the embodiments of the present application will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other accompanying drawings can be obtained based on these drawings without creative efforts.

[0012] Figure 1 It is a schematic block diagram of a non-linear processing method based on an all-optical computing architecture provided by an embodiment of the present application.

[0013] Figure 2 It is another schematic block diagram of a non-linear processing method based on an all-optical computing architecture provided by an embodiment of the present application.

[0014] Figure 3 It is a schematic block diagram of a non-linear processing device based on an all-optical computing architecture provided by an embodiment of the present application. Detailed implementation manners

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts fall within the scope of protection of the present application.

[0016] With the rapid development of artificial intelligence technology, deep learning models have been widely used in many fields. However, in the process of model inference and training, a large number of complex non-linear functions and mathematical operations are involved, such as multiplication, division, exponential operations, and calculations of activation functions. These operations usually require a large amount of hardware resources and computing time. Especially in an optoelectronic hybrid computing architecture, since optical operations cannot directly complete non-linear function calculations, optoelectronic conversions need to be performed frequently, increasing the computing delay and hardware complexity.

[0017] In related technologies, it is mostly based on Digital Signal Processing (DSP) hardware or Application-Specific Integrated Circuit (ASIC) design, and attempts to improve computational efficiency through linear approximation, quantization, or pipelined architecture optimization. However, these methods still have problems such as high hardware overhead, high power consumption, and long computational latency in high-precision operations, making it difficult to meet the low-latency and high-bandwidth requirements of optical computing architectures. Moreover, in optoelectronic hybrid systems, frequent optoelectronic conversions will result in signal loss and computational latency. Especially in the field of deep learning, many common activation functions, such as Softmax, Sigmoid, etc., can be abstractly modeled through the exponential function. How to design a hardware-friendly optimization method to further reduce power consumption and hardware complexity while ensuring computational accuracy has become the difficulty and focus of current research.

[0018] In view of this, the embodiments of the present application provide a non-linear processing method and device based on an all-optical computing architecture, which can solve the problems of high hardware resource occupancy, low computational efficiency, and high power consumption caused by the calculation of complex non-linear functions (such as exponential operations, activation function calculations, etc.) during the deep learning inference and training processes. In addition, by using a microring resonator (MRR), a Mach-Zehnder interferometer (MZI), and phase modulation to replace electronic logic units, and utilizing wavelength division multiplexing and multi-ring resonance, it supports all-optical parallel processing of matrix blocks, reduces the number of optoelectronic conversions for non-linear function processing in optoelectronic hybrid computing, significantly reduces latency, and improves overall performance.

[0019] 1. An all-optical computing architecture is a computing system based on optical devices and optical signal processing. Its core idea is to use the characteristics of light (such as high-speed transmission, parallel processing, low power consumption, etc.) to execute computing tasks, rather than relying on traditional electronic computing technologies. The goal of the all-optical computing architecture is to directly complete data transmission, storage, and processing in the optical domain by reducing or eliminating the optoelectronic conversion process, thereby improving computational efficiency, reducing power consumption, and reducing latency.

[0020] The following will mainly introduce several key optical devices involved in the embodiments of the present application.

[0021] Microring Resonator (MRR): It is usually composed of one or more ring waveguides coupled with straight waveguides. The ring waveguide is the core part, generally having a size of micrometers or even smaller. The material of the waveguide can be optical materials such as silicon, silicon nitride, and silicon dioxide. The straight waveguide is used to input and output optical signals and is connected to the ring waveguide through a specific coupling region. The design of the coupling region is crucial for the transmission efficiency of optical signals between the straight waveguide and the ring waveguide. In some complex microring resonator structures, there may also be multiple ring waveguides coupled to each other or integrated with other optical components such as gratings and modulators to achieve more complex optical functions. MRR can usually be applied in various fields. For example, in the field of optical communication, MRR can be used as a wavelength division multiplexing system and an optical switch. For another example, in the field of optical computing, MRR can be used for logical operations and optical storage.

[0022] Mach-Zehnder Interferometer (MZI): It is mainly composed of two beam splitters (BS) and two mirrors. The beam splitter is generally a 50:50 beam splitter, that is, it can evenly divide the incident light into two beams of equal intensity. The beam splitter usually uses optical coating technology to form a specific thin film structure on the surface of glass or other optical materials to achieve the light splitting function. The role of the mirror is to change the propagation direction of light so that the two beams of light split by the beam splitter can propagate along different paths and then converge again. The mirror usually has a high reflectivity to ensure that the energy loss of light during reflection is small. The two optical paths formed by the two beam splitters and the two mirrors form the interference arms. The lengths of the two interference arms can be different. Usually, the length of one interference arm is fixed, and the length of the other interference arm can be adjusted in some way to introduce an optical path difference. According to the principle of light interference, when the two beams of light satisfy a certain phase relationship, constructive interference or destructive interference will occur. If the optical path difference between the two beams of light is an integer multiple of the light wavelength, the two beams of light will undergo constructive interference, and a strong optical signal will be obtained at the output end; if the optical path difference is a half-integer multiple of the light wavelength, destructive interference will occur, and the optical signal at the output end will be weakened or even zero. By adjusting the length of one of the interference arms or the medium characteristics, etc., the optical path difference between the two beams of light can be changed, thereby controlling the movement of the interference fringes and the intensity of the interference output.

[0023] Figure 1FIG. 0 shows a schematic block diagram of a non - linear processing method based on an all - optical computing architecture according to an embodiment of the present application. This non - linear processing method is implemented based on an all - optical computing architecture, that is, each step in this non - linear processing method 100 is realized by a computer controlling the input of optical signals into various optical devices. For example, this non - linear processing method is realized by a processor in a computer controlling the input of optical signals into an MRR and an MZI. Optionally, as Figure 1 shown, this non - linear processing method 100 includes the following parts or all of the content.

