A Low-Complexity Hardware Implementation Method for Millimeter-Wave Multi-User Hybrid Precoding Based on Equivalent Channel Dimensionality Reduction

By adopting a low-complexity millimeter-wave multi-user hybrid precoding method based on equivalent channel dimensionality reduction, the problem of complex channel matrix processing in the prior art is solved, and efficient real-time processing on the FPGA platform is achieved, which is suitable for the hardware configuration of RF phase shifter networks.

CN123074536APending Publication Date: 2026-11-10CHONGQING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611241679.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing millimeter-wave multi-user MIMO-OFDM systems in broadband multi-user scenarios suffer from problems such as complex channel matrix processing, difficulty in coordinating analog domain constant modulus constraints and digital domain matrix decomposition, and difficulty in controlling fixed-point output errors. These issues lead to limitations in real-time baseband processing due to hardware resource consumption and computational delay.

Method used

A low-complexity millimeter-wave multi-user hybrid precoding method based on equivalent channel dimensionality reduction is adopted. Through angle domain constant modulus simulation weight generation, two-level equivalent channel matrix calculation, low-dimensional SVD digital weight generation, and multi-subcarrier fixed-point output, combined with the pipelined processing flow of FPGA, data access and buffer scheduling, constant modulus simulation weight construction, equivalent channel dimensionality reduction calculation, and SVD digital weight extraction are realized, which is adapted to RF phase shifter networks.

Benefits of technology

It reduces the complexity of multi-subcarrier matrix data scheduling, reduces hardware resource consumption and wiring congestion, and improves the system's real-time processing capability and hardware deployment efficiency, making it suitable for the real-time baseband processing requirements of FPGA platforms.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074536A_ABST
    Figure CN123074536A_ABST
Patent Text Reader

Abstract

This invention discloses a low-complexity millimeter-wave multi-user hybrid precoding hardware implementation method based on equivalent channel dimensionality reduction, belonging to the field of wireless communication technology. The method includes five steps: data access and buffer scheduling, constant modulus analog weight construction, equivalent channel dimensionality reduction calculation, SVD digital weight extraction, and multi-subcarrier fixed-point output. First, user angle information and multi-subcarrier channel data are received, and buffer organization and subcarrier scheduling are completed. Second, the user angle information is fixed-point mapped to phase step values, and a constant modulus complex array response vector is generated through phase accumulation and CORDIC operations to construct the transmitter's analog precoding weights and the receiver's analog combining weights. Then, a two-level complex multiplication-accumulation structure is used to complete the channel matrix dimensionality reduction, obtaining a low-dimensional equivalent channel matrix. Further, the low-dimensional equivalent channel matrix is ​​subjected to Jacobi rotation-based SVD pipelined decomposition to extract the digital precoding matrix and digital combining matrix. Finally, fixed-point format alignment, rounding, saturation processing, and pipelined output are completed according to subcarrier indices and matrix element indices. This method balances computational accuracy, resource consumption, and real-time processing requirements, and is suitable for millimeter-wave MIMO-OFDM baseband signal processing design and development.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and in particular relates to a low-complexity millimeter-wave multi-user hybrid precoding hardware implementation method based on equivalent channel dimensionality reduction. Background Technology

[0002] Millimeter-wave multi-user multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems can improve the spectral efficiency and link capacity of wireless communication systems through massive MIMO antenna arrays, spatial multiplexing, and wideband multi-subcarrier transmission, making them an important technology direction for high-throughput, low-latency wireless communication. Hybrid precoding, by combining analog domain beamforming and digital domain precoding, achieves a trade-off between system performance, number of RF links, and hardware complexity, making it suitable for millimeter-wave massive MIMO antenna systems.

[0003] Existing hybrid precoding methods mainly include codebook search, orthogonal matching pursuit, alternating minimization, equivalent channel, block diagonalization, and singular value decomposition. While these methods can improve system spectral efficiency or reduce algorithm complexity to some extent, they still require handling complex channel matrices corresponding to multiple OFDM subcarriers, analog domain constant modulus constraints, digital domain matrix decomposition, matrix iterative updates, and fixed-point quantization in broadband multi-user scenarios. If implemented using a software serial approach or a low-parallelism hardware approach, these methods are easily limited by computational latency, memory access, data scheduling, and hardware resource consumption, making it difficult to meet the requirements of real-time baseband processing.

[0004] Field-Programmable Gate Arrays (FPGAs), with their reconfigurability, parallel processing, and on-chip data organization capabilities, have been widely used in high-speed communication baseband processing. For precoding processes involving multi-level matrix operations and continuous data transfer, implementation tailored to hardware architecture characteristics helps achieve a balance between algorithm performance and hardware overhead. By rationally designing data buffering, computational pipelining, fixed-point representation, and resource reuse methods, the system's real-time processing capabilities and hardware deployment efficiency can be improved, providing support for the implementation of hybrid precoding engineering in millimeter-wave multi-user communication systems. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose a low-complexity millimeter-wave multi-user hybrid precoding hardware implementation method based on equivalent channel dimensionality reduction. This method addresses the problems of complex multi-user multi-subcarrier data organization, large complex matrix operations, difficulty in coordinating analog domain constant modulus constraints and digital domain matrix decomposition, and difficulty in controlling fixed-point output errors. It transforms the angle domain constant modulus analog weight generation, two-level equivalent channel matrix calculation, low-dimensional SVD digital weight generation, and multi-subcarrier fixed-point output organization into a pipelined process suitable for FPGA implementation.

