A hierarchical cyclic left shift-based FFT data rearrangement method based on FPGA

By using a hierarchical cyclic left shift address transformation method, the problems of storage resource and bandwidth consumption in existing FFT data rearrangement technology are solved, realizing efficient and low-resource-consumption FFT data rearrangement, which is suitable for various FFT architectures and high-frequency scenarios.

CN123115014APending Publication Date: 2026-11-17JIANGSU YU SPACE TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611259362.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing FFT data rearrangement technology has shortcomings in terms of storage resource consumption, memory access bandwidth, computation latency and architectural versatility, and cannot simultaneously meet these requirements, especially in high-point-count scenarios where timing constraints are difficult to satisfy.

Method used

It adopts a hierarchical circular left shift method, which avoids data movement by performing a 1-bit circular left shift on the data address after each butterfly operation. It achieves multi-level data reordering only through address logic transformation and supports FFT with any number of integer powers.

Benefits of technology

Significantly reduces storage bandwidth usage, saves FPGA on-chip resources, improves computing performance, adapts to multiple FFT architectures, meets high-frequency timing requirements, reduces overall computing latency, and simplifies software post-processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115014A_ABST
    Figure CN123115014A_ABST
Patent Text Reader

Abstract

This invention relates to the field of FPGA hardware acceleration technology for digital signal processing, specifically to an FFT data rearrangement method based on FPGA hierarchical cyclic left shift. The original sequential data addresses are natural binary sequences. After a single-stage butterfly operation, the target addresses required for data pairing are equivalent to the sequence after cyclically shifting the original addresses left by 1 bit. Traditional rearrangement schemes rely on data read / write copying to complete the sorting, with multi-level operations continuously consuming BRAM and DDR bus bandwidth. This invention does not move the data entities stored internally in the BRAM throughout the process, only modifying the hardware addressing addresses. There are no additional data read / write operations, reducing storage bandwidth consumption to less than 10% of the original scheme. In multi-channel parallel FFT operation scenarios, bandwidth saturation issues do not occur, significantly improving the system's maximum data throughput rate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of FPGA hardware acceleration technology for digital signal processing, specifically to an FFT data rearrangement method based on FPGA hierarchical cyclic left shift. Background Technology

[0002] Radix-2 FFT Traditional Data Rearrangement Principles and Hardware Implementation

[0003] For points For radix-2 decimation-time FFT, the original input time-domain signal needs to be sorted in bit-reverse order before being fed into the first-stage butterfly array. After each butterfly iteration, the intermediate output data also suffers from local disorder, requiring the next-stage butterfly unit to read paired, ordered data to perform complex number addition and subtraction of twitch factors. Mainstream hardware reordering implementations in the industry fall into three categories:

[0004] The first type is the dual-buffered RAM data exchange and rearrangement scheme. This scheme configures two on-chip BRAMs of equal capacity as ping-pong caches. The output data of the first-level butterfly operation is written to cache A. The hardware moves the out-of-order data to cache B in paired order through address traversal and data copying. The next-level butterfly operation reads ordered data from cache B. After the operation is completed, the two caches switch read and write roles, and the multi-level rearrangement is completed in a loop. This scheme has simple logic and easy timing convergence, and is the default implementation method of commercial FPGA FFT IP cores. However, it has significant drawbacks: First, it requires twice the capacity of on-chip storage resources, and the BRAM resource consumption directly doubles in high-point-count FFT scenarios; second, multi-level data movement continuously occupies storage read and write bandwidth. When the system performs multi-channel parallel FFT operations simultaneously, the storage controller bandwidth is saturated, directly limiting the maximum throughput rate of the system; third, the data copying and exchange process introduces additional read and write latency, increasing the overall FFT operation time.

[0005] The second type is the single-RAM co-address bit-reverse addressing scheme. This scheme uses only a single BRAM block to store data, does not set up double buffering, and directly reads the reverse address to complete the pairing by pre-compiling a bit-reverse address table. This scheme saves half of the storage resources, but it requires the pre-generation and storage of a full reverse address lookup table. The larger the number of points, the more register and ROM resources the lookup table occupies. At the same time, each data read requires a table lookup to convert the address, increasing the combinational logic path delay. In high-frequency FPGA designs, timing violations are very likely to occur, requiring the insertion of additional pipeline registers, further consuming logic resources.

