Baseband processor based on general calculation co-fusion and control method

Abstract

The invention relates to the communication field and the AI field, in particular to a general calculation co-fusion-based baseband processor and a control method, and the general calculation co-fusion-based baseband processor comprises a signal input module, a signal preprocessing module, a dynamic precision adjustable PE module, a signal decoding module, a neural network preprocessing module and an AI result output module. MIMO equalization and AI calculation share a shared PE unit, an optimal scheduling algorithm is provided to realize dynamic allocation of the PE unit, and the problems that a traditional baseband processor cannot realize resource sharing of communication optimization and neural network calculation and cannot dynamically allocate resources according to communication requirements are solved.

Description

Technical Field

[0001] The present invention relates to the fields of communications and AI, and in particular to a baseband processor based on universal computing integration. Background Art

[0002] With the continuous evolution of communications, 6G will go beyond the interconnection of people and things to the intelligent interconnection of all things. To achieve this, the 6G end-to-end communication system needs to support AI (Artificial Intelligence) and ML (Machine Learning) computing well to achieve optimal efficiency. Therefore, to truly support inclusive intelligence, it is necessary to provide high-performance connections, high-capacity computing and rich storage resources at the edge of the network.

[0003] 6G needs to provide high-performance wireless connections that are comparable to optical fiber in speed, with Tbps-level peak rates, 10-100Gbps experience rates, sub-millisecond latency, and 10 times the connection density of 5G. As a result, the baseband signal processing of communications is required to achieve extremely high processing rates, extremely high parallelism, great computing power, and extremely high energy efficiency. One of the current challenges is that integrated circuits are facing performance bottlenecks due to the advent of the post-Moore era. Therefore, it will be difficult to design and implement future 6G baseband algorithms simply through traditional resource stacking.

[0004] With the development of 6G and AI, the boundaries between communications and computing will gradually blur in the future. Communication networks and devices will support AI and ML computing, and baseband processors will increasingly integrate AI / ML algorithms. These algorithms have put forward new requirements on the computing power of baseband processors.

[0005] The computing resources used for communication and AI in traditional baseband processors are completely separate and independent, with limited computing power, unable to complete intensive computing tasks in AI, and low resource utilization. If computing resources can be shared for communication and AI computing, computing power can be greatly improved, and hardware resource utilization can be improved.

[0006] However, there are huge challenges in realizing the sharing of computing resources for communication and AI computing in the baseband. On the one hand, traditional baseband processors are not flexible enough, and the hardware is relatively fixed, which cannot adapt to flexible baseband processing algorithms for different communication qualities. Secondly, realizing the sharing of communication and AI computing resources in the baseband requires that communication and AI computing have a common computing paradigm, but the similarity between communication baseband algorithms and AI algorithm calculations is limited.

[0007] Finally, traditional baseband processors do not have a computing resource allocation mechanism and are unable to schedule computing resources according to actual communication conditions and hardware constraints. Summary of the invention

[0008] The purpose of the present invention is to provide a baseband processor based on universal computing integration, realize dynamic allocation of PE units (processing elements), and realize computing resource sharing of communication computing and neural network computing, thereby solving the problem that traditional baseband processors cannot realize dynamic allocation of computing resources according to communication needs, and cannot realize communication optimization and computing resource sharing of neural network computing.

[0009] In order to achieve the purpose of the above invention, the present invention provides the following technical solutions:

[0010] The baseband processor based on universal computing includes:

[0011] A signal input module, used for receiving communication signals;

[0012] A signal preprocessing module performs signal preprocessing on the communication signal received by the signal input module, and then transmits the preprocessed communication signal to the dynamic precision adjustable PE module;

[0013] Computing resource scheduling module, used to schedule the dynamically precision adjustable PE module;

[0014] Dynamic precision adjustable PE module for MIMO (Multiple Input Multiple Output) equalization and / or AI calculation;

[0015] The dynamic precision adjustable PE module comprises: a control unit, a storage unit and a calculation unit, wherein the calculation unit comprises a plurality of PE units; the control unit controls the calculation unit to read the storage unit data for calculation processing;

[0016] A signal decoding module, used for receiving the communication signal after MIMO equalization by the dynamic precision adjustable PE module, and performing decoding operation;

[0017] A neural network preprocessing module is used to preprocess the information content obtained by the decoding operation to obtain a signal for neural network calculation; then, the signal is input into the dynamic precision adjustable PE module to perform AI calculation and obtain an AI calculation result;

[0018] The AI ​​result output module is used to output the AI ​​calculation results.

[0019] The baseband processor of the present invention integrates common computing modes, realizes the shared allocation of PE units required for MIMO equalization computing and AI computing, realizes the sharing of the computing power of the baseband processor, and enables the dynamic allocation of computing resources for MIMO equalization and resources for AI computing, reduces the computing resource demand restrictions, and improves the utilization of computing resources. The final computing efficiency is optimized, the hardware overhead is minimized, and the design pressure of the baseband processor is greatly reduced.