[0006] To achieve the above objectives, the technical solution adopted in this invention is: a low-complexity millimeter-wave multi-user hybrid precoding hardware implementation method based on equivalent channel dimensionality reduction. This method takes user angle information and multi-subcarrier channel data as input, and uses the target fully digital precoding matrix of each subcarrier as the iteration target when iterative update processing is enabled; it completes data access, partition management, and processing order control through interface and cache scheduling; it generates constant-modulus analog domain weights through angle fixed-point mapping, CORDIC phase step calculation, phase accumulation, and CORDIC rotation operations; it reduces the dimensionality of subsequent digital domain processing through a locally parallel, globally serial two-level complex multiplication and accumulation structure; it generates digital precoding matrices and digital merging matrices through SVD pipelined decomposition based on bilateral Jacobi rotation; when iterative update processing is enabled, it updates the transmitter's analog precoding matrix and repeats the equivalent channel calculation and SVD digital weight extraction until the preset number of analog weight update iterations is reached; and it continuously outputs the multi-subcarrier weights through a fixed-point output module. The method includes the following steps:

[0007] S101: Data Access and Cache Scheduling. The system receives departure angle information, arrival angle information, multi-subcarrier channel data, and configuration control information for each user through the FPGA top-level interface. When iterative update processing is enabled, it receives the target all-digital precoding matrix for each subcarrier. User angle information and multi-subcarrier channel data are written to the input cache. When iterative update processing is enabled, the target all-digital precoding matrix is ​​also written to the input cache. Intermediate matrix data, SVD decomposition result data, and output weight data generated during processing are written to their respective caches. Cache read / write addresses, data validity identifiers, and processing order control signals are generated based on user index, subcarrier index, and matrix element index.

[0008] S102: Construction of Constant Mode Simulation Weights. The departure angle and arrival angle information for each user are mapped to fixed-point phase steps. Phase accumulation is performed based on the antenna element index, and a complex array response vector satisfying constant mode constraints is generated through CORDIC operations. Transmitter simulation precoding weights and a transmitter simulation precoding matrix are constructed based on the complex array response vector corresponding to the departure angle. Receiver simulation combining weights and a receiver simulation combining matrix are constructed based on the complex array response vector corresponding to the arrival angle. When the number of RF links exceeds the number of users, extended constant mode weights corresponding to the remaining RF links are generated through a pseudo-random phase sequence and written into the RF vector buffer.

[0009] S103: Equivalent Channel Dimensionality Reduction Calculation. Based on the transmitter's analog precoding matrix, receiver's analog combining matrix, and multi-subcarrier channel data, a two-level complex multiplication and accumulation structure with local parallelism and overall serialism is adopted. First, the product of each subcarrier channel matrix and the transmitter's analog precoding matrix is ​​calculated to obtain the transmitter's dimension-reduced intermediate matrix. Then, the product of the conjugate transpose of the receiver's analog combining matrix and the transmitter's dimension-reduced intermediate matrix is ​​calculated to obtain the low-dimensional equivalent channel matrix corresponding to each OFDM subcarrier.

[0010] S104: SVD Digital Weight Extraction. The low-dimensional equivalent channel matrix corresponding to each OFDM subcarrier is input into the SVD pipeline decomposition module based on bilateral Jacobi rotation. Through multiple rounds of Jacobi rotation iteration, the left singular matrix, singular value matrix, and right singular matrix are obtained. According to the number of data streams, column vectors corresponding to larger singular values ​​are extracted from the right and left singular matrices respectively, forming a digital precoding matrix and a digital combining matrix. The digital precoding matrix is ​​power normalized so that the product of the transmitter's analog precoding matrix and the digital precoding matrix satisfies a preset transmit power constraint. The digital precoding matrix and the digital combining matrix are then written into the decomposition result cache.

[0011] S105: Multi-subcarrier fixed-point output. The digital precoding matrix, digital combining matrix, transmitter analog precoding weights, and receiver analog combining weights are subjected to bit-width expansion, fixed-point format alignment, rounding, truncation, and saturation processing, and output sequentially according to the subcarrier index and matrix element index. Specifically, the output data, data validity flag, subcarrier index flag, and matrix element index flag are pipelined with delay alignment to ensure that the output data is synchronized with each flag signal.

[0012] Based on the above technical solution, the transmitter simulates the precoding matrix. The matrix elements satisfy the following constant modulus constraint:

[0013]

[0014] The elements of the analog merging matrix at the receiving end satisfy the following constant modulus constraint:

[0015]

[0016] in, and These are the constant-mode amplitudes at the transmitting and receiving ends, respectively. and These are the phases of the corresponding matrix elements. and These represent the row index and column index of the matrix element, respectively. The imaginary unit, It is a natural constant. This constraint is used to characterize that the analog domain weights are mainly formed by phase control, adapting to the hardware implementation of RF phase shifter networks.

[0017] Based on the above technical solution, the equivalent channel matrix calculation stage can map the high-dimensional channel matrix, the transmitter's simulated precoding weights, and the receiver's simulated combining weights into a low-dimensional equivalent channel matrix. For the first... The equivalent channel relationship of each OFDM subcarrier can be expressed as:

[0018]

[0019] in, Indicates the first The multi-user channel matrix is ​​formed by vertically concatenating the user channel matrices corresponding to each subcarrier according to the user index. This represents the transmitter's analog precoding matrix. This represents the block diagonal receiver simulation merge matrix, composed of the simulated merge submatrices of each user receiver according to their user indices. Its conjugate transpose is used for receiver merging. Indicates the first The low-dimensional equivalent channel matrix corresponding to each subcarrier. For ease of hardware pipelining, the above calculation is broken down into two stages of complex matrix multiplication:

[0020]

[0021] in, This represents the intermediate matrix after dimensionality reduction at the transmitter. The first-level matrix operation completes the transmitter dimensionality reduction between the channel data and the analog precoding weights, while the second-level matrix operation completes the receiver merging between the analog merging weights and the intermediate matrix.

[0022] In one specific implementation, the transmitting end is equipped with 144 antennas and the receiving end with 4 antennas, with 3 radio frequency links configured at both ends. For each subcarrier, according to the aforementioned two-level equivalent channel relationship, the transmitting end first uses simulated precoding weights to complete the dimensionality reduction, and then the receiving end uses simulated combining weights to complete the combining, so that the multi-user channel matrix is ​​successively reduced from 4 rows and 144 columns to a 4 rows and 3 columns intermediate matrix and a 3 rows and 3 columns low-dimensional equivalent channel matrix. Subsequent singular value decomposition is only performed on the 3 rows and 3 columns low-dimensional equivalent channel matrix, reducing the number of input complex elements from 576 to 9, which is 1 / 64 of the original.

