FFT processor, FFT calculation method, system on chip, integrated circuit and sensor
By using FFT kernel cascading and iterative butterfly operation structure, the contradiction between large number of points and high throughput of FFT computing devices in radar signal processing systems is resolved, realizing flexible FFT calculation, meeting the needs of various radar applications, and improving computing efficiency and throughput.
Patent Information
- Application Number
- CN202511310655.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-09-13
- Filing Date
- 2025-09-12
- Publication Date
- 2026-03-13
AI Technical Summary
In existing radar signal processing systems, FFT computing devices struggle to balance large point counts and high throughput, resulting in a significant contradiction between hardware resource and power consumption requirements, making it difficult to meet the flexible computing needs of various radar applications.
It employs a serial cascaded approach with multiple FFT cores, combining large point counts, pipelined and independent parallel modes, and achieves flexible FFT computation through iterative and cascaded butterfly operation structures, supporting FFT computation requirements with various radix and point counts.
It enables efficient and flexible FFT calculation to meet the needs of various radar applications with limited hardware resources, thereby improving the throughput and computational efficiency of radar signal processing.
Smart Images

Figure CN121658762A_ABST
Abstract
Description
Technical Field
[0001] This article relates to radar signal processing technology, particularly an FFT processor, FFT calculation method, radar signal processing system-on-a-chip, integrated circuit, and electromagnetic wave sensor. Background Technology
[0002] Radar is an electronic system that uses electromagnetic waves to detect targets and acquire information such as their distance, speed, and angle. Since its invention in the early 20th century, radar technology has been widely used in military, civilian, and scientific research fields, becoming one of the core tools of modern detection and sensing. As an important component of radar technology, radar target detection systems aim to extract target information from complex electromagnetic environments and achieve target detection, tracking, and identification.
[0003] The Digital Fourier Transform (DFT) is a well-defined discrete digital signal operation where the input and output are both discrete signal vectors of equal length. The Fast Fourier Transform (FFT) is an implementation of the DFT. In radar signal processing, the FFT is one of the core algorithms due to its efficiency and versatility. It is involved in many radar signal processing stages, such as pulse compression, Doppler processing, high-resolution imaging, and beamforming.
[0004] Each radar signal processing system-on-a-chip integrates at least one FFT processor to meet the needs of relevant components in performing FFT calculations. Summary of the Invention
[0005] This disclosure provides an FFT processor, an FFT calculation method, a radar signal processing system-on-a-chip, an integrated circuit, and an electromagnetic wave sensor. The FFT processor expands the number of FFT calculation points by serially cascading multiple FFT cores. Based on the independently configurable butterfly operation structure contained in each FFT core, it can implement FFT cores with various bases as needed, fully meeting the different FFT calculation requirements of relevant functional modules or operators in the radar chip.
[0006] This disclosure provides an FFT processor, which may include two cascaded FFT cores. The FFT processor may have at least two of the following modes: a large-point FFT mode, a pipelined mode, and an independent parallel mode. When the FFT processor is in the large-point FFT mode, the two FFT cores (i.e., dual cores) are configured to decompose an N-point FFT into two FFTs. When the FFT processor is in the pipelined mode, the cascaded FFT core is configured to perform a first FFT, and the cascaded FFT core is configured to perform a second FFT based on the result of the first FFT. And when the FFT processor is in the independent parallel mode, the two FFT cores are configured to independently and in parallel process data from different channels. This disclosure also provides an FFT processor including two cascaded FFT cores, which has at least two of the following operating modes: a large-point FFT mode, a pipelined mode, and an independent parallel mode. When the FFT processor is in the large-point FFT mode, the two FFT cores are configured to decompose an N-point FFT into two FFTs; When the FFT processor is in the pipelined mode, of the two FFT cores, the cascaded FFT core is configured to perform distance FFT, and the cascaded FFT core is configured to perform Doppler FFT; and... When the FFT processor is in the independent parallel mode, the two FFT cores are configured to process data from different channels independently and in parallel.
[0007] This disclosure also provides an FFT processor, including: At least two FFT cores are cascaded in series, and the maximum cardinality supported by the FFT cores is . FFT calculation; The FFT kernel includes at least one butterfly operation structure, and the FFT kernel is configured to operate according to a set kernel radix. Determine the number of butterfly operation structures in this FFT kernel that participate in the FFT calculation and the cardinality of each butterfly operation structure; <= , where i represents the i-th FFT kernel.
[0008] This disclosure also provides an FFT calculation method, the method comprising: Based on the number of FFT points to be processed and the computing power parameters of the FFT processor, the number of iterations for performing the FFT calculation and the actual computing configuration parameters of the FFT processor in each iteration are determined; wherein, the FFT processor includes at least two FFT cores cascaded in series, and the maximum cardinality supported by the FFT core is [missing information]. FFT calculation; Based on the number of iterations, perform the following steps in each iteration: Obtain the actual computational configuration parameters for this iteration and determine the core operation radix for each FFT core. ; According to the nuclear operation base Configure the corresponding FFT core, start the FFT processor, and use the at least two FFT cores to serially execute the FFT calculation for this iteration; Wherein, the kernel operation radix of the FFT kernel Used to determine the number of butterfly operation structures participating in this FFT calculation in this FFT kernel and the cardinality of each butterfly operation structure; The computing power parameters include: maximum supported FFT points. The actual calculation configuration parameters include: the actual number of FFT points calculated. ;
[0009]
[0010] I represents the number of FFT cores. <= , where i represents the i-th FFT kernel.