[0020] Furthermore, the neural network preprocessing module preprocesses the information content of the completed decoding operation, that is, normalizes the data. Preferably, the normalized data is then transmitted to the dynamic precision adjustable PE module.

[0021] Further, the MIMO equalization is: the dynamic precision adjustable PE module uses a linear minimum mean square error algorithm to equalize the communication signal after the signal preprocessing;

[0022] The AI ​​calculation is: performing neural network calculation on the information content of the data normalization process to obtain a digital sequence.

[0023] Furthermore, the computing resource scheduling module can control the dynamic precision adjustable PE module to allocate PE unit resources to complete the calculation according to the communication quality, hardware constraints and task requirements.

[0024] Furthermore, the PE unit includes at least a complex multiplier and an arithmetic logic unit; when the PE unit performs MIMO equalization and AI calculation, the calculation accuracy is adjustable.

[0025] Furthermore, the AI ​​result output module receives the AI ​​calculation result, restores the AI ​​calculation result into corresponding information content, and displays it.

[0026] Furthermore, the computing resource scheduling module can control the number of PE units allocated by the dynamic precision adjustable PE module to MIMO equalization and the number of PE units allocated to AI calculation.

[0027] Furthermore, the dynamic precision adjustable PE module dynamically allocates PE units for calculation according to the control signal given by the computing resource scheduling module.

[0028] The present invention also provides a method for performing common computing and integration control using the above-mentioned baseband processor, which includes two stages, namely a communication computing stage and an AI computing stage;

[0029] The first stage, the communication and computing stage:

[0030] receiving communication signal input;

[0031] Preprocess the input signal;

[0032] The pre-processed signal is sent to the dynamic precision adjustable PE module for MIMO equalization;

[0033] Decoding the communication signal after MIMO equalization;

[0034] The second stage, AI computing stage:

[0035] Send the decoded signal to the neural network preprocessing module for preprocessing;

[0036] The signal after neural network preprocessing is sent to the dynamic precision adjustable PE module for AI calculation;

[0037] Output the results of AI calculation.

[0038] Furthermore, the channel capacity of the MIMO communication system is calculated, the time required for MIMO equalization is calculated according to the channel capacity, and the number of PE units allocated to MIMO equalization is determined.

[0039] Further, the number of PE units allocated to AI calculation is determined according to the number of PE units allocated to MIMO equalization.

[0040] Preferably, the calculation formula of the channel capacity is as follows:

[0041]

[0042] Where C is the channel capacity; k1 and ρ are coefficient factors; N r is the number of columns of the channel matrix (i.e. the number of receiving antennas); is the quantization error matrix H q The average singular value of is the average eigenvalue of the covariance matrix Q; Q d is the quantization bit width of the calculation element; R R Normalize the receiving Hermit matrix to the unit diagonal element; is the variance of additive white Gaussian noise.

[0043] Preferably, the calculation formula for the time required for MIMO equalization is as follows:

[0044]

[0045] Among them, T cycle is the time required for MIMO equalization; D is the amount of data input for communication; C is the channel capacity.

[0046] Preferably, the calculation formula for the number of PE units allocated to MIMO equalization is as follows:

[0047]

[0048] Among them, P com is the number of PE units allocated to MIMO equalization; ε is the resources required for MIMO equalization; N t is the number of rows of the channel matrix, i.e. the number of transmitting antennas; T cycle is the time required for MIMO equalization; ξ(Q) is the resource overhead consumed by a single PE unit when performing MIMO equalization when the actual bit width of the calculation element is Q;

[0049] The calculation formula for the number of PE units allocated to AI calculation is as follows:

[0050] P nnc =PP com

[0051] Among them, P nnc is the number of PE units allocated to AI calculation, P is the total number of PE units; P com is the number of PE units allocated to MIMO equalization.

[0052] The universal computing integration of the present invention refers to the universal allocation of comprehensive computing resources to achieve the integration of signal processing and AI computing. The two share the same PE array and computing resources, reducing the development pressure of the baseband processor.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention can use the same PE unit for MIMO equalization and AI calculation, which solves the problem that communication optimization and neural network calculation of traditional baseband processors cannot share resources, and solves the problem that traditional baseband processors cannot dynamically allocate computing resources according to communication needs.

[0055] The present invention is based on a baseband processor with universal computing integration and adopts a dynamic PE resource scheduling algorithm to allocate computing resources according to constraints such as communication quality, task requirements, and hardware requirements, so that the final computing efficiency is optimized and the hardware overhead is minimized, greatly reducing the design pressure of the baseband processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a schematic diagram of the framework structure of a baseband processor based on universal computing integration;

[0057] Figure 2 It is a schematic diagram of the structure of a dynamically precision adjustable PE module of a baseband processor based on universal computing integration.

[0058] Figure 3 This is the working timing diagram of the baseband processor based on universal computing integration. DETAILED DESCRIPTION

[0059] A baseband processor with universal computing and integration mainly includes the following modules:

[0060] Signal input module: used to receive communication signal input.