[0023] In the above configuration, the first stage uses 12 parallel complex multiply-accumulate units, corresponding to the 12 independent elements of the intermediate matrix, and accumulates them clockwise along the 144 antenna dimensions. The processing latency of the first stage, including delays such as complex multiplication pipeline, is approximately 150 clock cycles. The internal accumulation length of the second stage is 4 clock cycles. This time-division multiplexing method, which is locally parallel and globally serial, achieves the aforementioned dimensionality reduction while avoiding the large number of complex multipliers and wiring resources required for fully parallel operations on high-dimensional matrices.

[0024] Based on the above technical solution, the digital weight generation stage can perform singular value decomposition on the low-dimensional equivalent channel matrix corresponding to each subcarrier, and its matrix decomposition relationship can be expressed as:

[0025]

[0026] in, It is a left singular matrix. It is a singular value matrix. This is a right singular matrix. The number of singular values ​​corresponding to larger singular values ​​equals the number of data streams. The singular vectors can be used to generate digital precoding matrices:

[0027] The digital precoding matrix and the digital merging matrix can be represented as:

[0028]

[0029] in, Indicates the first The digital precoding matrix corresponding to each subcarrier The number of data streams is represented by the digital precoding matrix, which consists of column vectors in the right singular matrix corresponding to the larger singular values, with the number of vectors equal to the number of data streams. The digital merging matrix... Left singular matrix The matrix is ​​composed of column vectors extracted according to the same index, and divided into user-defined digital merged sub-matrices according to user indexes and output separately; the digital precoding matrix is ​​power normalized so that the product of the transmitter's analog precoding matrix and the digital precoding matrix satisfies the preset transmit power constraint.

[0030] Based on the above technical solution, an iterative update process is further included after step S104. The transmitter analog precoding matrix generated in step S102 is used as the initial matrix for the 0th iteration, and the iteration index is set to 0; in the... In this iteration, steps S103 and S104 calculate the low-dimensional equivalent channel matrix and the digital precoding matrix of each subcarrier based on the current transmitter analog precoding matrix. When the preset number of simulation weight update iterations is reached, step S105 is executed; when If the preset number of simulation weight update iterations has not been reached, an update matrix is ​​constructed based on the current digital precoding matrix and the target all-digital precoding matrix. This involves sequentially performing conjugate transpose, matrix multiplication, small-scale matrix inversion, and constant modulus projection to obtain the transmitter simulation precoding matrix for the next iteration. The process then returns to step S103. The update relationship can be expressed as:

[0031]

[0032] in, Indicates the first The transmitter simulation precoding matrix obtained in the next iteration This represents the iteration count index, initialized to 0. This represents the initial transmitter simulation precoding matrix generated in step S102. This represents the joint target matrix formed by horizontally concatenating the target fully digital precoding matrices of each subcarrier in subcarrier index order. Indicates by the first In the next iteration, the current digital precoding matrices of each subcarrier are horizontally concatenated in the same order to form a joint digital precoding matrix. This indicates phase extraction or constant mode projection processing under constant mode constraints; when the preset number of simulation weight update iterations is reached, step S105 outputs the current corresponding simulation domain weights and digital domain weights.

[0033] Based on the above technical solution, the interface and cache scheduling stage includes angle information input, channel data input, target matrix input, configuration control, and weight output interfaces; the cache includes an input cache, intermediate matrix cache, decomposition result cache, radio frequency vector cache, and output cache, and performs partitioned read / write and processing order control according to user index, subcarrier index, and matrix element index; the input cache is used to save the full digital precoding matrix of each subcarrier target when iterative update processing is enabled, and the decomposition result cache is also used to save the digital weights and intermediate data of each iteration.

[0034] The present invention has the following beneficial effects:

[0035] (1) The present invention organizes the millimeter-wave multi-user MIMO-OFDM hybrid precoding process into an FPGA pipelined processing structure of angle domain constant modulus weight generation, two-level equivalent channel matrix calculation, low-dimensional SVD digital weight generation and multi-subcarrier fixed-point output, thereby improving the deployability of the hybrid precoding algorithm on resource-constrained FPGAs.

[0036] (2) This invention enables angle information, channel matrix, intermediate matrix, SVD decomposition results and output weights to maintain correspondence between different processing stages by using cache partitioning and processing order control for user index, subcarrier index and matrix element index, thereby reducing the complexity of multi-subcarrier matrix data scheduling.

[0037] (3) The present invention generates analog domain weights by angle fixed-point mapping, CORDIC phase step calculation, phase accumulation and CORDIC rotation operation, so that the analog precoding weights and analog merging weights satisfy the RF phase shifter network constraints and can be adapted to hardware configurations where the number of RF links is greater than the number of users.

[0038] (4) This invention transforms high-dimensional channel matrix processing into low-dimensional matrix processing through a two-stage complex multiplication-accumulation structure and a low-dimensional equivalent channel matrix decomposition method, avoiding excessive multiplier resource consumption and wiring congestion caused by directly using fully parallel high-dimensional complex matrix multiplication. In the above specific configuration, the number of input complex elements for subsequent matrix decomposition is reduced from 576 to 9, which is 1 / 64 of the original.

[0039] (5) This invention enables digital domain weights to be output in the order of OFDM subcarriers through unified fixed-point format, pipelined output and resource reuse processing, and enables analog domain weights to be output synchronously or reused as shared weights under the corresponding user or frame configuration, which is suitable for deployment on FPGA prototype verification platforms with real-time processing requirements. Attached Figure Description

[0040] Figure 1 This diagram illustrates the steps of the low-complexity millimeter-wave multi-user hybrid precoding hardware implementation method based on equivalent channel dimensionality reduction proposed in this invention.