[0006] The third category is the off-chip DDR cache rearrangement scheme. For FFTs with 262,144 points or more, the on-chip BRAM capacity is insufficient. Traditional solutions write intermediate data to external DDR memory, relying on DDR to complete large-scale data sorting and movement. DDR read and write operations have access latency of tens of clock cycles, and the external bus bandwidth is limited. High-point FFT multi-level rearrangement will cause a large amount of bus congestion, which is completely unable to meet the microsecond-level real-time processing requirements of radar and communication systems.

[0007] Limitations of existing optimization schemes

[0008] In recent years, the industry has proposed a number of improvement techniques to address the overhead of FFT rearrangement, such as block-based local rearrangement, rotation factor and address table reuse, and multi-port BRAM parallel addressing. However, all of the above optimization methods have underlying shortcomings:

[0009] All existing solutions cannot completely avoid data entity migration operations; they can only reduce the number of migrations and the size of the migrated data blocks, but storage bandwidth overhead cannot be eliminated.

[0010] Block rearrangement technology is only applicable to FFT with fixed segment length, and cannot be compatible with architectures with continuously variable number of points from 128 to 262144, so its versatility is poor.

[0011] Multi-port storage hardware is expensive, and on-chip dual-port and quad-port BRAM resources on FPGAs are scarce. Large-scale deployment will significantly increase the cost of chip selection.

[0012] The bit-reverse lookup table logic grows exponentially with the number of FFT points. In scenarios with extremely large number of points, the combinational logic level is too high, making it difficult to meet the timing constraints of high frequencies of 200MHz and above.

[0013] In the applicant's previously developed 128-point hardware-reused variable-point FFT architecture, a 7-level butterfly operation was used as the upper limit for a single block segment, dividing the 18-level 262,144-point FFT into three time-division multiplexed operations. If a traditional double-buffered data movement and rearrangement scheme is used, each sub-operation requires multiple data exchanges. After the three segmented operations are superimposed, the storage bandwidth overhead increases by 3 times, the overall operation latency of the 262,144-point FFT increases by more than 40%, and the BRAM resource utilization rate increases from 15% to 30%, significantly weakening the resource-saving advantage brought by hardware multiplexing.

[0014] In summary, existing FFT data rearrangement technology cannot simultaneously address the four key performance indicators of storage resource consumption, memory access bandwidth, computation latency, and architectural versatility. The industry urgently needs a lightweight hardware implementation solution that can complete multi-level data reordering without data migration and solely through address logic transformation. Summary of the Invention

[0015] To address the shortcomings of existing FPGA FFT data rearrangement schemes, which rely on large-scale data movement, incur high storage bandwidth overhead, consume significant on-chip storage resources, face stringent timing constraints, and cannot adapt to variable-point block-multiplexed FFT architectures, this invention provides an FPGA-based hierarchical cyclic left-shift FFT data rearrangement method, supporting hardware structure, computer-readable storage medium, and FPGA electronic device. The core idea of ​​this invention is to change only the storage address without moving the stored data. After each butterfly operation, a 1-bit cyclic left shift is performed on the data-bound binary address, and the next level directly uses the shifted address for addressing, eliminating bandwidth loss caused by data movement at its source. After all operations are completed, a single batch address compensation shift restores the natural sequence, achieving ordered data reading across all scenarios using radix-2 FFT with extremely low logic resource overhead, balancing hardware resources, storage bandwidth, and real-time computing performance.

[0016] Core technical principle of this invention

[0017] for A point-radix-2 decimation-time FFT is used, with a data storage address width of m bits. The original sequential data addresses are natural binary sequences. After one butterfly operation, the target addresses required for data pairing are equivalent to the sequence after the original addresses are cyclically shifted left by 1 bit. In each iteration of the butterfly operation, the address is synchronously shifted left by 1 bit. After m levels of operations, the address sequence has undergone a cumulative m cyclic left shifts, resulting in a fixed offset and disordered data. At this point, only one m cyclic left shift compensation correction is needed to restore the addresses to their natural order, outputting standard ordered spectral data.

[0018] This invention deeply binds address shifting operations with butterfly operations in a hierarchical manner, updates the addressing address in real time in a hierarchical manner, and keeps the data in a fixed storage location inside the BRAM throughout the process. There is no data reading, copying, or writing behavior, which completely eliminates the need for additional hardware resources such as double buffering, lookup tables, and external DDR cache.