[0061] Signal preprocessing module: used to preprocess the input signal; preferably, the signal preprocessing includes one or more operations such as precoding, modulation, serial-to-parallel conversion, and parallel-to-serial conversion.

[0062] Computing resource scheduling module: used to schedule the dynamically adjustable precision PE module. This module receives information such as communication quality, task requirements, hardware constraints, etc. from the signal preprocessing and neural network preprocessing modules, and generates corresponding control signals to allocate PE resources under the guidance of the optimal scheduling algorithm, so as to minimize the total computing time, maximize the computing efficiency, and minimize the hardware overhead.

[0063] Dynamic precision adjustable PE module: The module integrates the control unit, computing unit and storage unit. The computing unit is composed of a series of flexibly configurable PE units. The control unit can flexibly configure the PE units according to constraints such as communication quality, hardware constraints, and task requirements to complete communication MIMO equalization and AI calculation. Under the guidance of the control signal from the computing resource scheduling module, the PE module performs MIMO equalization and AI calculation with appropriate resources.

[0064] Signal decoding module: used to decode the communication signal calculated by the PE module.

[0065] Neural network preprocessing module: used to preprocess the decoded communication information so that it can adapt to the subsequent AI calculation.

[0066] AI result output module: used to output the AI ​​calculation results of the PE module.

[0067] The working method of the above baseband processor includes two stages:

[0068] The first stage is the communication calculation stage, which mainly receives external signals and processes them. The processing includes at least one of coding, filtering, modulation, and MIMO equalization; preferably, filtering, modulation, and MIMO equalization. The main steps of the first stage are as follows:

[0069] 1. The signal input module accepts external communication signal input.

[0070] 2. The signal preprocessing module preprocesses the input signal, including encoding, modulation, filtering and other operations.

[0071] 3. The pre-processed signal is sent to the dynamically accurate adjustable PE module (ie, PE array) for MIMO equalization, completing the relevant calculations of communication.

[0072] 4. The signal decoding module decodes the communication signal after MIMO equalization.

[0073] The second stage is mainly the AI ​​calculation stage (i.e. neural network calculation), which mainly performs AI reasoning tasks on the signal after MIMO equalization. The main steps of the second stage are as follows:

[0074] 5. The decoded signal will be sent to the neural network preprocessing module for preprocessing to prepare for the subsequent AI calculation.

[0075] 6. The signal after neural network preprocessing is sent to the dynamically accurate and adjustable PE module (i.e., PE array) for AI calculation.

[0076] 7. The signal output module outputs the signal calculated by AI.

[0077] During the process, the computing resource scheduling module controls the dynamic precision and can adjust the resources allocated to the PE module for MIMO equalization and AI computing.

[0078] The present invention also provides an optimal computing resource scheduling algorithm, which takes task completion time as the goal and can allocate computing resources according to constraints such as communication quality, task requirements, hardware requirements, etc., so that the total computing time is minimized, the computing efficiency is maximized, and the hardware overhead is minimized.

[0079] A baseband processor resource scheduling algorithm comprises the following steps:

[0080] S1. Estimate the channel capacity model of the MIMO communication system, determine the communication calculation time and resource model of the baseband processor according to the channel capacity, and then determine the number of PE units allocated to MIMO equalization according to the communication calculation time and resource model.

[0081] S2. Determine the AI ​​computing time and resource model, and determine the number of PE units for AI computing based on the number of PE units evenly allocated by MIMO.

[0082] Furthermore, the channel capacity model of the MIMO communication system is as follows,

[0083] With N t Transmitting antenna and N r MIMO system with receiving antennas, N t ≤N r , the received symbol vector y is expressed as follows:

[0084] y=Hx+n (1)

[0085] H is a N r ×N t The complex channel matrix of x is the column vector containing the complex modulated signal, n is the additive white Gaussian noise with variance σ 2 = N0, the modulation symbol belongs to a size of 2 M A set of , M is the number of bits per modulation symbol.

[0086] For linear detectors, the MIMO channel is considered as a linear system. The estimation of the transmitted symbol vector x is achieved by directly multiplying the received symbol vector y with the estimation matrix G to cancel the effect of the channel. The estimation matrix G is usually based on criteria such as Minimum Mean Square Error (MMSE) and can be calculated by a large number of matrix operations involving intensive data path calculations. In Linear Minimum Mean-Square Error Detection (LMMSE), the detector minimizes the mean square error of the estimated symbols after equalization, which is defined as:

[0087]

[0088] Estimate the transmitted symbol vector using the LMMSE equalization method The calculation formula is:

[0089]

[0090] In the equilibrium stage, N t ×N r Matrix (H H H+N0I) to find the inverse, and then add H H Perform two matrix-vector multiplications with the received symbol vector y.