[0041] Figure 2 This is a diagram of the FPGA hardware implementation architecture of the hybrid precoding method of the present invention, wherein, Figure 2 (a) shows the overall architecture of the FPGA hardware implementation of the hybrid precoding method. Figure 2 (b) is an FPGA architecture for a two-level equivalent channel dimensionality reduction calculation module. Figure 2 (c) is the FPGA architecture of the low-dimensional equivalent channel Jacobi SVD decomposition module.

[0042] Figure 3 This is a timing simulation result diagram of the core module of the present invention, wherein, Figure 3 (a) shows the timing simulation results of the two-level equivalent channel dimensionality reduction calculation module. Figure 3 (b) shows the timing simulation results of the Jacobi SVD decomposition module for the low-dimensional equivalent channel. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the FPGA hardware implementation process of the present invention, and are not intended to limit the present invention to a specific device model, interface name, cache depth, fixed bit width, or hardware implementation form.

[0044] like Figure 1 As shown in the figure, this embodiment of a low-complexity millimeter-wave multi-user hybrid precoding hardware implementation method based on equivalent channel dimensionality reduction includes the following steps:

[0045] S101: Data Access and Cache Scheduling. The system receives departure angle information, arrival angle information, multi-subcarrier channel data, and configuration control information for each user through the FPGA top-level interface. When iterative update processing is enabled, it receives the target all-digital precoding matrix for each subcarrier. User angle information and multi-subcarrier channel data are written to the input cache. When iterative update processing is enabled, the target all-digital precoding matrix is ​​also written to the input cache. Intermediate matrix data, SVD decomposition result data, and output weight data generated during processing are written to their respective caches. Cache read / write addresses, data validity identifiers, and processing order control signals are generated based on user index, subcarrier index, and matrix element index.

[0046] S102: Construction of Constant Mode Simulation Weights. The departure angle and arrival angle information for each user are mapped to fixed-point phase steps. Phase accumulation is performed based on the antenna element index, and a complex array response vector satisfying constant mode constraints is generated through CORDIC operations. Transmitter simulation precoding weights and a transmitter simulation precoding matrix are constructed based on the complex array response vector corresponding to the departure angle. Receiver simulation combining weights and a receiver simulation combining matrix are constructed based on the complex array response vector corresponding to the arrival angle. When the number of RF links exceeds the number of users, extended constant mode weights corresponding to the remaining RF links are generated through a pseudo-random phase sequence and written into the RF vector buffer.

[0047] S103: Equivalent Channel Dimensionality Reduction Calculation. Based on the transmitter's analog precoding matrix, receiver's analog combining matrix, and multi-subcarrier channel data, a two-level complex multiplication and accumulation structure with local parallelism and overall serialism is adopted. First, the product of each subcarrier channel matrix and the transmitter's analog precoding matrix is ​​calculated to obtain the transmitter's dimension-reduced intermediate matrix. Then, the product of the conjugate transpose of the receiver's analog combining matrix and the transmitter's dimension-reduced intermediate matrix is ​​calculated to obtain the low-dimensional equivalent channel matrix corresponding to each OFDM subcarrier.

[0048] S104: SVD Digital Weight Extraction. The low-dimensional equivalent channel matrix corresponding to each OFDM subcarrier is input into the SVD pipeline decomposition module based on bilateral Jacobi rotation. Through multiple rounds of Jacobi rotation iteration, the left singular matrix, singular value matrix, and right singular matrix are obtained. According to the number of data streams, column vectors corresponding to larger singular values ​​are extracted from the right and left singular matrices respectively, forming a digital precoding matrix and a digital combining matrix. The digital precoding matrix is ​​power normalized so that the product of the transmitter's analog precoding matrix and the digital precoding matrix satisfies a preset transmit power constraint. The digital precoding matrix and the digital combining matrix are then written into the decomposition result cache.

[0049] S105: Multi-subcarrier fixed-point output. The digital precoding matrix, digital combining matrix, transmitter analog precoding weights, and receiver analog combining weights are subjected to bit-width expansion, fixed-point format alignment, rounding, truncation, and saturation processing, and output sequentially according to the subcarrier index and matrix element index. Specifically, the output data, data validity flag, subcarrier index flag, and matrix element index flag are pipelined with delay alignment to ensure that the output data is synchronized with each flag signal.

[0050] In step S101, the FPGA top-level interface includes an angle information input interface, a channel data input interface, a target matrix input interface, a configuration control interface, and a weight output interface. The angle information input interface receives the departure angle and arrival angle information for each user; the channel data input interface receives the complex channel matrices corresponding to different OFDM subcarriers; the target matrix input interface receives the target all-digital precoding matrix for each subcarrier when iterative update processing is enabled; the configuration control interface receives configuration parameters such as the number of users, the number of RF links, the number of data streams, the number of subcarriers, the matrix dimension, the number of SVD iterations, the number of analog weight update iterations, and the fixed-point output bit width; the weight output interface outputs the analog precoding weights, the analog combining weights, the digital precoding matrix, and the digital combining matrix.

[0051] The buffer scheduling process includes an input buffer, an intermediate matrix buffer, a decomposition result buffer, a radio frequency vector buffer, and an output buffer. The input buffer stores angle information, the multi-subcarrier channel matrix, and the fully digital precoding matrices of each subcarrier target received when iterative update processing is enabled. The intermediate matrix buffer stores the transmitter's dimensionality-reduced intermediate matrix and the low-dimensional equivalent channel matrix. The decomposition result buffer stores the left singular matrix, singular value matrix, right singular matrix, extracted digital precoding matrix, and digital merging matrix obtained from SVD decomposition, as well as the intermediate results of each iteration. The radio frequency vector buffer stores constant modulus analog weights and extended radio frequency link weights. The output buffer organizes the output weight data according to the subcarrier order and matrix element order.