[0019] Complete technical solution of the present invention

[0020] Solution 1: An FFT data rearrangement method based on FPGA hierarchical cyclic left shift

[0021] This method runs on FPGA programmable logic hardware and is compatible with radix-2 FFTs of any integer power of 2. The complete execution steps are as follows:

[0022] Step S1: Obtain the intermediate FFT result data cached in the on-chip storage unit.

[0023] After the FPGA butterfly arithmetic array completes a single-level complex number butterfly calculation, it writes the intermediate complex number result to the on-chip BRAM memory cell. The hardware logic synchronously latches the original binary memory address corresponding to each set of complex number data, establishing a one-to-one binding mapping relationship between data and address, and outputs the bound address information to the address translation module. This step only performs a single data write operation; there is no data exchange or copying process.

[0024] Step S2: After a single-level butterfly operation, perform a 1-bit circular left shift transformation of the storage address.

[0025] After each complete butterfly operation, the address translation module synchronously and in parallel performs a 1-bit circular left shift operation on all bound memory addresses. The underlying execution logic is divided into three steps:

[0026] (1) Cache the highest bit: Extract the highest bit of the current binary address and store it in a 1-bit shift buffer register;

[0027] (2) Shift the low-order bits left: shift the remaining m-1 bits of the address (excluding the most significant bit) to the left by 1 bit, leaving the least significant bit empty.

[0028] (3) Cyclic filling of low bits: The original highest bit in the cache register is filled into the lowest bit of the address to generate the new address after the current level transformation.

[0029] All three steps described above are implemented using combinational logic, completing all address translations in parallel within a single clock cycle without pipeline delay. After the translation is complete, the original data remains stored in BRAM in the same location; only the address index used for hardware addressing is updated.

[0030] Step S3: The next-level butterfly operation uses the transformed address to directly address and read the ordered paired data.

[0031] When the next-level butterfly array initiates the data reading process, the storage controller directly receives the address after a circular left shift output by the address translation module and reads the intermediate data stored in the BRAM based on this address. Since the address has undergone hierarchical circular left shift permutation, the read data naturally satisfies the pairing order required by this level of butterfly operation, eliminating the need for additional sorting, cache switching, or data entity relocation.

[0032] Step S4: After all butterfly operations are completed, perform a supplementary circular left shift to correct the address sequence.

[0033] After all iterations of the m-level complete butterfly operation are completed, the hardware reads the pre-stored total number of butterfly levels m, calculates the total number of address compensation shifts m times, and performs supplementary cyclic left shift operations on the storage addresses corresponding to all final results in batches. After compensation shift correction, the address sequence is restored to its natural binary order. At this point, the data in the BRAM can be directly read to output the standard FFT spectrum result without disorder, without the need for secondary sorting processing by the backend software.

[0034] Furthermore, this method has extremely strong architectural compatibility:

[0035] Points adaptation: Supported

[0036] 128≤N≤262144

[0037] all The number of points can be directly matched with the time-division multiplexed block variable point FFT architecture. Each segment operation independently performs hierarchical address circular left shift, and the address offset is temporarily stored after the segment operation is completed. After all segment operations are completed, a total compensation shift is performed uniformly.

[0038] Architecture Adaptation: Compatible with three mainstream FPGA hardware architectures: fully parallel pipelined FFT, single butterfly unit serial iterative FFT, and block time-division multiplexing FFT. Only the butterfly output and memory addressing interfaces need to be connected, without changing the original butterfly operation pipeline timing.

[0039] Option 2: A hardware structure for FFT data rearrangement based on hierarchical cyclic left shift

[0040] This hardware architecture is entirely built using FPGA programmable logic units, which can be independently embedded as IP cores in any FFT operation architecture. It includes three core functional modules, which are interconnected using synchronous timing signals:

[0041] Get Module

[0042] The module's input is directly connected to the complex output port of the butterfly arithmetic array, and internally integrates a multi-channel data buffer register and an address latch register. The real and imaginary complex data output from each stage of the butterfly arithmetic operation are stored in the buffer register. The original binary address allocated when synchronously latching the data and writing it to the BRAM is used to bind the data and address into a set of signals, which are then output in parallel to the address translation module. The module has built-in synchronous timing control logic, ensuring complete synchronization with the butterfly arithmetic clock and eliminating cross-clock domain interaction overhead.