[0091] In actual working environments, due to limited hardware resources, the computational complexity of the communication system is limited. Therefore, to achieve an efficient communication system, it is necessary to optimize the computational complexity of the communication system by quantizing the channel matrix while ensuring the communication quality. The quantization of the channel matrix can be expressed as:

[0092] H≈H q =Q(H) (4)

[0093] Q(·) is a quantization function. Quantization greatly reduces the complexity of the system. By quantizing the N-bit channel matrix to an NM-bit matrix, the complexity of the multiplier is reduced to N(NM). Without loss of generality, the quantization function is equivalent to the following equation based on approximate calculation:

[0094] H=Hq +E q (5)

[0095] H q is the quantization matrix, E q is the quantization error matrix.

[0096] For an ideal baseband MIMO communication system, the transmitted signal can be expressed as equation (3), where the channel model H is given by Given, R R and R T are the normalized receive and transmit correlation Hermitian matrices with unit diagonal elements, respectively. Due to the existence of a lossless feedback link, that is, CSIT and CSIR are the same, both the transmitter and the receiver can obtain R R R T .H ω The elements of which are independent and identically distributed are N c (0, 1). x represents the transmission signal, which is given by:

[0097] x=Ws (6)

[0098] s is the transmitted symbol and W is the precoding matrix. The precoding matrix is:

[0099] W=(H H H+σ 2 I) -1 H H (7)

[0100] I is the unit matrix. The demodulated signal received by the receiving end is:

[0101] s rev =Wy (8)

[0102] For the channel matrix H, there are precision errors caused by quantization truncation in actual hardware implementation. For the convenience of theoretical analysis, these precision errors are equivalently expressed as rounding noise equations. Therefore, the quantized channel matrix H with quantization error q The relationship between it and the ideal channel matrix H can be expressed as:

[0103]

[0104] Where E ω is the rounding noise matrix. The rounding noise matrix E ω is a mean value μ q , the variance is The random matrix satisfies the following probability density function:

[0105]

[0106] Rounding noise Eω It also satisfies:

[0107]

[0108] Q represents the quantization bit width of the computation element.

[0109] From this we can deduce the rounding noise E ω The mean μ q for:

[0110]

[0111] Rounding noise E ω Variance for

[0112]

[0113] Where Ω is the length of the quantization interval, which is equal to 2 -(Q-1) Based on the channel quantization error model established by equations (1) and (5), the output of the channel can be written as:

[0114] y=H q x+E q x+n (14)

[0115] Therefore, the total noise of the system is expressed as n total =E q x+n, mean is 0, covariance matrix:

[0116]

[0117] Where: expectation and x, n, E ω related; due to the matrix E ω The elements of are independent and identically distributed We use the following lemma:

[0118]

[0119] Received signal y, transmitted signal x and quantized channel H q The mutual information between can be constrained as:

[0120]

[0121] Here, Q = E(xx H ), E represents the statistical expectation, and the upper bound of the mutual information can be expressed as

[0122]

[0123] make for The eigenvalue ofq The singular values ​​of According to the lemma, we can transform the lower bound expression of mutual information into:

[0124]

[0125] κ1 and ρ are coefficient factors, Q is the covariance matrix, is the singular value of the matrix (·), N r is the number of columns of the channel matrix (N r is the number of receiving antennas, calculated as the number of columns of the channel matrix), Q represents the quantization bit width of the calculation element, is the average singular value of the matrix (·), is the average eigenvalue of the matrix (·). The eigenvalues ​​of the total noise covariance matrix are

[0126]

[0127] ρ is a coefficient factor, ρ = tr(R T E(xx H )).

[0128] The upper bound of mutual information can be expressed as:

[0129]

[0130] For Gaussian inputs, the gap between the lower and upper bounds is usually small, so the channel capacity of the system is given by

[0131]

[0132] Substituting equation (20) into equation (23), the channel capacity of the system can be expressed as:

[0133]

[0134] is the matrix H q The average singular value of and The matrices Q and R are R The average eigenvalue of .

[0135] κ1 and ρ are coefficient factors, Q is the covariance matrix, is the singular value of the matrix (·), N r is the number of columns of the channel matrix, Q represents the quantization bit width of the calculation element, is the average singular value of the matrix (·), is the average eigenvalue of the matrix (·).

[0136] 2.1 Communication Computing Time and Resource Model

[0137] The channel capacity of the system can be expressed as:

[0138]

[0139] is the matrix H q The average singular value of and The matrices Q and R are R The average eigenvalue of .

[0140] For communication channels Its conjugate matrix H H Matrix operations with itself H require times multiplication and addition operations, matrix multiplication H H The resources required for H calculation are as follows:

[0141]

[0142] ξ(Q D ) represents the resource overhead consumed by a single PE unit. D is the bit width of the input data.

[0143] In the pulse array implementation of matrix multiplication, each PE unit needs to contain at least one complex multiplier and one arithmetic logic unit (ALU). The total number of PEs required for the communication process is:

[0144]

[0145] T cycle is the number of clock cycles. Communication T cycle The calculation time is limited by the amount of data and channel capacity, as follows:

[0146]

[0147] D is the amount of data input for communication, and C is the channel capacity.