[0052] like Figure 2 As shown in (a), the overall architecture of this FPGA is similar to... Figure 1 The method described is as follows: First, the system generates and caches the constant modulus weights of the analog domain for the transmitter and receiver based on the user's departure angle and arrival angle. Then, it uses the analog domain weights to perform dimensionality reduction on the transmitter side and merging on the receiver side for each subcarrier channel to obtain a low-dimensional equivalent channel. Subsequently, it performs JacobiSVD decomposition on the low-dimensional equivalent channel to extract digital precoding weights and digital merging weights. If iterative updates are enabled, the transmitter's analog precoding weights are updated by the target all-digital precoding matrix and the current digital precoding weights, and fed back to the dimensionality reduction path of the equivalent channel. Finally, it outputs the analog domain weights and digital domain weights after fixed-point processing.

[0053] in, Figure 2 (b) is a two-level equivalent channel dimensionality reduction calculation module, which focuses on demonstrating how to compress a high-dimensional channel into a low-dimensional equivalent channel by simulating precoding weights at the transmitter and simulating combining weights at the receiver; Figure 2 (c) is the Jacobi SVD decomposition module for the low-dimensional equivalent channel, focusing on the singular value decomposition and digital weight extraction of the low-dimensional matrix. In the simulated weight generation section, the number of RF links is denoted as... The number of data streams per subcarrier is denoted as When the number of RF links exceeds the number of users, additional RF link phases are used to generate the remaining constant-mode weights, which are then written into the cache along with the analog weights corresponding to the user angles. Figure 2 (b) shows the dimensionality reduction module call.

[0054] The elements of the transmitter's analog precoding matrix satisfy:

[0055]

[0056] The elements of the analog merging matrix at the receiving end satisfy:

[0057]

[0058] in, and These are the constant-mode amplitudes at the transmitting and receiving ends, respectively. and These are the phases of the corresponding matrix elements. and These represent the row index and column index of the matrix element, respectively. The imaginary unit, It is a natural constant. This constraint is used to characterize that the analog domain weights are mainly formed by phase control, adapting to the hardware implementation of RF phase shifter networks.

[0059] The second-stage CORDIC operation outputs two orthogonal I / Q signals, which, after amplitude compensation and fixed-point format alignment, form a complex array response vector that satisfies the constant modulus constraint. The complex array response vector adopts a 32-bit packed format, where the high 16 bits correspond to the imaginary part and the low 16 bits correspond to the real part, and is written to the RF vector buffer.

[0060] Furthermore, when the number of RF links at the transmitting end exceeds the number of users, the RF vector buffer path generates constant-mode phase vectors corresponding to the remaining RF links through a pseudo-random phase sequence, and writes them into the FIFO together with the array response vectors corresponding to the user angles. In this embodiment, the number of transmitting antennas is 144, the number of users is 2, the number of RF links is 3, and the third RF link is generated by a pseudo-random phase sequence.

[0061] like Figure 2 As shown in (b), the equivalent channel calculation module receives a multi-subcarrier channel matrix, a transmitter-simulated precoding matrix, and a receiver-simulated combining matrix. In the attached diagram, H_eq_real, H_eq_imag, and H_eq_valid represent the real part, imaginary part, and data validity identifier of the low-dimensional equivalent channel matrix, respectively. The meaning of H_eq in the attached diagram is the same as that of H_eff in the text regarding the low-dimensional equivalent channel matrix. In this embodiment, to facilitate RTL-level verification, each user channel matrix is ​​vertically concatenated according to its user index to form a multi-user channel matrix. After fixed-point conversion, it is written to the on-chip ROM as a COE file. The ROM simulates the channel data input interface according to the subcarrier index, continuously providing channel data to the equivalent channel calculation module. For the first... The low-dimensional equivalent channel matrix of the OFDM subcarriers is represented as follows:

[0062]

[0063] in, Indicates the first The multi-user channel matrix is ​​formed by vertically concatenating the user channel matrices corresponding to each subcarrier according to the user index. This represents the transmitter's analog precoding matrix. This represents the block diagonal receiver simulation merge matrix, composed of the simulated merge submatrices of each user receiver according to their user indices. Its conjugate transpose is used for receiver merging. Indicates the first The low-dimensional equivalent channel matrix corresponding to each subcarrier. For ease of hardware pipelining, the above calculation is broken down into two stages of complex matrix multiplication:

[0064]

[0065] in, The first-level transmitter dimensionality reduction result; the second-level calculation is the receiver's simulated merging matrix and... The product of, i.e. The above formula is used to illustrate the data path of the two matrix multipliers in this step. The actual hardware implementation adopts a structure that reuses the time-sharing ROM / RAM and the complex multiplication and accumulation unit.

[0066] Figure 2 In (b), the first-level matrix multiplier adopts a time-division multiplexed multiply-accumulate architecture with local parallelism and overall serial processing. In one implementation, 12 complex multiply-accumulate units are instantiated in parallel within the module, simultaneously calculating 12 independent elements in the target matrix; the 144 rows of data from the channel matrix and the 144 rows of data from the transmitter's analog precoding matrix are input serially and accumulated over 144 clock cycles to obtain the final result. Each complex multiply-accumulate unit executes according to the following complex multiplication relationship:

[0067]

[0068] In the formula, , , , Let these represent the real and imaginary parts of the two complex number inputs, respectively. The unit is the imaginary unit. The transmit counter and receive counter are used to control the input transmission and the multiplication output reception, respectively. The accumulator has a reserved protection bit to prevent overflow.

[0069] The second-stage matrix multiplier is structurally similar to the first-stage matrix multiplier, except that the input matrix dimension is reduced, and the inner accumulation length is shortened from the transmitter antenna dimension to the receiver RF combining dimension. In this embodiment, the number of transmitter RF links and receiver RF links are both 3, and the output low-dimensional equivalent channel matrix is ​​a 3×3 square matrix. This module sequentially outputs the low-dimensional equivalent channel matrices of multiple subcarriers. During output, the matrix element index indicates the position of the current output element in the matrix, the subcarrier index indicates the number of the currently processed subcarrier, and the data validity flag indicates that the current output data is valid. Figure 3 (a) The timing simulation results of the two-level equivalent channel dimensionality reduction calculation module are given, where H_eq_real and H_eq_imag represent the real and imaginary parts of the elements of the low-dimensional equivalent channel matrix, respectively, valid_out represents the output valid identifier, and element_idx and subcarrier_idx represent the matrix element position and the current subcarrier number, respectively.