[0043] Address translation module

[0044] The address translation module is the core innovative hardware unit of this invention, which is divided into two parts: parallel shift sub-logic and compensated shift control sub-logic.

[0045] (1) Parallel shift sub-logic: It consists of a multiplexer and a shift register array. It receives the original address output by the acquisition module, completes the 1-bit cyclic left shift of the entire address in parallel in a single clock cycle, and outputs the transformed address to the read address module.

[0046] (2) Compensation shift control sub-logic: Built-in series register is used to store the total number of butterfly series m of FFT; when the hardware detects the flag signal that all butterfly operations are completed, it automatically generates m cyclic left shift control pulses, performs compensation shift correction on the final storage address in batches, and outputs the natural order address sequence.

[0047] This module consumes only a small amount of registers and combinational logic resources, and does not require BRAM or ROM to store address tables, resulting in extremely low hardware resource overhead.

[0048] Read addressing module

[0049] The addressing module bidirectionally interfaces with the address translation module and the on-chip BRAM memory controller, while simultaneously transmitting ordered complex data to the next-level butterfly arithmetic array. The module receives the translated address signal and directly forwards it to the BRAM read / write controller to complete the addressing and read operation. The ordered paired data output from the BRAM does not require buffering and flows directly into the input port of the next-level butterfly unit, establishing a zero-latency data transmission link. The module does not have an internal data exchange buffer, eliminating data movement operations at the hardware level.

[0050] Option 3: Computer-readable storage medium

[0051] This storage medium can be a hardware carrier such as non-volatile memory, Flash, or hard disk. The hardware description language program, control instruction set, and logic configuration code that implement the rearrangement method of this invention are embedded within the medium. When the program and instructions in the storage medium are loaded and parsed by the FPGA's supporting processor and hardware synthesis tools, the above-mentioned hierarchical cyclic left shift rearrangement hardware structure can be synthesized within the FPGA to fully execute the hierarchical address shift and data rearrangement without data movement.

[0052] Option 4: An FPGA electronic device

[0053] This electronic device consists of a Xilinx / Intel series FPGA programmable logic chip, an on-chip BRAM memory array, an external DDR4 memory, a high-speed ADC / DAC data interface, and a power and clock management module. The internal memory stores FFT operation control instructions and address circular left shift logic configuration code. After the FPGA chip loads the instructions, it generates a butterfly-shaped operation array and rearranges the hardware structure according to this invention. The external high-speed interface completes time-domain signal input and spectrum result output, enabling independent real-time variable-point FFT acceleration operations in radar and communication scenarios.

[0054] Beneficial effects of the invention

[0055] Completely eliminate storage bandwidth bottlenecks caused by data migration

[0056] Traditional rearrangement schemes rely on data read, write, and copying to complete the sorting, and multi-level operations continuously occupy the bandwidth of BRAM and DDR bus. This invention does not move the data entities stored inside BRAM throughout the process, only modifies the hardware addressing address, and has no additional data read and write operations. The storage bandwidth usage is reduced to less than 10% of the original scheme. There will be no bandwidth saturation problem in multi-way parallel FFT operation scenarios, which greatly improves the maximum data throughput rate of the system.

[0057] Significantly saves on-chip memory and logic resources in FPGAs

[0058] This invention eliminates the need for dual-buffered ping-pong BRAM and large-capacity bit-reversed address lookup ROM, relying solely on a small number of shift registers and multiplexers to implement address translation logic. In the same 262,144-point FFT scenario, compared to traditional dual-buffered rearrangement schemes, on-chip BRAM resource usage is reduced by 52%, and address lookup table-related logic resource consumption is reduced by 90%. With the accompanying time-division multiplexed block FFT architecture, the overall chip resource utilization rate can be controlled within 15%, and there are no idle hardware resources for low-point operations.

[0059] Excellent timing performance; high-frequency FPGA designs are easy to converge.

[0060] The address circular left shift operation is implemented using a single-level combinational logic, eliminating the long data paths caused by multi-level table lookups and multi-level caching, resulting in minimal combinational logic latency. At a standard 200MHz FPGA operating frequency, timing constraints can be met without inserting pipelined registers, eliminating the risk of setup and hold time violations, thus adapting to the high-frequency hardware design requirements of radar and communication systems.