[0148] 2.2 AI Computing Time and Resource Model

[0149] Neural networks can be mapped to specific hardware architectures through data flow graphs. In the data flow graph, each node represents a neural network operation, and each edge represents the direction of data flow during the neural network reasoning process. The reasoning time of neural network hardware can be expressed as:

[0150]

[0151] T represents the clock cycle, d represents the pipeline depth, and B is the batch size of the input data.

[0152] Γ is the topological matrix of the data flow graph, where each column represents a node in the data flow graph, each row represents an edge in the data flow graph, and each element represents the amount of data consumed or generated by the current node on the current edge, which is limited by the number of PEs in the hardware architecture. Each node in the data flow graph represents an operation in the neural network, that is, a layer, and the edges in the data flow graph represent the connection order of different layers in the neural network.

[0153] W is the workload vector matrix, where the i-th element represents the dimension of the feature matrix of the hardware module corresponding to the i-th node in the input data flow graph, which is determined by the size of the data input to the neural network and the structure of the neural network.

[0154] By dividing and rounding element by element, the resulting matrix Represents the number of clock cycles required for the hardware to complete the work corresponding to each element in the pipeline. The maximum value between the pipeline depth and the loop count represents the total number of cycles required to complete a prediction. Multiply the total number of cycles by the clock cycle length to get the total inference time.

[0155] To minimize the inference time, we need to determine the optimal workload vector W and topology matrix Γ.

[0156] For a specific neural network structure, given the dimension of the input data, the workload matrix can be completely determined, and the workload matrix can be directly derived

[0157]

[0158] For a specific neural network architecture, the overall structure of the data flow graph can be determined directly. Therefore, the dimension of the topology matrix Γ can be fully determined. The variable is the number of PEs assigned to each layer (convolution, pooling, and nonlinear activation).

[0159] To simplify the representation, the edges that are not related to each node in the original matrix are omitted. The original Γ can be expressed as

[0160]

[0161] P j represents the number of PEs required for the jth layer of the neural network. For a specific neural network pipeline partitioning strategy, the minimum inference time of the neural network is

[0162]

[0163] The sum of the i-th part represents the number of clock cycles corresponding to the number of layers of the neural network contained in the i-th stage of the pipeline.

[0164] For a certain neural network, a heuristic algorithm can be used to search for the optimal pipeline partitioning strategy to solve the above equation, and then The largest element of can be expressed as

[0165]

[0166] The value of α depends on the specific neural network structure and hardware architecture.

[0167] D represents the number of elements in the neural network input data feature matrix, Where W i Represents the optimal i-th pipeline stage.

[0168] P nn represents the number of PEs applied to the neural network. Then, the prediction time of the neural network can be calculated as

[0169]

[0170] The total time for the entire system to complete the communication and neural network computing tasks can be expressed as

[0171]

[0172] C represents the channel capacity, which is a function of the noise and the function of the quantization bit width Q, which is given by the previous channel capacity model, k represents a coefficient satisfying k<1, P represents the number of PEs in the entire system, and kP represents the number of PEs allocated to neural network calculations.

[0173] The minimum time model for communication and neural network calculation is

[0174]

[0175] ε and ξ(Q) are given. Since the PE units are assigned to two tasks, neural network calculation and communication, kP represents the number of PE units assigned to neural network calculation, and (1-k)P represents the number of PE units assigned to communication. It is ensured that the number of PEs allocated to communication meets the computing resource requirements of the communication part.

[0176] To minimize the total time, the values ​​of Q and k can be adjusted. k can have an upper bound

[0177]

[0178] The minimum total time is

[0179]

[0180] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above-mentioned subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0181] Example 1

[0182] like Figure 1 As shown, an embodiment of the present invention provides a baseband processor based on universal computing and integration, comprising: a signal input module, a signal preprocessing module, a PE resource scheduling module, a dynamic precision adjustable PE module, a signal decoding module, a neural network preprocessing module, and an AI result output module;

[0183] The signal input module is used to receive communication signals.

[0184] Specifically, the signal input module is connected to the signal preprocessing module, and transmits the received communication signal to the signal preprocessing module.

[0185] The signal preprocessing module performs signal preprocessing on the communication signal, and transmits the communication signal after the signal preprocessing to the dynamic precision adjustable PE module in a MIMO transmission mode.

[0186] Specifically, the signal preprocessing includes encoding, modulation, and filtering.

[0187] Specifically, the MIMO transmission mode is a mode with N t transmit antennas and N r receive antennas and (N t ≤N r ) of a MIMO transmission system; the N t 、N r is a positive integer, N t Represents the number of transmitting antennas, N r Represents the number of receiving antennas.

[0188] Specifically, the signal preprocessing module is connected to the PE resource scheduling module, and transmits the communication signal after the signal preprocessing to the PE resource scheduling module.

[0189] The dynamic precision can adjust the PE module to achieve dynamic allocation of PE units for MIMO equalization or neural network calculation; Figure 2 As shown, the dynamic precision adjustable PE module includes a control module, a storage module and a computing unit; the computing unit includes several PE units; the MIMO equalization and the neural network calculation share the PE unit.