[0024] S110, input an optical signal matrix block X into the micro - ring resonator MRR, and output a first optical signal, where the first optical signal is the maximum optical signal in the optical signal matrix block X.

[0025] S120, input the optical signal matrix block X and the first optical signal into a cascaded Mach - Zehnder interferometer MZI and a series of multiple micro - ring resonators MRR in sequence, and obtain a second optical signal corresponding to any optical signal in the optical signal matrix block X. The second optical signal is the optical signal corresponding to the exponential power t with base 2, where t=(x i -x max )·log 2 e,x i is the value of the any optical signal, and x max is the value of the maximum optical signal.

[0026] S130, input the second optical signal into a low - wavelength channel and a high - wavelength channel in sequence, and output a third optical signal corresponding to the any optical signal. The third optical signal is the optical signal corresponding to the local exponential value 2 t of the any optical signal, where, the low - wavelength channel includes a Mach - Zehnder interferometer MZI and a micro - ring resonator MRR, and is used to output the optical signal corresponding to 2 r , r is the fractional part of t, and the high - wavelength channel includes q cascaded Mach - Zehnder interferometers MZI, and is used to output the optical signal corresponding to 2 q , q is the integer part of t.

[0027] S140, input the third optical signal corresponding to each optical signal in the optical signal matrix block X into a series of multiple micro - ring resonators MRR, and output a fourth optical signal. The fourth optical signal is the sum of the third signals corresponding to all optical signals in the optical signal matrix block X, and the fourth optical signal is the optical signal corresponding to the local exponential value of the optical signal matrix block X.

[0028] S150, modulate the third signal corresponding to any of the optical signals into an optical signal corresponding to a first wavelength and modulate the fourth optical signal into an optical signal corresponding to a second wavelength, the optical signal corresponding to the first wavelength being the corresponding optical signal, and the optical signal corresponding to the second wavelength being the corresponding optical signal, k 1 and v 1 parameters of the first wavelength, k 2 and v 2 being parameters of the second wavelength, v 1 and v 2 being constants less than 1.

[0029] S160, based on the magnitudes of v 1 and v 2 , sequentially input the optical signal corresponding to the first wavelength and the optical signal corresponding to the second wavelength into a microring resonator MRR array and a cascaded Mach-Zehnder interferometer MZI, and output a fifth optical signal corresponding to any of the optical signals, the fifth optical signal being the optical signal corresponding to the ratio of N 1 and N 2 . Among them, if v 1 -v 2 is greater than or equal to 0, the fifth optical signal is the corresponding optical signal; if v 1 -v 2 is less than 0, the fifth signal is the corresponding optical signal.

[0030] It should be noted that Figure 1 the illustrated embodiment takes the calculation of the Softmax function for the original optical signal as an example, but those skilled in the art understand that the embodiments of the present application can also implement the calculation of other non-linear functions for the original optical signal. For example, the Sigmoid function or the Tanh function. It only needs to adjust Figure 1 the order of each step and the combination method of various optical devices in the illustrated embodiment. The embodiments of the present application do not limit this.

[0031] Next, taking the Softmax function, that is, as an example, the technical solution shown in Figure 1 will be described in detail. First, based on mathematical principles, the exponential function with base e in this Softmax function can be equivalently converted into an operation that can be achieved by optical devices through the following conversion.

[0032] First, use the base change formula to change the exponential with base e (the natural constant) to an exponential with base 2:

[0033]

[0034] Among them, x i is the value of any optical signal in the optical signal matrix block X, and x max is the value of the maximum optical signal in the optical signal matrix block X. t is the exponential power to the base 2 corresponding to any optical signal, that is, t = (x i - x max )·log 2 e.

[0035] Secondly, convert t to:

[0036] t = I·log 2 e ≈ I + (I >> 1) + (I >> 3) - (I >> 4) (2);

[0037] Among them, I = x i - x max , thus, after obtaining the values of each optical signal and the value of the maximum optical signal in the optical signal matrix block X, the exponential power t corresponding to each optical signal can be calculated through shift and addition / subtraction operations.

[0038] After obtaining the exponential power t, it is necessary to further calculate 2 t . Considering that t may be a non-integer, 2 t may not be fully realized through shift operations. At this time, t can be decomposed into an integer part r and a fractional part q. Therefore, 2 t can be converted into the following formula:

[0039] 2 t = 2 q+r = 2 q ·2 r = (q.r) 2 (3);

[0040] Among them, (q.r) 2 is the binary bit stream representation form of 2 q ·2 r .

[0041] According to the Taylor expansion, further numerical approximation is made for 2 r , and the following can be obtained:

[0042] 2 r ≈ 1 + (r >> 1) (4);

[0043] Therefore, 2 t can be converted into the following formula:

[0044] 2 t ≈ (1 + (r >> 1)) << q (5).

[0045] So far, through the derivation of the above formulas (1) to (5), the exponential function with base e can be converted into shift operations and simple addition and subtraction operations, and these shift operations and simple addition and subtraction operations can be completely completed by optical devices, specifically as follows:

[0046] In step S110, the optical signal matrix block X can be input into the rear micro-ring resonator MRR to screen out the maximum optical signal in the optical signal matrix block X. It should be noted that step S110 may not be an essential step, but can be selectively executed based on the actual situation. This screening of the maximum optical signal can be used to avoid numerical overflow by subtracting the value of the maximum optical signal in subsequent steps.