[0061] It is highly versatile and adaptable to a full range of variable point counts and various FFT architectures.

[0062] This invention uniformly supports FFTs with all integer powers of 2 from 128 to 262144, and can be seamlessly embedded into three mainstream FFT hardware architectures: fully parallel, serial pipelined, and time-division multiplexed block. It is highly compatible with the applicant's previously proposed 128-point hardware multiplexed variable-point FFT architecture. The segmented sub-operations are independently executed with hierarchical address shifting, and unified address compensation is performed at the end of each segment, forming a complete low-resource, high-speed, large-point FFT hardware solution.

[0063] Reduce overall system latency and improve real-time processing capabilities

[0064] By eliminating the extra clock cycle overhead caused by multi-level data movement, cache swapping, and table lookup, the overall FFT computation time is reduced by more than 35% at the same main frequency. In actual tests, on a Xilinx XC7K325T FPGA with a main frequency of 200MHz, the total computation time of a complete 262,144-point FFT was reduced from 34.6ms in the traditional solution to 22.45ms, which fully meets the millisecond-level real-time spectrum analysis target of radar.

[0065] Reduce post-processing overhead of upper-layer software

[0066] After all butterfly operations are completed, the natural order spectrum data is directly output through hardware address compensation shift, eliminating the need for backend CPUs and embedded processors to execute bit reordering software algorithms, reducing data interaction and software computation latency between chips, and simplifying the overall system hardware and software design.

[0067] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0068] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of an 8-point FFT butterfly pattern of the present invention;

[0070] Figure 2 This is a schematic diagram of the 8-point FFT cyclic left shift address of the present invention;

[0071] Figure 3 This is a schematic diagram of the 8-point FFT butterfly pattern of the present invention. Detailed Implementation

[0072] according to Figures 1-3 As shown, the present invention provides two sets of typical embodiments: a basic verification embodiment of small-scale 8-point fixed-point FFT and an engineering implementation embodiment of 128~262144 variable-point block multiplexing FFT. The hardware platform uniformly adopts Xilinx XC7K325T FPGA chip, with a system global clock frequency of 200MHz and an on-chip BRAM bit width configuration of 256bit, supporting parallel reading and writing of complex data.

[0073] Example 1: Verification of Hierarchical Address Cyclic Left Shift Rearrangement using 8-point Radix-2 Time-decimation FFT

[0074] 8-point FFT satisfies The total number of butterfly stages m=3, the address width is 3 bits, and the original natural address sequence is: 000, 001, 010, 011, 100, 101, 110, 111. Refer to the appendix for the complete execution flow. Figure 3 :

[0075] Step 1: First-level butterfly operation execution and initial left shift of address.

[0076] The raw time-domain data is written to the BRAM at natural addresses and then fed into the first-level butterfly array to complete complex number operations. After the operations are completed, the address transformation module synchronously performs a 1-bit circular left shift on all 3-bit addresses.

[0077] Original address 000 → Shift left by 1 bit → 000; Original address 001 → Shift left by 1 bit → 010; Original address 010 → Shift left by 1 bit → 100; Original address 011 → Shift left by 1 bit → 110; Original address 100 → Shift left by 1 bit → 001; Original address 101 → Shift left by 1 bit → 011; Original address 110 → Shift left by 1 bit → 101; Original address 111 → Shift left by 1 bit → 111.

[0078] The next-level butterfly array directly uses the shifted address to read BRAM data. The read data is automatically matched with the pairing order of the second-level butterfly array, and the data storage location remains unchanged.

[0079] Step 2: Second-level butterfly operation and double left shift of address

[0080] After the second-level butterfly operation iteration is completed, the address transformation module performs a 1-bit circular left shift on the current storage address to generate a new address for the third-level operation to read. There is no data copying or swapping operation throughout the process.

[0081] Step 3: After the third-level butterfly operation is completed, address compensation correction is performed.

[0082] After all three levels of butterfly operations are completed, the hardware reads the total number of levels m=3 and performs three round-shift compensation operations on all memory addresses to correct the sequence offset caused by the multi-level shifts. After the correction is completed, the BRAM memory addresses are restored to the natural binary order of 000~111, and the ordered 8-point FFT spectrum result is directly read and output.