[0190] Specifically, the MIMO equalization is used to optimize the communication signal that has completed the signal preprocessing; the MIMO equalization will use the linear minimum mean square error algorithm to equalize the communication signal that has completed the signal preprocessing; the neural network calculation will perform neural network calculation on the information content that has completed the data normalization processing to obtain a digital sequence; the MIMO equalization and the neural network calculation have a common calculation paradigm and can time-share multiplex the PE unit; the main purpose of the communication optimization is to reduce inter-code crosstalk.

[0191] Specifically, the MIMO equalization and the neural network calculation have a common calculation paradigm, and the PE unit can be time-division multiplexed, including: the essence of the MIMO equalization is to perform matrix multiplication operations, and the essence of the neural network calculation is also to perform matrix multiplication operations. The MIMO equalization and the neural network calculation have a common calculation paradigm, namely, matrix multiplication operations, and both the MIMO equalization and the neural network calculation can use PE units to perform matrix multiplication operations.

[0192] Specifically, the PE unit at least includes a complex multiplier and an arithmetic logic unit. When the PE unit performs MIMO equalization and neural network calculation, the calculation accuracy of multiplication calculation and addition calculation can be adjusted.

[0193] Specifically, Figure 2 As shown, the calculation accuracy of the multiplication and addition calculations can be adjusted, including: the accuracy of the multiplication calculation is dynamic, 2 bits multiplied by 2 bits is 4 bits, 4 bits multiplied by 4 bits is 8 bits, and so on, 8 bits multiplied by 8 bits is 16 bits, and the accuracy of the addition calculation is also dynamic, and the addition may overflow, so the addition of two 16 bits is temporarily stored in 20 bits, so the PE unit uses 20 bits to save the intermediate results when doing partial sum accumulation, and then cuts back to 16 bits. The use of adjustable calculation accuracy can save overhead when high-precision calculation is not required. Figure 2 The control module is connected to the storage module, which means that the control module can transmit the control signal to the storage module and control the resource allocation of PE. The two storage modules respectively store two elements of matrix multiplication to be executed. The storage module is connected to the PE array, which means that the storage module can transmit the data to the specified PE unit.

[0194] Specifically, the dynamically-adjustable PE module is connected to the signal decoding module, and transmits the communication signal after the MIMO equalization to the signal decoding module.

[0195] Specifically, the dynamic precision adjustable PE module is connected to the neural network output module, and transmits the calculation result of the neural network calculation to the neural network output module.

[0196] The signal decoding module performs a decoding operation on the communication signal that has completed the MIMO equalization.

[0197] Specifically, the signal decoding module is connected to the neural network preprocessing module, and transmits the communication signal that completes the decoding operation to the neural network preprocessing module.

[0198] The neural network preprocessing module performs data normalization processing on the information content carried by the communication signal that completes the decoding operation, and transmits the information content that has completed the data normalization processing to the dynamic precision adjustable module.

[0199] Specifically, the neural network preprocessing module is connected to the dynamic precision adjustable PE module, and transmits the information content of the completed data normalization processing to the dynamic precision adjustable PE module.

[0200] The AI ​​result output module receives the digital sequence obtained by completing the neural network calculation, restores the digital sequence into corresponding information content and displays the information content.

[0201] The PE resource scheduling module is used to receive the communication signal that has completed the signal preprocessing, and obtain the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation based on the optimal scheduling algorithm, and transmit the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation to the dynamic precision adjustable PE module.

[0202] Specifically, the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation are used to dynamically allocate PE units to the dynamic precision adjustable PE module.

[0203] Specifically, the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation are used to guide the dynamic precision adjustable PE module to dynamically allocate PE units, including: the dynamic precision adjustable PE module receives the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation; the control module of the dynamic precision adjustable PE module dynamically allocates PE units according to the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation.

[0204] Specifically, the control module of the dynamically-adjustable PE module dynamically allocates PE units according to the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation, including: the control module allocates the same number of PE units as the number of PE units allocated to MIMO equalization or the number of PE units allocated to neural network calculation, for MIMO equalization or neural network calculation.

[0205] Specifically, the method of obtaining the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation based on the optimal scheduling algorithm includes: calculating the channel capacity under the MIMO transmission mode; calculating the time required for MIMO equalization based on the channel capacity; and obtaining the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation based on the time required for MIMO equalization and the resources required for MIMO equalization.

[0206] Specifically, the calculating of the channel capacity under the MIMO transmission mode includes:

[0207] For N t transmit antennas and Nr receive antennas and (N t ≤N r ), the received symbol vector y can be expressed as:

[0208] y=Hx+n (1)

[0209] Where y is the received symbol vector, H is N r ×N t The channel matrix of , x is the transmitted symbol vector, n is the additive Gaussian white noise, and its variance is The modulation symbol belongs to a size of 2 M A set of , M is the number of bits per modulation symbol.