[0047] After obtaining the maximum optical signal, the calculation of the exponential power t needs to be performed. That is, in step S120, the maximum optical signal and the optical signal matrix block X can be sequentially input into multiple MRRs connected in cascade and in series by MZI. The MZI cascade is used to implement the shift operation in the above formula (2), and the multiple MRRs connected in series are used to implement the addition and subtraction operations in formula (2). Finally, the optical signal corresponding to the exponential power t converted to base 2 for each optical signal in the optical signal matrix block X is output. That is, each optical signal in the optical signal matrix block X corresponds to a second optical signal, and the second optical signal is the optical signal corresponding to the exponential power t in formula 2. Through this method, the calculation of the exponential power is simplified to addition and shift operations, greatly reducing the consumption of hardware resources and the computational complexity.

[0048] Furthermore, in step S130, first, each second optical signal corresponding to the optical signal output in step S120 is sequentially input into the low-wavelength channel and the high-wavelength channel by using wavelength demultiplexing. Among them, the low-wavelength channel processes the second optical signal through MZI phase modulation and adjustment of the resonance frequency of the MRR, and then obtains the optical signal corresponding to formula (4). That is, each optical signal in the optical signal matrix block X after passing through the low-wavelength channel can output an optical signal corresponding to 2 r correspondingly. Further, the optical signal passing through the low-wavelength channel can be input into the high-wavelength channel. The high-wavelength channel uses q MZIs connected in cascade to achieve fast left shift processing and can obtain the optical signal corresponding to formula (5). That is, each optical signal in the optical signal matrix block X after passing through the low-wavelength channel and the high-wavelength channel in sequence can output an optical signal corresponding to 2 t correspondingly. That is to say, the output of step S130 is the optical signal obtained after performing the exponential function operation of e on each optical signal in the optical signal matrix block X. In step S130, the integer part can be realized by cascading q MZIs to perform fast left shift calculation of 2 q Combined with the approximate processing of the decimal part, the computational complexity is greatly reduced, and the requirement for complex hardware arithmetic units in the traditional method is significantly reduced.

[0049] To complete the calculation of the softmax function, it is also necessary to further calculate the sum of the optical signals after performing the exponential function of e on all the optical signals in the optical signal matrix block X. That is, in step S140, the optical signals after performing the exponential function of e on each optical signal of the input optical signal matrix block X obtained in step S130 can be allocated to different wavelengths λ 1 、λ 2 、λ 3 、……、λ n 。 Through multiple series-connected MRRs of the fiber optic combiner, the resonant wavelength is set to λ i 。 When the input optical signal matches the resonance condition, the MRR couples the light into the ring cavity and accumulates it. Finally, the output optical signal is the sum of the optical signals obtained after performing the exponential function of e on all the optical signals in the optical signal matrix block X in step S130.

[0050] Next, combining the ideas of logarithm and antilogarithm, based on the resonant wavelength selectivity of the MRR and the optical intensity ratio calculation realized by the interference modulation of the MZI, the exponential division operation can be realized by using simple shift operations. Specifically, we modulate the third optical signal corresponding to each optical signal obtained in step S130 into an optical signal with the first wavelength λ 1 , and modulate the fourth optical signal corresponding to the optical signal matrix block X obtained in step S140 into an optical signal with the second wavelength λ 2 。 After inputting the two modulated optical signals into the MRR array and realizing the shift operation through MZI cascading, the optical signal output after any optical signal in the optical signal matrix block X passes through the softmax function can be obtained.

[0051] The specific derivation process is as follows:

[0052] First, the numerator term in the softmax function can be expressed as The denominator term can be expressed as where k 1 and v 1 are parameters of λ 1 , k 2 and v 2 are parameters of λ 2 , v 1 and v 2 are constants less than 1. That is, once λ 1 and λ 2 are known, k 1 , v 1 , k 2 and v 2 can be obtained. The value Q of the optical signal output after any optical signal in the optical signal matrix block X passes through the softmax function:

[0053]

[0054] Take the logarithm of Q:

[0055]

[0056] Combined with the Taylor expansion, we have:

[0057] log 2 Q≈(k 1 -k 2 -1)+(1+v 1 -v 2 ) (8);

[0058] Then perform the antilogarithm operation:

[0059]

[0060] Finally, based on the magnitude of v 1 -v 2 , Q can be simplified into the following two cases:

[0061]

[0062] First, v 1 -v 2 can be calculated through the interference effect of the MZI. Then, the resonant wavelength of the MRR will shift according to the exponential part of the input optical intensity. By detecting the shift amount of the resonant wavelength, the exponential power difference of k 1 -k 2 can be obtained. In addition, the addition and subtraction operations in formula (10) can be calculated using the MRR array. Finally, by cascading the corresponding number of MZIs according to the exponential term to shift the corresponding number of bits to the left, division can be achieved through simple shift operations of the binary bit stream, thereby efficiently updating the value of the non-linear function in the optical computing architecture and improving the overall performance of the architecture.