[0083] Performance comparison: This embodiment adopts a hierarchical cyclic left shift rearrangement scheme, which consumes only 12 register resources; the traditional double-buffered rearrangement scheme requires two 8-bit capacity BRAMs, which consumes an additional 128 bits of storage resources, and adds 8 data shift read and write operations, increasing the overall operation latency by 12 clock cycles.

[0084] Example 2: Adapting to a time-division multiplexing block FFT architecture with 128~262144 points

[0085] This embodiment is equipped with the prior patented 128-point hardware multiplexing variable-point FFT computing architecture. Taking a 262,144-point FFT as an example... The total number of butterfly levels m=18, which is divided into three segments based on the upper limit of level 7: level 7, level 7, and level 4. The segmented rearrangement execution flow is as follows:

[0086] First segment operation (7-level butterfly): After each level operation is completed, a 1-bit address is cyclically shifted left. The 7-level operation accumulates 7 address shifts. After the operation of this segment is completed, the address offset of this segment is temporarily stored, and the data is kept in BRAM without being moved.

[0087] The second stage of operation (7-level butterfly): reuses the same set of 128-point FFT basic operation units, and each level inside continues to perform a 1-bit circular left shift of the address, with a total of 14 shifts.

[0088] The third operation (level 4 butterfly, with operation truncation logic enabled): only level 4 butterfly operations are executed, with each level's synchronous address shifted left in a loop, for a total of 18 shifts; all three operations are completed.

[0089] Global address compensation correction: Based on the total number of levels m=18, perform 18 supplementary cyclic left shifts on all memory addresses to cancel the sequence offset caused by the three-stage hierarchical shifts in one go, and output 262144 points of natural order FFT spectrum data.

[0090] Hardware performance metrics: At a 200MHz clock frequency, the entire block FFT + hierarchical cyclic left shift rearrangement architecture completes 262,144-point FFT in 22.45ms; the FPGA on-chip BRAM resource utilization rate is 15%, and the memory bus bandwidth utilization is only 8% of the traditional rearrangement scheme; there is no external DDR memory interaction, no external bus latency, and the requirements of airborne radar real-time signal processing are met.

[0091] Example 3: 1024-point FFT adaptation test in a fully parallel pipeline

[0092] The 1024-point FFT has a total of 10 stages (m=10), and the rearranged hardware structure of this invention is embedded in a fully parallel 10-stage pipelined butterfly architecture. After each pipeline output, the address is synchronously updated and cyclically shifted left, ensuring uninterrupted pipeline operation without buffering or pauses. The pipeline output hardware automatically performs 10 address compensation shifts, continuously outputting an ordered spectral data stream. The system data throughput reaches 200 MSPS, a 32% improvement compared to traditional table lookup reverse addressing schemes, with no timing violations.

[0093] Supplementary Explanation of Hardware Logic Implementation

[0094] The address translation module of this invention is entirely written in VerilogHDL hardware description language, with no complex timing logic. It only uses continuous assignment statements to implement the combinational logic of circular left shift, with a code size of less than 100 lines and extremely low logic resource consumption after synthesis. For addresses with arbitrary bit width m, it can automatically adapt to different FFT points through parameterized design. The IP core can be parameterized to configure the address bit width and the total number of butterfly stages, making it highly versatile and directly integrateable into various FPGA signal processing projects.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for FFT data rearrangement based on FPGA hierarchical cyclic left shift, characterized in that, The method includes: Acquire the intermediate FFT result data to be processed. The intermediate FFT result data is completely stored in the FPGA on-chip memory unit BRAM. Each set of butterfly operation output data is bound to a unique binary storage address. After the hardware logic completes any level of radix-2 butterfly operation, a 1-bit circular left shift operation is synchronously performed on all storage addresses corresponding to the intermediate results of the current batch. When the next level of butterfly operation is started and the input paired data is read, the storage controller directly calls the address after the circular left shift transformation to complete the addressing and reading. There is no need to perform read, write, move, copy, or exchange operations on the data entities inside the on-chip storage unit throughout the entire process. After all butterfly operation series iterations are completed, the address compensation shift number is calculated based on the total number of butterfly series m of FFT. A supplementary cyclic left shift operation is performed on the storage address of the final output data to offset the sequence offset caused by multi-level address transformation, so that the output spectrum data is restored to its natural order.