[0070] like Figure 2 As shown in (c), the SVD decomposition module is generated by the MATLAB Simulink and HDL Coder toolchain, employing the Square Jacobi SVD HDL Optimized structure optimized for square matrix SVD. This module is based on the bilateral Jacobi algorithm, receiving the low-dimensional equivalent channel matrix and performing the following:

[0071]

[0072] in, It is a left singular matrix. It is a singular value matrix. This is a right singular matrix. In this embodiment, singular vectors corresponding to larger singular values, with a number equal to the number of data streams, are extracted to generate the transmitter digital precoding matrix and the receiver digital combining matrix:

[0073] The digital precoding matrix can be represented as:

[0074]

[0075] in, Indicates the number of data streams. Represents a matrix consisting of column vectors from the right singular matrix, each corresponding to a larger singular value, with a number equal to the number of data streams; receiver digital combining matrix. The output consists of column vectors extracted from the left singular matrix according to the same index, and is divided into user-defined digital merged submatrices according to user indices for separate output. The digital precoding matrix is ​​power normalized to ensure that the product of the transmitter's analog precoding matrix and the digital precoding matrix satisfies a preset transmit power constraint. This module is based on the bilateral Jacobi algorithm, iteratively rotating and updating multiple two-dimensional submatrices of the input matrix to gradually converge the matrix to a diagonal form, thus outputting the singular value matrix and the left and right singular matrices. Input and output data streams are flow-controlled via start, valid, and ready signals, and the decomposition of each subcarrier matrix is ​​triggered by matrix processing start control. Figure 3 (b) The timing simulation results of the Jacobi SVD decomposition module for the low-dimensional equivalent channel are given, where U_real, U_imag, S_real, S_imag, V_real and V_imag represent the real and imaginary parts of the elements of the left singular matrix, singular value matrix and right singular matrix, respectively, UsV_valid indicates that the output data is valid, and sub_cnt and matrix_cnt represent the matrix element count and matrix group count, respectively.

[0076] like Figure 2As shown in (a), after obtaining the current digital precoding matrix, the iterative update module corrects the transmitter analog precoding weights according to the alternating update idea. The transmitter analog precoding matrix generated in step S102 is used as the initial matrix for the 0th iteration. In each iteration, the module sequentially performs the conjugate transpose of the joint digital precoding matrix, matrix multiplication, small-scale matrix inversion, and constant modulus projection to generate the transmitter analog precoding matrix for the next iteration. When the preset number of analog weight update iterations has not been reached, the updated matrix is ​​fed back to the equivalent channel calculation module to recalculate the low-dimensional equivalent channel matrix and update the digital weights. After the preset number of analog weight update iterations has been reached, the current corresponding analog domain weights and digital domain weights are sent to the multi-subcarrier fixed-point output module.

[0077] Specifically, in the first In this iteration, an update matrix is ​​constructed based on the joint target all-digital precoding matrix and the joint digital precoding matrix obtained in this iteration. The transmitter simulation precoding matrix for the next iteration is obtained by inverting a small-scale matrix.

[0078]

[0079] in, Indicates the first The transmitter simulation precoding matrix obtained in the next iteration This represents the iteration count index, initialized to 0. This represents the initial transmitter simulation precoding matrix generated in step S102. This represents the joint target matrix formed by horizontally concatenating the target fully digital precoding matrices of each subcarrier in subcarrier index order. Indicates by the first In the next iteration, the current digital precoding matrices of each subcarrier are horizontally concatenated in the same order to form a joint digital precoding matrix. This indicates phase extraction or constant mode projection processing under constant mode constraints; when the preset number of simulation weight update iterations is reached, step S105 outputs the current corresponding simulation domain weights and digital domain weights.

[0080] The target all-digital precoding matrix is ​​written to the input buffer through the target matrix input interface and read by the iterative update module according to the matrix element index, serving as the target matrix for the iterative update of the transmitter's analog precoding matrix. During the iterative update process, FbbH represents... Fbb2 means F_inv means Fopt means Frf means .

[0081] like Figure 2As shown in (a), the matrix inversion module uses the adjoint matrix inversion method to invert a small-scale 3×3 complex matrix. For the iteration number index, the th The joint digital precoding matrix of the next iteration The input matrix is ​​formed by horizontally concatenating the current digital precoding matrices of each subcarrier in subcarrier index order. The relationship for calculating its inverse matrix is ​​as follows:

[0082]

[0083] in, Representation matrix The adjoint matrix, Representation matrix The determinant of; when At that time, the matrix inversion module sequentially completes the calculation of the cofactor, the generation of the adjoint matrix, and the calculation of the determinant. The division of each matrix element is performed by a single complex division unit in a time-division multiplexing manner to obtain the input matrix. The inverse matrix. In the matrix inversion module, A_real and A_imag represent the real and imaginary parts of the input matrix, respectively, and inv_real and inv_imag represent the real and imaginary parts of the inverse matrix, respectively.

[0084] In step S105, the multi-subcarrier fixed-point output module sequentially reads the digital precoding matrix and digital combining matrix according to the subcarrier index and matrix element index, and reuses the corresponding transmitter analog precoding weights and receiver analog combining weights within the same configuration period. Before output, each complex matrix element undergoes bit-width expansion, fixed-point format alignment, rounding, and saturation processing, and the data validity identifier is aligned with the matrix element position and subcarrier number using pipelined delay. For each subcarrier, the output module outputs the corresponding digital precoding matrix and digital combining matrix sequentially according to the matrix row and column order.

[0085] The FPGA computation output can be compared with the MATLAB floating-point simulation results. For the simulated precoding matrix, the mean square error of the array elements is maintained at ±30°. The mean square error of the real and imaginary parts of each subcarrier matrix element remains at the same level for both digital precoding and digital combining matrices. Order of magnitude. Furthermore, using a preset setpoint format and 10 Jacobi iterations, the Vivado hardware simulation output matrix is ​​close to the MATLAB floating-point theoretical value, with the absolute error of each matrix element controlled within... Within.