[0063] Generally, in some high-speed optical communication systems, to improve the speed and efficiency of signal processing, the optical signal is split into multiple parallel processing channels, each channel is responsible for processing a part of the signal, and finally the processed signals are combined. This can make full use of the parallel processing capabilities of multiple processing units to improve the overall processing speed of the system. The above steps S110 to S150 only describe the operation process of the softmax function for the optical signal matrix block X in one channel. In fact, each channel can refer to the above steps S110 to S150 to implement. To further optimize the effect after each optical signal performs the softmax function, the above x max can be replaced with the value x' of the maximum optical signal among all optical signalsmax , that is, the global maximum value. That is, as Figure 2 shown, the non-linear processing method 100 may further include the following steps:

[0064] S165, storing the value of the first optical signal corresponding to each optical signal matrix block X in the plurality of optical signal matrix blocks X in an optical buffer ring (OBR), and outputting a sixth signal, where the sixth signal is the maximum value x' of the values of the first optical signals corresponding to the plurality of optical signal matrix blocks X max The corresponding optical signal stores the values of the first optical signals corresponding to the plurality of optical signal matrix blocks X in an optical buffer ring (OBR), and outputs a sixth signal, where the sixth signal is the maximum value x' of the values of the first optical signals corresponding to the plurality of optical signal matrix blocks X max corresponding optical signal;

[0065] S170, sequentially inputting the sixth signal and the first signal corresponding to any one of the optical signal matrix blocks X into a cascaded Mach-Zehnder interferometer MZI and a plurality of cascaded micro-ring resonators MRR in series to obtain a seventh signal corresponding to any one of the optical signal matrix blocks X, where the seventh signal is an optical signal corresponding to an exponential power t' with base 2, where t' = x max - x' max ;

[0066] S175, inputting the seventh signal into the low-wavelength channel and the high-wavelength channel in sequence, and outputting an eighth optical signal corresponding to any one of the optical signal matrix blocks X, where the eighth optical signal is 2 t′ corresponding optical signal, where the low-wavelength channel is used to output 2 r′ corresponding optical signal, r' is the fractional part of t', and the high-wavelength channel includes q' cascaded Mach-Zehnder interferometers MZI for outputting 2 q′ corresponding optical signal, where q' is the integer part of t';

[0067] S180, modulating the fifth optical signal corresponding to any one optical signal in the optical signal matrix block X into an optical signal corresponding to a third wavelength and modulating the eighth optical signal corresponding to the optical signal matrix block X into an optical signal corresponding to a fourth wavelength, where the optical signal corresponding to the third wavelength is corresponding optical signal, and the optical signal corresponding to the fourth wavelength is corresponding optical signal, k 3 and v 3 are parameters of the third wavelength, k 4 and v4 is the parameter of the fourth wavelength, v 3 and v 4 are constants less than 1;

[0068] S185, based on v 3 and v 4 the optical signals corresponding to the third wavelength and the optical signals corresponding to the fourth wavelength are sequentially input into the micro-ring resonator MRR array and the cascaded Mach-Zehnder interferometer MZI according to their magnitudes, and the ninth optical signal corresponding to any optical signal in the optical signal matrix block X is output, and the ninth optical signal is the optical signal corresponding to the product of N 3 and N 4 and the ninth optical signal is the optical signal corresponding to the global exponential value of any optical signal in the optical signal matrix block X, where if v 3 + v 4 is less than 1, the ninth optical signal is the corresponding optical signal; if v 3 + v 4 is greater than or equal to 1, the ninth optical signal is the corresponding optical signal.

[0069] Specifically, in step S110, the value x max of the maximum optical signal in the optical signal matrix block X of each channel is stored in an optical storage device, for example, an optical buffer ring (OBR). In step S165, the OBR dynamically adjusts the resonant wavelength through thermo-optic tuning, thereby dynamically updating the current global maximum x' max , that is, the global maximum optical signal is selected from the maximum optical signals output from multiple channels, thereby supporting the synchronization and consistency of multi-module operations.

[0070] Convert each optical signal corresponding to into the following formula:

[0071]

[0072] When calculating , it is similar to calculating , that is, first convert it into an exponent with base 2:

[0073]

[0074] where x max is the value of the maximum optical signal in the optical signal matrix block X, x' max is the value of the maximum optical signal in multiple optical signal matrix blocks X, that is, the global maximum, and t' is the exponentiation with base 2 corresponding to any optical signal matrix block X, that is, t' = (x max - x′max )·log 2 e。

[0075] Secondly, convert t' to:

[0076] t = I'·log 2 e ≈ I'+(I' >> 1)+(I' >> 3)-(I' >> 4) (13);

[0077] where I' = x max - x' max , thus, after obtaining the value of the maximum optical signal in the optical signal matrix block X and the values of the maximum optical signals in multiple optical signal matrix blocks X, the exponential power t' corresponding to each optical signal can be calculated through shift and addition / subtraction operations.

[0078] After obtaining the exponential power t', it is necessary to further calculate 2 t′ , considering that t' may be a non-integer, 2 t′ may not be fully realized through shift operations. At this time, t' can be decomposed into an integer part r' and a fractional part q'. Therefore, t' can be converted into the following formula:

[0079] 2 t′ = 2 q′+r′ = 2 q′ ·2 r′ = (q'.r') 2 (14);

[0080] where (q'.r') 2 is the binary bit stream representation form of 2 q′ ·2 r′ .

[0081] According to the Taylor expansion, further numerical approximation of 2 r′ can be obtained:

[0082] 2 r′ ≈ 1+(r' >> 1) (15);

[0083] Therefore, 2 t can be converted into the following formula:

[0084] 2 t′ ≈ (1+(r' >> 1)) << q' (16).

[0085] So far, through the derivation of the above formulas (12) to (16), the exponential function with base e can be converted into shift operations and simple addition / subtraction operations, and these shift operations and simple addition / subtraction operations can be completely completed by optical devices, specifically as follows:

[0086] In step S170, the maximum optical signal in any optical signal matrix block X and the maximum optical signals in multiple optical signal matrix blocks X can be sequentially input into a cascaded MZI and multiple cascaded MRRs respectively. The cascaded MZI is used to implement the shift operation in the above formula (13), and the multiple cascaded MRRs are used to implement the addition and subtraction operations in formula (13). Finally, the optical signal corresponding to the exponentiation to the base 2, i.e., t', is output for each optical signal matrix block X. That is, each optical signal matrix block X corresponds to a seventh optical signal, and the seventh optical signal is the optical signal corresponding to the exponent t' in formula (13). Through this method, the calculation of the exponent is simplified to addition and shift operations, greatly reducing the hardware resource consumption and operation complexity.