[0210] In actual working environments, the computational complexity of communication systems is limited due to limited hardware resources. Therefore, to achieve an efficient communication system, it is necessary to optimize the computational complexity of the communication system by quantizing the channel matrix while ensuring the communication quality.

[0211] The quantization of the channel matrix can be expressed as:

[0212] H=H q +E q (2)

[0213] Among them, H q is the quantization matrix, E q is the quantization error matrix.

[0214] Quantization matrix H qAnd the quantization error matrix E q It can be expressed as:

[0215]

[0216] Among them, R R is the unit diagonal element normalized receiver Hermitian matrix, R T is the unit diagonal element normalized emission Hermitian matrix, H w is the channel precoding matrix, E w is the rounding noise matrix.

[0217] Among them, the rounding noise matrix E w is a mean value μ q , the variance is The random matrix satisfies the following probability density function:

[0218]

[0219] And the rounded noise matrix E w Also meets:

[0220]

[0221] Among them, Q d is a quantization bit width of a calculation element, where the calculation element is an element of a channel matrix.

[0222] Then the rounded noise matrix E w The mean μ q for:

[0223]

[0224] Rounding noise matrix E w Variance for:

[0225]

[0226] Where Ω is the length of the quantized interval, which is equal to

[0227] From (1) and (2), the received symbol vector y can be expressed as:

[0228] y=H q x+E q x+n

[0229] The total noise can be expressed as:

[0230] n total =E q x+n

[0231] Then the covariance matrix of the total noise can be expressed as:

[0232]

[0233] R T is the covariance matrix of the normalized emission Hermit matrix with unit diagonal elements, R R is the covariance matrix of the Hermit matrix with unit diagonal elements, The dimension is N r ×N r The identity matrix of .

[0234] Then the eigenvalues ​​of the covariance matrix of the total noise are:

[0235]

[0236] ρ is a coefficient factor, ρ = tr(R T (xx H ).

[0237] Receive symbol vector y, transmit symbol vector x and quantization matrix H q The mutual information between can be constrained as:

[0238]

[0239] σ i (·) represents the eigenvalue of the matrix (·), H q represents the quantization matrix, is the eigenvalue of the covariance matrix Q, is the eigenvalue of the total noise covariance matrix.

[0240] For Gaussian inputs, the gap between the lower and upper bounds is usually small, so the channel capacity can be expressed as:

[0241]

[0242] Substituting the eigenvalue of the covariance matrix of the total noise into the equation, we can obtain:

[0243]

[0244] The calculation formula of the channel capacity is as follows:

[0245]

[0246] Where C is the channel capacity; k1 and ρ are coefficient factors; N r is the number of rows of the channel matrix, i.e. the number of receiving antennas; is the quantization error matrix H q The average singular value of is the average eigenvalue of the covariance matrix Q; Q d is the quantization bit width of the calculation element; R R Normalize the receiving Hermit matrix to the unit diagonal element; is the variance of additive white Gaussian noise.

[0247] Specifically, according to the channel capacity, the time required for MIMO equalization is calculated, including:

[0248] The time required for MIMO equalization is related to the amount of data input and channel capacity of the communication. The specific calculation formula is as follows:

[0249]

[0250] Among them, T cycle is the time required for MIMO equalization; D is the amount of data input for communication; C is the channel capacity.

[0251] Specifically, according to the time required for MIMO equalization and the resources required for MIMO equalization, the number of PE units allocated to MIMO equalization and the number of PE units allocated to neural network calculation are obtained, including:

[0252] Assume that the channel matrix H is obtained under MIMO transmission mode, where the channel matrix H is a matrix with N r The number of rows and N t The channel matrix H and its conjugate matrix H H To perform matrix multiplication, a total of times multiplication and The calculation formula for the resources required to achieve MIMO equalization is as follows:

[0253]

[0254] Among them, ε is the resource required for MIMO equalization, N t is the number of columns of the channel matrix H, N r is the number of rows in the channel matrix, ξ(Q d ) is the quantization bit width of the calculation element, Q d When performing MIMO equalization, the resource overhead of a single PE unit is calculated; the calculation elements are all elements for performing MIMO equalization.

[0255] The calculation formula for the number of PE units allocated to MIMO equalization is as follows:

[0256]

[0257] Among them, P comis the number of PE units allocated to MIMO equalization; ε is the resources required for MIMO equalization; N t is the number of rows of the channel matrix, i.e. the number of transmitting antennas; T cycle is the time required for MIMO equalization; ξ(Q) is the resource overhead consumed by a single PE unit when performing MIMO equalization when the actual bit width of the calculation element is Q; the calculation element is all the elements for MIMO equalization.

[0258] The number of PE units allocated to the neural network calculation is restricted by the number of PE units required for MIMO equalization. Because the PE units allocated to MIMO equalization must meet the number of PE units required for MIMO equalization, the calculation formula for the number of PE units allocated to the neural network calculation is as follows:

[0259] P nnc =PP com

[0260] Among them, P nnc is the number of PE units allocated to neural network calculations, P is the total number of PE units; P com The number of PE units required for MIMO equalization.