2. The FFT data rearrangement method based on FPGA hierarchical cyclic left shift according to claim 1, characterized in that, Perform a 1-bit circular left shift operation on the storage address of the intermediate result, including: Extract the highest bit of the binary address as a cyclic shift buffer bit; shift the remaining low bits of the address to the left by 1 bit; fill the highest bit of the buffer into the lowest bit of the address, completing a single 1-bit address cyclic replacement, and completing all address synchronization transformations within a single clock cycle.

3. The FFT data rearrangement method based on FPGA hierarchical cyclic left shift according to claim 1, characterized in that, The number of bits in the supplementary circular left shift operation corresponds one-to-one with the total number of butterfly levels m in the FFT; Let the number of FFT transformation points be N, satisfying N=2m, and the total number of compensation shifts be m. Only one batch address shift correction is performed after all butterfly operations are completed, without increasing the memory access control overhead of intermediate hierarchical operations.

4. The FFT data rearrangement method based on FPGA hierarchical cyclic left shift according to any one of claims 1 to 3, characterized in that, The method is applicable to radix-2 time decimation FFT and radix-2 frequency decimation FFT operations with any number of integer powers of 2 satisfying N=2m, and supports FFT block reuse operations with a large span of variable number of points from 128 to 262144.

5. The FFT data rearrangement method based on FPGA hierarchical cyclic left shift according to any one of claims 1 to 3, characterized in that, The method can be seamlessly embedded into any of the following architectures: fully parallel pipelined FFT architecture, single butterfly unit serial iterative FFT architecture, and time-division multiplexing-based block variable-point FFT architecture, without modifying the original butterfly operation hardware pipeline timing.

6. A hardware structure for FFT data rearrangement based on hierarchical cyclic left shift, deployed inside an FPGA chip, characterized in that, The structure includes: The acquisition module is used to connect to the output port of the butterfly operation array, cache the FFT intermediate result data generated by each stage of butterfly operation, synchronously latch the original binary storage address corresponding to each set of data, and output the data to the address transformation module after binding the data with the address. The address transformation module is directly connected to the acquisition module and has built-in parallel cyclic left shift logic and final-stage compensation shift control logic. It is used to synchronously perform a 1-bit cyclic left shift operation on all intermediate result storage addresses after each butterfly operation is completed. After all butterfly operations are completed, a compensation shift control signal is generated according to the total number of butterfly levels to perform supplementary cyclic left shifts on the final storage addresses in batches. The read addressing module connects to the output of the address translation module at one end and to the on-chip BRAM memory controller and the input port of the next-level butterfly arithmetic array at the other end. It is used to directly forward the translated address to the memory controller when reading input data in the next-level butterfly arithmetic operation, and automatically read data that conforms to the butterfly pairing order, without configuring a data exchange buffer or executing data transfer instructions.

7. The hardware structure for FFT data rearrangement based on hierarchical cyclic left shift according to claim 6, characterized in that, The address translation module has a built-in multiplexer and shift register array, which can complete the 1-bit circular left shift operation of all addresses in parallel in a single clock cycle. The address transformation module is independently configured with a shift bit register to store the total number of butterfly stages m. After the operation is completed, it automatically generates m cyclic left shift control instructions to complete the address sequence compensation and correction.

8. The hardware structure for FFT data rearrangement based on hierarchical cyclic left shift according to claim 7, characterized in that, The underlying logic of the address transformation module for performing a single 1-bit cyclic left shift is as follows: the highest bit of the address binary is temporarily stored, the remaining bits of the address are shifted left by one bit, the temporarily stored highest bit is filled into the lowest bit of the address, and the new address after transformation is output.

9. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set. The at least one instruction, the at least one program, the code set, or the instruction set is parsed, loaded, and executed by the FPGA-supporting embedded processor or hardware logic to implement the FPGA-based hierarchical cyclic left shift FFT data rearrangement method as described in any one of claims 1 to 5.

10. An FPGA electronic device, characterized in that, The electronic device includes an FPGA programmable logic processor, an on-chip BRAM memory unit, an external DDR memory, and a high-speed data input / output interface; The memory stores FFT operation control logic and address cyclic left shift rearrangement instructions. These instructions are loaded by the FPGA processor and executed in parallel in hardware to implement the FPGA-based hierarchical cyclic left shift FFT data rearrangement method as described in any one of claims 1 to 5.