[0087] Further, in step S175, first, each seventh optical signal corresponding to the optical signal matrix block X output in step S170 is sequentially input into a low-wavelength channel and a high-wavelength channel by using wavelength demultiplexing. Among them, the low-wavelength channel processes the seventh optical signal by adjusting the MZI phase modulation and the resonance frequency of the MRR, and then obtains the optical signal corresponding to formula (15), that is, after passing through the low-wavelength channel, an optical signal corresponding to 2 r′ is output for any optical signal matrix block X. Further, the optical signal passing through the low-wavelength channel can be input into the high-wavelength channel, and the high-wavelength channel uses q' cascaded MZIs to achieve fast left-shift processing and can obtain the optical signal corresponding to formula (16). That is, after passing through the low-wavelength channel and the high-wavelength channel in sequence, each optical signal matrix block X can output an optical signal corresponding to 2 t′ . That is to say, the output of step S175 is the optical signal obtained after performing the exponential function operation of e on the maximum optical signal in each optical signal matrix block X. In step S175, the integer part can be realized by cascading q' MZIs for fast left-shift calculation of 2 q′ , combined with the approximate processing of the fractional part, greatly reducing the calculation complexity and significantly reducing the requirement for complex hardware operation units in the traditional method.

[0088] Next, by combining the ideas of logarithm and antilogarithm, and implementing optical intensity multiplication calculation based on the resonance wavelength selectivity of the MRR and the interference modulation of the MZI, the exponential multiplication operation can be realized by using simple shift operations. Specifically, we modulate the third optical signal corresponding to any optical signal in each optical signal matrix block X obtained in step S140 into an optical signal with a third wavelength λ 3 , and modulate the eighth optical signal corresponding to each optical signal matrix block X obtained in step S175 into an optical signal with a fourth wavelength λ 4The optical signals. After modulating the two optical signals, input them into the MRR array, and then realize the shift operation through MZI cascading, the ninth optical signal corresponding to any optical signal in the optical signal matrix block X can be obtained, that is The corresponding optical signal.

[0089] The specific derivation process is as follows:

[0090] First, represent the optical signal of the third wavelength as The optical signal of the fourth wavelength is represented as Where k 3 And v 3 Are the parameters of λ 3 K 4 And v 4 Are the parameters of λ 4 V 3 And v 4 Are constants less than 1, that is, once λ 3 And λ 4 Are known, k 3 And v 3 K 4 And v 4 Can be obtained. Then the value M of the ninth optical signal corresponding to any optical signal in the optical signal matrix block X:

[0091]

[0092] Similarly, taking the logarithm of M, Taylor expansion approximation, and antilogarithm can obtain:

[0093]

[0094] Then calculate v 3 + v 4 Through the interference effect of MZI. According to the situation of v 3 + v 4 It can be divided into two cases:

[0095]

[0096] The principles and calculation processes of exponential multiplication operation and exponential division operation are similar. For example, v 3 + v 4 Can be calculated through the interference effect of MZI. Then, the resonant wavelength of the MRR will shift according to the exponential part of the input optical intensity. By detecting the shift amount of the resonant wavelength, k 3 + k 4The exponential power difference. In addition, the addition and subtraction operations in formula (19) can be calculated using the MRR array. Finally, according to the exponential term, cascade the corresponding number of MZIs to shift the corresponding number of bits to the left. In this way, division can be achieved through simple shifting operations of the binary bit stream, thereby efficiently updating the value of the non-linear function in the optical computing architecture and improving the overall performance of the architecture.

[0097] After updating to it is also necessary to update to

[0098] where sum' is the accumulated sum of the global exponential values of each optical signal matrix block X, and sum is the accumulated sum of the local exponential values of each optical signal matrix block X.

[0099] Since the operation of sum has been implemented in the above step S140, after calculating the accumulated sum of the local exponential values in the above optical signal matrix block X can be multiplied by it to obtain the accumulated sum of the global exponential values of each optical signal matrix block X.

[0100] Furthermore, after obtaining the optical signals corresponding to and the optical signals corresponding to step S150 and S160 can be executed again. In this way, after considering the global maximum optical signal, the optical signals after performing the softmax function operation on each optical signal in the optical signal matrix block X can be output, thereby enhancing the stability of numerical calculations and avoiding numerical overflow problems in exponential operations caused by too large input values.

[0101] In some embodiments, after obtaining the exponential power t corresponding to each optical signal in the optical signal matrix block X, storing it in the optical storage device can bypass repeated exponential operations, thereby greatly reducing the computational complexity and consumption of hardware resources.

[0102] In other embodiments, after obtaining the local exponential values of each optical signal in the optical signal matrix block X, they can also be stored in the optical storage device, and after obtaining the global exponential values of each optical signal in the optical signal matrix block X, the local exponential values stored in the optical storage device can be replaced with the global exponential values.

[0103] In some other embodiments, after obtaining the local softmax function values of each optical signal matrix block X, they can also be stored in an optical storage device. After obtaining the global softmax function values of each optical signal matrix block X, the local softmax function values stored in the optical storage device can be replaced with the global softmax function values.

[0104] Figure 3 FIG. shows a schematic block diagram of a non-linear processing device based on an all-optical computing architecture provided by an embodiment of the present application. As Figure 3 shown, the non-linear processing device 200 includes the following parts or all of the content.

[0105] The maximum optical selection module 210 is configured to input the optical signal matrix block X into the microring resonator MRR and output a first optical signal, where the first optical signal is the maximum optical signal in the optical signal matrix block X.