[0086] In one implementation, the FPGA hardware platform uses the Xilinx ZYNQ-7000 series XC7Z020CLG400-2 chip, with a system master clock constraint of 100 MHz. Synthesis results show that 88 18K BRAMs are used (36.7%), 97 DSP48Es (44.1%), 29420 FFs (27.7%), and 28796 LUTs (54.1%). Static timing analysis after placement and routing shows that the overall system WNS is +3.12 ns and TNS is 0 ns, which meets the timing convergence requirements under the current master clock constraint. In a fully serial cyclic configuration processing 128 subcarriers, the complete pipeline consumes approximately 129,000 clock cycles, corresponding to a single full-band precoding weight update time of approximately 1.29 ms at a master clock of 100 MHz.

[0087] The above embodiments are used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions or modifications made based on the technical concept of the present invention regarding module division, interface naming, buffer depth, fixed-point bit width, SVD implementation method, matrix inversion method, number of RF links, number of subcarriers, resource reuse method, or FPGA device model shall fall within the scope of protection of the present invention.

Claims

1. A low-complexity millimeter-wave multi-user hybrid precoding hardware implementation method based on equivalent channel dimensionality reduction, characterized in that, Includes the following steps: S101: Data Access and Buffer Scheduling. The system receives departure angle information, arrival angle information, multi-subcarrier channel data, and configuration control information for each user through the FPGA top-level interface. When iterative update processing is enabled, it receives the target all-digital precoding matrix for each subcarrier. User angle information and multi-subcarrier channel data are written into the input buffer. When iterative update processing is enabled, the target all-digital precoding matrix is ​​also written into the input buffer. Intermediate matrix data, SVD decomposition result data, and output weight data generated during processing are written into their respective buffers. Generate cache read / write addresses, data validity identifiers, and processing order control signals based on user indexes, subcarrier indexes, and matrix element indexes; S102: Construction of Constant Mode Simulation Weights. The departure angle and arrival angle information for each user are mapped to fixed-point phase steps. Phase accumulation is performed based on the antenna element index, and a complex array response vector satisfying constant mode constraints is generated through CORDIC operations. Transmitter simulation precoding weights and a transmitter simulation precoding matrix are constructed based on the complex array response vector corresponding to the departure angle. Receiver simulation combining weights and a receiver simulation combining matrix are constructed based on the complex array response vector corresponding to the arrival angle. When the number of RF links exceeds the number of users, extended constant mode weights corresponding to the remaining RF links are generated through a pseudo-random phase sequence and written into the RF vector buffer. S103: Equivalent Channel Dimensionality Reduction Calculation. Based on the transmitter's analog precoding matrix, receiver's analog combining matrix, and multi-subcarrier channel data, a two-level complex multiplication and accumulation structure with local parallelism and overall serialism is adopted. First, the product of each subcarrier channel matrix and the transmitter's analog precoding matrix is ​​calculated to obtain the transmitter's dimension-reduced intermediate matrix. Then, the product of the conjugate transpose of the receiver's analog combining matrix and the transmitter's dimension-reduced intermediate matrix is ​​calculated to obtain the low-dimensional equivalent channel matrix corresponding to each OFDM subcarrier. S104: SVD Digital Weight Extraction. The low-dimensional equivalent channel matrix corresponding to each OFDM subcarrier is input into the SVD pipeline decomposition module based on bilateral Jacobi rotation. Through multiple rounds of Jacobi rotation iteration, the left singular matrix, singular value matrix, and right singular matrix are obtained. According to the number of data streams, column vectors corresponding to larger singular values ​​are extracted from the right and left singular matrices respectively, forming a digital precoding matrix and a digital combining matrix. The digital precoding matrix is ​​power normalized so that the product of the transmitter's analog precoding matrix and the digital precoding matrix satisfies a preset transmit power constraint. The digital precoding matrix and the digital combining matrix are then written into the decomposition result cache. S105: Multi-subcarrier fixed-point output. The digital precoding matrix, digital combining matrix, transmitter analog precoding weights, and receiver analog combining weights are subjected to bit-width expansion, fixed-point format alignment, rounding, truncation, and saturation processing, and output sequentially according to the subcarrier index and matrix element index. Specifically, the output data, data validity flag, subcarrier index flag, and matrix element index flag are pipelined with delay alignment to ensure that the output data is synchronized with each flag signal.

2. The method according to claim 1, characterized in that, In step S101, the FPGA top-level interface includes an angle information input interface, a channel data input interface, a target matrix input interface, a configuration control interface, and a weight output interface. The target matrix input interface is used to receive the full digital precoding matrix of each subcarrier target when iterative update processing is enabled. The configuration control information includes at least one of the following: number of users, number of radio frequency links, number of data streams, number of subcarriers, equivalent channel matrix dimension, number of SVD iterations, number of simulated weight update iterations, and fixed-point output bit width. The buffer includes an input buffer, an intermediate matrix buffer, a decomposition result buffer, a radio frequency vector buffer, and an output buffer. The input buffer is also used to store the full digital precoding matrix of each subcarrier target when iterative update processing is enabled.

3. The method according to claim 1, characterized in that, In step S102, the matrix elements of the transmitter's analog precoding matrix satisfy the following constant modulus constraint: The elements of the analog merging matrix at the receiving end satisfy: in, and These are the constant-mode amplitudes at the transmitting and receiving ends, respectively. and These are the phases of the corresponding matrix elements. and These represent the row index and column index of the matrix element, respectively. The imaginary unit, It is a natural constant. This constraint is used to characterize that the analog domain weights are mainly formed by phase control, adapting to the hardware implementation of RF phase shifter networks.