[0261] like Figure 3 This is a working timing diagram of the baseband processor based on universal computing. The communication time and neural network prediction time are carried out in sequence. By optimizing the number of PE units allocated to MIMO equalization and neural network algorithm calculation, the overall time is minimized, and the resource allocation of the two is optimized to achieve the purpose of optimal efficiency.

Claims

1. A baseband processor based on universal computing, characterized in that: include: A signal input module, used for receiving communication signals; A signal preprocessing module performs signal preprocessing on the communication signal received by the signal input module, and then transmits the preprocessed communication signal to the dynamic precision adjustable PE module; Computing resource scheduling module, used to schedule the dynamically precision adjustable PE module; Dynamic precision adjustable PE module for MIMO equalization and / or AI calculation; The dynamic precision adjustable PE module comprises: a control unit, a storage unit and a calculation unit, wherein the calculation unit comprises a plurality of PE units; the control unit controls the calculation unit to read the storage unit data for calculation processing; A signal decoding module, used for receiving the communication signal after MIMO equalization by the dynamic precision adjustable PE module, and performing decoding operation; A neural network preprocessing module is used to preprocess the information content obtained by the decoding operation to obtain a signal for neural network calculation; then, the signal is input into the dynamic precision adjustable PE module to perform AI calculation and obtain an AI calculation result; The AI ​​result output module is used to output the AI ​​calculation results.

2. The baseband processor based on universal computing and integration according to claim 1, characterized in that: The MIMO equalization is: the dynamic precision adjustable PE module uses a linear minimum mean square error algorithm to equalize the communication signal after signal preprocessing; The AI ​​calculation is: performing neural network calculation on the information content of the data normalization process to obtain a digital sequence.

3. The baseband processor based on universal computing and integration according to claim 1, characterized in that: The computing resource scheduling module can control the dynamic precision adjustable PE module to allocate PE unit resources and complete the calculation according to the communication quality, hardware constraints and task requirements.

4. The baseband processor based on universal computing and integration according to claim 3 is characterized in that: The computing resource scheduling module can control the number of PE units allocated to MIMO equalization by the dynamic precision adjustable PE module and the number of PE units allocated to AI calculation.

5. The baseband processor based on universal computing and integration according to claim 1, characterized in that: The PE unit at least includes a complex multiplier and an arithmetic logic unit; when the PE unit performs MIMO equalization and AI calculation, the calculation accuracy is adjustable.

6. A method for performing integrated computing and control using the baseband processor according to any one of claims 1 to 5, characterized in that: It includes two stages: communication computing stage and AI computing stage; The first stage, the communication and computing stage: receiving communication signal input; Preprocess the input signal; The pre-processed signal is sent to the dynamic precision adjustable PE module for MIMO equalization; Decoding the communication signal after MIMO equalization; The second stage, AI computing stage: Send the decoded signal to the neural network preprocessing module for preprocessing; The signal after neural network preprocessing is sent to the dynamic precision adjustable PE module for AI calculation; Output the results of AI calculation.

7. The method for performing common computing and integration control of a baseband processor according to claim 6, characterized in that: Calculate the channel capacity of the MIMO communication system, determine the time required to calculate MIMO equalization based on the channel capacity, and determine the number of PE units allocated to MIMO equalization.

8. The method for performing common computing and integration control of a baseband processor according to claim 7, characterized in that: The calculation formula of the channel capacity is as follows: Where C is the channel capacity; k1 and ρ are coefficient factors; N r is the number of rows of the channel matrix, i.e. the number of receiving antennas; is the quantization error matrix H q The average singular value of is the average eigenvalue of the covariance matrix Q; Qd is the quantization bit width of the calculation element; RR is the unit diagonal element normalized receiving Hermit matrix; is the variance of additive Gaussian white noise; The calculation formula for the time required for MIMO equalization is as follows: Among them, T cycle is the time required for MIMO equalization; D is the amount of data input for communication; C is the channel capacity; The calculation formula for the number of PE units allocated to MIMO equalization is as follows: Among them, P com is the number of PE units allocated to MIMO equalization; ε is the resources required for MIMO equalization; Nt is the number of columns in the channel matrix, i.e., the number of transmitting antennas; T cycle is the time required for MIMO equalization; ξ(Q) is the resource overhead consumed by a single PE unit when performing MIMO equalization when the actual bit width of the calculation element is Q.

9. The method for performing common computing and integration control of a baseband processor according to claim 8, characterized in that: The number of PE units allocated to AI calculation is determined according to the number of PE units allocated to MIMO equalization.

10. The method for performing common computing and integration control of a baseband processor according to claim 9, characterized in that: The calculation formula for the number of PE units allocated to AI calculation is as follows: P nnc =P-P com Among them, P nnc is the number of PE units allocated to AI calculation, P is the total number of PE units; P com is the number of PE units allocated to MIMO equalization.

Images