[0106] The optical phase shift module 220 is configured to sequentially input the optical signal matrix block X and the first optical signal into a cascaded Mach-Zehnder interferometer MZI and a series of multiple microring resonators MRR to obtain a second optical signal corresponding to any optical signal in the optical signal matrix block X, where the second optical signal is the optical signal corresponding to the exponential power t with base 2, where t = (x i -x max )·log 2 e, x i is the value of the any optical signal, and x max is the value of the maximum optical signal.

[0107] The optical approximate exponential module 230 is configured to sequentially input the second optical signal into a low-wavelength channel and a high-wavelength channel and output a third optical signal corresponding to the any optical signal, where the third optical signal is the optical signal corresponding to the local exponential value 2 t corresponding to the any optical signal, where the low-wavelength channel includes a Mach-Zehnder interferometer MZI and a microring resonator MRR for outputting the optical signal corresponding to 2 r corresponding to the any optical signal, r is the fractional part of t, and the high-wavelength channel includes q cascaded Mach-Zehnder interferometers MZI for outputting the optical signal corresponding to 2 q corresponding to the any optical signal, and q is the integer part of t.

[0108] An optical wavelength division multiplexing and accumulation module 240 is configured to input the third optical signals corresponding to each optical signal in the optical signal matrix block X into a plurality of cascaded micro-ring resonators MRR, and output a fourth optical signal, where the fourth optical signal is the sum of the third signals corresponding to all optical signals in the optical signal matrix block X, and the fourth optical signal is the optical signal corresponding to the local exponential value of the optical signal matrix block X.

[0109] An optical division and shift module 250 is configured to modulate the third signal corresponding to any one of the optical signals into an optical signal corresponding to a first wavelength and modulate the fourth optical signal into an optical signal corresponding to a second wavelength. The optical signal corresponding to the first wavelength is the corresponding optical signal, and the optical signal corresponding to the second wavelength is the corresponding optical signal, k 1 and v 1 are parameters of the first wavelength, k 2 and v 2 are parameters of the second wavelength, v 1 and v 2 are constants less than 1; and based on the magnitudes of v 1 and v 2 , the optical signal corresponding to the first wavelength and the optical signal corresponding to the second wavelength are sequentially input into a micro-ring resonator MRR array and a cascaded Mach-Zehnder interferometer MZI, and a fifth optical signal corresponding to any one of the optical signals is output. The fifth optical signal is the optical signal corresponding to the ratio of N 1 and N 2 . Wherein, if v 1 -v 2 is greater than or equal to 0, the fifth optical signal is the corresponding optical signal; if v 1 -v 2 is less than 0, the fifth signal is the corresponding optical signal.

[0110] In a possible embodiment, the maximum optical selection module 210 is further configured to store the values of the first optical signals corresponding to each optical signal matrix block X in a plurality of the optical signal matrix blocks X in an optical buffer ring OBR, and output a sixth signal, where the sixth signal is the optical signal corresponding to the maximum value x' max among the values of the first optical signals corresponding to the plurality of optical signal matrix blocks X.

[0111] The optical phase shift module 220 is further configured to: sequentially input the sixth signal and the first signal corresponding to any one of the optical signal matrix blocks X into a cascaded Mach-Zehnder interferometer MZI and a plurality of serially connected micro-ring resonators MRRs, to obtain a seventh signal corresponding to any one of the optical signal matrix blocks X, where the seventh signal is an optical signal corresponding to the exponential power t' with base 2, and t' = x max -x' max .

[0112] The optical approximate exponential module 230 is further configured to: input the seventh signal into the low-wavelength channel and the high-wavelength channel in sequence, and output an eighth optical signal corresponding to any one of the optical signal matrix blocks X, where the eighth optical signal is an optical signal corresponding to 2 t′ . Among them, The low-wavelength channel is configured to output an optical signal corresponding to 2 r′ , where r' is the fractional part of t', and the high-wavelength channel includes q' cascaded Mach-Zehnder interferometers MZIs, and is configured to output an optical signal corresponding to 2 q′ , where q' is the integer part of t'.

[0113] The non-linear processing device 200 further includes: an optical multiplication shift module 260, configured to modulate the fifth optical signal corresponding to any one of the optical signals in the optical signal matrix block X into an optical signal corresponding to a third wavelength, and modulate the eighth optical signal corresponding to the optical signal matrix block X into an optical signal corresponding to a fourth wavelength. The optical signal corresponding to the third wavelength is . The optical signal corresponding to the fourth wavelength is . k 3 and v 3 are parameters of the third wavelength, k 4 and v 4 are parameters of the fourth wavelength, and v 3 and v 4 are constants less than 1; and based on the magnitudes of v 3 and v 4 , sequentially input the optical signal corresponding to the third wavelength and the optical signal corresponding to the fourth wavelength into a micro-ring resonator MRR array and a cascaded Mach-Zehnder interferometer MZI, and output a ninth optical signal corresponding to any one of the optical signals in the optical signal matrix block X. The ninth optical signal is an optical signal corresponding to the product of N 3 and N 4 , and the ninth optical signal is an optical signal corresponding to the global exponential value of any one of the optical signals in the optical signal matrix block X. Among them, if v 3 +v 4 is less than 1, the ninth optical signal is . If v3 +v 4 Greater than or equal to 1, the ninth optical signal is The corresponding optical signal.

[0114] Based on the same concept, an embodiment of the present application also provides another non-linear processing device based on an all-optical computing architecture. The non-linear processing device includes a memory for storing computer program instructions and a processor for executing the program instructions. When the program instructions are executed by the processor, the non-linear processing device is caused to execute the above various method embodiments.