4. The method according to claim 1, characterized in that, In step S103, for the first OFDM subcarriers, the low-dimensional equivalent channel matrix satisfies: in, Indicates the first The low-dimensional equivalent channel matrix corresponding to each subcarrier. Indicates the first The multi-user channel matrix is ​​formed by vertically concatenating the user channel matrices corresponding to each subcarrier according to the user index. This represents the transmitter's analog precoding matrix. This represents a block diagonal receiver-side simulated merging matrix composed of simulated merging submatrices from each user receiver, indexed by user. Its conjugate transpose is used for receiver-side merging. The two-stage complex multiplication-accumulation structure includes a first-stage transmitter-side dimensionality-reducing matrix multiplier and a second-stage receiver-side merging matrix multiplier. in, The intermediate matrix after dimensionality reduction at the transmitting end is represented. Under the implementation configuration of 144 antennas at the transmitting end, 4 antennas at the receiving end, and 3 radio frequency links at each end, the two-stage complex multiply-accumulate structure first performs equivalent dimensionality reduction at the transmitting end, and then performs equivalent dimensionality reduction at the receiving end, reducing the 4-row 144-column multi-user channel matrix to a 4-row 3-column intermediate matrix and a 3-row 3-column low-dimensional equivalent channel matrix in sequence. The first-stage transmitting end dimensionality reduction matrix multiplier and the second-stage receiving end merging matrix multiplier both adopt a time-division multiplexing structure composed of complex multiplication units and complex accumulation units.

5. The method according to claim 4, characterized in that, The first-stage transmitter dimension reduction matrix multiplier instantiates 12 complex multiplication and accumulation units in parallel to compute the 12 independent elements of the intermediate matrix in parallel, and performs time-division accumulation of the input matrix elements along 144 antenna dimensions; each complex multiplication satisfies: in, and These are the real and imaginary parts of the first complex number input, respectively. and These are the real and imaginary parts of the second complex input, respectively, where j is the imaginary unit. The accumulated result is written to the intermediate matrix buffer after being accumulated with guard bits, saturated truncation bits, and aligned with fixed-point format.

6. The method according to claim 1, characterized in that, In step S104, for the first Singular value decomposition is performed on the low-dimensional equivalent channel matrix corresponding to each OFDM subcarrier: And based on the number of data streams, extract the corresponding column vectors of the right singular matrix to generate the digital precoding matrix: in, It is a left singular matrix. It is a singular value matrix. It is a right singular matrix. Indicates the first The digital precoding matrix corresponding to each subcarrier The number of data streams is represented by the digital precoding matrix, which consists of column vectors in the right singular matrix corresponding to the larger singular values, with the number of vectors equal to the number of data streams. The digital merging matrix... Left singular matrix The matrix is ​​composed of column vectors extracted according to the same index, and divided into user-defined digital merged sub-matrices according to user indexes for separate output. The digital precoding matrix is ​​power normalized to ensure that the product of the transmitter's analog precoding matrix and the digital precoding matrix satisfies a preset transmit power constraint. The SVD pipelined decomposition module has FIFO buffers at both the input and output ends, and completes the ordered scheduling of the multi-subcarrier matrix through start signals, valid signals, and ready signals.

7. The method according to claim 1, characterized in that, After step S104, an iterative update process is also included: the transmitter simulation precoding matrix generated in step S102 is used as the initial transmitter simulation precoding matrix for the 0th iteration, and the initial value of the iteration index is set to 0; In each iteration, steps S103 and S104 are executed based on the current transmitter analog precoding matrix to obtain the current digital precoding matrix. When the preset number of analog weight update iterations has not been reached, an update matrix is ​​constructed based on the target all-digital precoding matrix and the current digital precoding matrix. Conjugate transpose, matrix multiplication, small-scale matrix inversion, and constant modulus projection are executed sequentially to obtain the transmitter analog precoding matrix for the next iteration, and the process returns to step S103. After the preset number of analog weight update iterations has been reached, step S105 is executed.

8. The method according to claim 7, characterized in that, The inversion of the small-scale matrix is ​​performed using the adjoint matrix inversion method; whereby... For the iteration number index, the th The joint digital precoding matrix of the next iteration The input matrix is ​​formed by horizontally concatenating the current digital precoding matrices of each subcarrier in subcarrier index order. Its inverse matrix satisfies: in, Representation matrix The adjoint matrix, Representation matrix The determinant of; when At that time, the adjoint matrix inversion method sequentially performs algebraic cofactor calculation, adjoint matrix generation, determinant calculation, and complex division processing, with the complex division processing being completed by a single division operation unit in time-division multiplexing.

9. The method according to claim 1, characterized in that, The fixed-point processing includes uniform fixed-point format representation, bit width expansion, guard bit accumulation, rounding, truncation and saturation processing of angle data, complex matrix data, SVD decomposition output data and output weight data; the output weight data adopts a 32-bit complex number packing format, in which the high 16 bits are the imaginary part and the low 16 bits are the real part.

10. A low-complexity millimeter-wave multi-user hybrid precoding hardware implementation system based on equivalent channel dimensionality reduction, characterized in that, The system includes an interface and buffer scheduling module, a constant modulus analog weight generation module, an equivalent channel matrix calculation module, an SVD digital weight generation module, an iterative update module, and a multi-subcarrier fixed-point output module. The interface and buffer scheduling module provides each module with angle information, channel data, a target all-digital precoding matrix, configuration control information, subcarrier indexes, matrix element indexes, and data validity identifiers. The constant modulus analog weight generation module outputs the transmitter's analog precoding weights and the receiver's analog combining weights. The equivalent channel matrix calculation module outputs a low-dimensional equivalent channel matrix based on the analog domain weights and multi-subcarrier channel data. The SVD digital weight generation module outputs a power-normalized digital precoding matrix and a digital combining matrix. The iterative update module is used to update the transmitter's analog precoding matrix based on the target all-digital precoding matrix and the current digital precoding matrix, and to feed the update result back to the equivalent channel matrix calculation module before the preset number of analog weight update iterations is reached. The multi-subcarrier fixed-point output module is used to perform fixed-point conversion on the analog domain weights and digital domain weights and output them in subcarrier order after reaching a preset number of analog weight update iterations.