[0115] It should be noted that the detailed content of the device-side embodiments can be referred to the method-side embodiments. For the sake of brevity, it will not be repeated here.

[0116] Based on the same concept, an embodiment of the present application also provides a computer-readable storage medium. The computer-readable storage medium includes a stored program. When the program runs, it controls the device where the computer-readable storage medium is located to execute the above various method embodiments.

[0117] Although the present application has been described with reference to the preferred embodiments, various improvements can be made to it and components therein can be replaced with equivalents without departing from the scope of the present application. In particular, as long as there is no structural conflict, the various technical features mentioned in each embodiment can be combined in any way. The present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A nonlinear processing method based on an all-optical computing architecture, characterized in that: The all-optical computing architecture includes a microring resonator (MRR) and a Mach-Zehnder interferometer (MZI), and the nonlinear processing method includes: Inputting the optical signal matrix block X into the microring resonator MRR, and outputting a first optical signal, wherein the first optical signal is the maximum optical signal in the optical signal matrix block X; The optical signal matrix block X and the first optical signal are sequentially input into a cascaded Mach-Zehnder interferometer MZI and a plurality of microring resonators MRR connected in series, and a second optical signal corresponding to any optical signal in the optical signal matrix block X is obtained, wherein the second optical signal is an optical signal corresponding to an exponential power t with a base of 2, wherein t=(x i -x max )·log2e,x i is the value of any optical signal, x max is the value of the maximum optical signal; The second optical signal is sequentially input into the low wavelength channel and the high wavelength channel, and a third optical signal corresponding to any of the optical signals is output, wherein the third optical signal is the local index value 2 of any of the optical signals. t The corresponding optical signal, where The low wavelength channel includes a Mach-Zehnder interferometer MZI and a microring resonator MRR, which are used to output 2 r The corresponding optical signal, r is the fractional part of t, and the high wavelength channel includes q cascaded Mach-Zehnder interferometers MZI, which are used to output 2 q The corresponding optical signal, q is the integer part of t; Input the third optical signal corresponding to each optical signal in the optical signal matrix block X into a plurality of microring resonators MRR connected in series, and output a fourth optical signal, wherein the fourth optical signal is the sum of the third signals corresponding to all optical signals in the optical signal matrix block X, and the fourth optical signal is an optical signal corresponding to a local index value of the optical signal matrix block X; The third signal corresponding to any one of the optical signals is modulated into an optical signal corresponding to a first wavelength, and the fourth optical signal is modulated into an optical signal corresponding to a second wavelength, wherein the optical signal corresponding to the first wavelength is The optical signal corresponding to (1+v1), the optical signal corresponding to the second wavelength is corresponding optical signal, k1 and v1 are parameters of the first wavelength, k2 and v2 are parameters of the second wavelength, and v1 and v2 are constants less than 1; Based on the size of v1 and v2, the optical signal corresponding to the first wavelength and the optical signal corresponding to the second wavelength are sequentially input into the microring resonator MRR array and the cascaded Mach-Zehnder interferometer MZI, and a fifth optical signal corresponding to any of the optical signals is output, wherein the fifth optical signal is an optical signal corresponding to the ratio of N1 to N2, wherein if v1-v2 is greater than or equal to 0, the fifth optical signal is The corresponding light signal; if v1-v2 is less than 0, the fifth signal is The corresponding optical signal.

2. The nonlinear processing method according to claim 1, characterized in that: The nonlinear processing method further comprises: The value of the first optical signal corresponding to each optical signal matrix block X in the plurality of optical signal matrix blocks X is stored in the optical buffer ring OBR, and a sixth signal is output, wherein the sixth signal is the maximum value x' among the values ​​of the first optical signal corresponding to the plurality of optical signal matrix blocks X max The corresponding optical signal; The sixth signal and the first signal corresponding to any of the optical signal matrix blocks X are sequentially input into the cascaded Mach-Zehnder interferometer MZI and the plurality of microring resonators MRR connected in series to obtain a seventh signal corresponding to any of the optical signal matrix blocks X, wherein the seventh signal is an optical signal corresponding to the exponential power t' with a base of 2, wherein t'=x max -x′ max ; The seventh signal is input into the low wavelength channel and the high wavelength channel in sequence, and an eighth optical signal corresponding to any optical signal matrix block X is output, and the eighth optical signal is 2 t' The corresponding optical signal, where The low wavelength channel is used to output 2 r' The corresponding optical signal, r' is the fractional part of t', and the high wavelength channel includes q' cascaded Mach-Zehnder interferometers MZI, which are used to output 2 q′ The corresponding optical signal, q' is the integer part of t'; The fifth optical signal corresponding to any optical signal in the optical signal matrix block X is modulated into an optical signal corresponding to a third wavelength, and the eighth optical signal corresponding to the optical signal matrix block X is modulated into an optical signal corresponding to a fourth wavelength. The optical signal corresponding to the third wavelength is The optical signal corresponding to the fourth wavelength is corresponding optical signal, k3 and v3 are parameters of the third wavelength, k4 and v4 are parameters of the fourth wavelength, and v3 and v4 are constants less than 1; Based on the size of v3 and v4, the optical signal corresponding to the third wavelength and the optical signal corresponding to the fourth wavelength are sequentially input into the microring resonator MRR array and the cascaded Mach-Zehnder interferometer MZI, and a ninth optical signal corresponding to any optical signal in the optical signal matrix block X is output, wherein the ninth optical signal is an optical signal corresponding to the product of N3 and N4, and the ninth optical signal is an optical signal corresponding to the global index value of any optical signal in the optical signal matrix block X, wherein if v3+v4 is less than 1, the ninth optical signal is Corresponding optical signal; if v3+v4 is greater than or equal to 1, the ninth optical signal is The corresponding optical signal.

3. A nonlinear processing device based on an all-optical computing architecture, characterized in that: The all-optical computing architecture includes a microring resonator MRR and a Mach-Zehnder interferometer MZI, and the nonlinear processing device includes: A maximum optical selection module, used to input the optical signal matrix block X into the microring resonator MRR, and output a first optical signal, where the first optical signal is the maximum optical signal in the optical signal matrix block X; An optical phase shift module is used to sequentially input the optical signal matrix block X and the first optical signal into a cascaded Mach-Zehnder interferometer MZI and a plurality of microring resonators MRR connected in series, to obtain a second optical signal corresponding to any optical signal in the optical signal matrix block X, wherein the second optical signal is an optical signal corresponding to an exponential power t with a base of 2, wherein t=(x i -x max )·log2e,x i is the value of any optical signal, x max is the value of the maximum optical signal; An optical approximate index module, used to input the second optical signal into the low wavelength channel and the high wavelength channel in sequence, and output a third optical signal corresponding to any of the optical signals, wherein the third optical signal is a local index value 2 of any of the optical signals t The corresponding optical signal, where The low wavelength channel includes a Mach-Zehnder interferometer MZI and a microring resonator MRR, which are used to output 2 r The corresponding optical signal, r is the fractional part of t, and the high wavelength channel includes q cascaded Mach-Zehnder interferometers MZI, which are used to output 2 q The corresponding optical signal, q is the integer part of t; An optical wavelength division multiplexing accumulation module, used to input the third optical signal corresponding to each optical signal in the optical signal matrix block X into a plurality of microring resonators MRR connected in series, and output a fourth optical signal, wherein the fourth optical signal is the sum of the third signals corresponding to all optical signals in the optical signal matrix block X, and the fourth optical signal is an optical signal corresponding to a local index value of the optical signal matrix block X; An optical division shift module is used to modulate the third signal corresponding to any one of the optical signals into an optical signal corresponding to a first wavelength and modulate the fourth optical signal into an optical signal corresponding to a second wavelength, wherein the optical signal corresponding to the first wavelength is The optical signal corresponding to the second wavelength is The optical signal corresponding to the first wavelength is k1 and v1, k2 and v2 are parameters of the second wavelength, and v1 and v2 are constants less than 1; and based on the size of v1 and v2, the optical signal corresponding to the first wavelength and the optical signal corresponding to the second wavelength are sequentially input into the microring resonator MRR array and the cascaded Mach-Zehnder interferometer MZI, and a fifth optical signal corresponding to any of the optical signals is output, and the fifth optical signal is an optical signal corresponding to the ratio of N1 to N2, wherein if v1-v2 is greater than or equal to 0, the fifth optical signal is The corresponding light signal; if v1-v2 is less than 0, the fifth signal is The corresponding optical signal.

4. The nonlinear processing device according to claim 3, characterized in that: The maximum light selection module is also used for: The value of the first optical signal corresponding to each optical signal matrix block X in the plurality of optical signal matrix blocks X is stored in the optical buffer ring OBR, and a sixth signal is output, wherein the sixth signal is the maximum value x' among the values ​​of the first optical signal corresponding to the plurality of optical signal matrix blocks X max The corresponding optical signal; The optical phase shift module is also used for: The sixth signal and the first signal corresponding to any of the optical signal matrix blocks X are sequentially input into the cascaded Mach-Zehnder interferometer MZI and the plurality of microring resonators MRR connected in series to obtain a seventh signal corresponding to any of the optical signal matrix blocks X, wherein the seventh signal is an optical signal corresponding to the exponential power t' with a base of 2, wherein t'=x max -x′ max ; The light proximity index module is also used to: The seventh signal is input into the low wavelength channel and the high wavelength channel in sequence, and an eighth optical signal corresponding to any optical signal matrix block X is output, and the eighth optical signal is 2 t' The corresponding optical signal, where The low wavelength channel is used to output 2 r' The corresponding optical signal, r' is the fractional part of t', and the high wavelength channel includes q' cascaded Mach-Zehnder interferometers MZI, which are used to output 2 q′ The corresponding optical signal, q' is the integer part of t'; The nonlinear processing device further comprises: Optical multiply-shift module for: The fifth optical signal corresponding to any optical signal in the optical signal matrix block X is modulated into an optical signal corresponding to a third wavelength, and the eighth optical signal corresponding to the optical signal matrix block X is modulated into an optical signal corresponding to a fourth wavelength. The optical signal corresponding to the third wavelength is The optical signal corresponding to the fourth wavelength is corresponding optical signal, k3 and v3 are parameters of the third wavelength, k4 and v4 are parameters of the fourth wavelength, and v3 and v4 are constants less than 1; and Based on the size of v3 and v4, the optical signal corresponding to the third wavelength and the optical signal corresponding to the fourth wavelength are sequentially input into the microring resonator MRR array and the cascaded Mach-Zehnder interferometer MZI, and a ninth optical signal corresponding to any optical signal in the optical signal matrix block X is output, wherein the ninth optical signal is an optical signal corresponding to the product of N3 and N4, and the ninth optical signal is an optical signal corresponding to the global index value of any optical signal in the optical signal matrix block X, wherein if v3+v4 is less than 1, the ninth optical signal is Corresponding optical signal; if v3+v4 is greater than or equal to 1, the ninth optical signal is The corresponding optical signal.

5. A nonlinear processing device based on an all-optical computing architecture, characterized in that: include: A processor, and a memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the nonlinear processing method based on the all-optical computing architecture as claimed in claim 1 or as described.

Images