[0011] This disclosure also provides a radar signal processing on-chip system, including: Controller, multiple operators, and memory; The controller is configured to activate at least one of the plurality of operators to perform an operator function. The plurality of operators include: an FFT processor as described in the embodiments of this disclosure, comprising two cascaded FFT cores, and having at least two operating modes among large-point FFT mode, pipeline mode, and independent parallel mode; The FFT processor is configured to retrieve a working mode from the FFT processor configuration data storage area set in the memory according to the operator start instruction from the controller, and perform FFT calculation according to the working mode; or, The plurality of operators include: an FFT processor comprising at least two FFT cores cascaded in series, as described in embodiments of this disclosure; The FFT processor is configured to obtain the actual computational configuration parameters for this iteration from the FFT processor configuration data storage area set in the memory according to the operator start instruction from the controller, and execute the FFT computation for this iteration according to the actual computational configuration parameters.
[0012] This disclosure also provides an integrated circuit, including: The digital signal processing module performs FFT calculations based on the FFT calculation method described in any embodiment of this disclosure.
[0013] This disclosure also provides an electromagnetic wave sensor, including: a carrier; The radar signal processing system-on-a-chip as described in any embodiment of this disclosure, or the integrated circuit as described in any embodiment of this disclosure, is disposed on the carrier. The antenna is mounted on the carrier and is either integrated into a single device with the radar signal processing system-on-a-chip or the integrated circuit, or is disposed separately. The antenna is connected to the radar signal processing system-on-chip or integrated circuit and is used to transmit the electromagnetic wave signal and / or receive the echo signal.
[0014] Compared with related technologies, the configurable general-purpose FFT processor provided in this application takes into account multiple factors such as chip hardware cost, computing efficiency, FFT processor versatility, and storage access bandwidth, providing radar chips with an FFT computing implementation scheme that fully meets the computing needs of various applications.
[0015] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. Other advantages of this application can be realized and obtained by means of the solutions described in the description and the accompanying drawings. Attached Figure Description
[0016] The accompanying drawings are used to provide an understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.
[0017] Figure 1a This application provides a three-level butterfly structure for implementing radix-8 FFT computation. Figure 1b This application provides a first-level butterfly structure for implementing radix-2 FFT computation. Figure 1c This application provides a two-level butterfly structure for implementing radix-4 FFT computation. Figure 2 A schematic diagram of an FFT processor structure provided in an embodiment of this application; Figure 3a A schematic diagram of an FFT kernel bypass result provided in an embodiment of this application; Figure 3b This is a schematic diagram of another FFT kernel bypass result provided in an embodiment of this application; Figure 4a and Figure 4b This application provides an embodiment of an FFT refactoring configuration and result diagram; Figure 5 A flowchart of an FFT calculation method provided in this application embodiment; Figure 6a A schematic diagram of a radix-4 FFT operation structure with an added set of multipliers provided in an embodiment of this application; Figure 6b for Figure 6a A simplified representation; Figure 7 A schematic diagram of a ping-pong storage area provided in an embodiment of this application; Figure 8 A schematic diagram of an inter-core rotation factor multiplication operation provided in this application embodiment; Figure 9 This application provides a schematic diagram of N-point FFT calculation data grouping as an embodiment. Figure 10a This application provides a schematic diagram of an inter-core data transposition process. Figure 10b This application provides a schematic diagram of data transposition processing between adjacent iteration rounds. Figure 11 This is a schematic diagram of a radar signal processing system-on-a-chip structure provided in an embodiment of this application. Detailed Implementation
[0018] This application describes several embodiments, but these descriptions are exemplary and not restrictive, and it will be apparent to those skilled in the art that many more embodiments and implementations are possible within the scope of the embodiments described herein. Although many possible combinations of features are shown in the drawings and discussed in the detailed description, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with, or may replace, any feature or element of any other embodiment.
[0019] This application includes and contemplates combinations of features and elements known to those skilled in the art. The embodiments, features, and elements disclosed herein can also be combined with any conventional features or elements to form a unique inventive scheme as defined by the claims. Any feature or element of any embodiment can also be combined with features or elements from other inventive schemes to form another unique inventive scheme as defined by the claims. Therefore, it should be understood that any feature shown and / or discussed in the embodiments of this application can be implemented individually or in any suitable combination. Therefore, the embodiments are not limited except by the limitations imposed by the appended claims and their equivalents. Furthermore, various modifications and changes can be made within the scope of the appended claims.
[0020] Furthermore, in describing representative embodiments, the specification may have presented methods and / or processes as a specific sequence of steps. However, the method or process should not be limited to the specific order of steps described herein, to the extent that it does not depend on such a specific order. As will be understood by those skilled in the art, other sequences of steps are also possible. Therefore, the specific order of steps set forth in the specification should not be construed as a limitation of the claims. Moreover, the claims concerning the method and / or process should not be limited to the steps performed in the written order, and those skilled in the art will readily understand that these orders can be varied and still remain within the spirit and scope of the embodiments of this application.
[0021] An FFT processor is a hardware or software module specifically designed for efficient DFT computation and is widely used in radar signal processing systems. The FFT processor is a dedicated computing unit for accelerating DFT calculations; its core idea is to use divide-and-conquer algorithms (such as the Cooley-Tukey algorithm) to reduce the complexity from O(n log n) to O(n log n). ) decreased to O( ), XThe number of points in the FFT is the radix. To improve computational efficiency, radar signal processing systems on a chip (SoC) typically employ a hardware-implemented FFT processor. The butterfly arithmetic structure is the core computational unit in the FFT processor, also known as the butterfly arithmetic unit, used to implement radix-N butterfly operations, where N = 2, 4, or 8, etc. A single-level (or multi-layer) radix-N butterfly arithmetic structure / unit can perform an N-point FFT calculation in a single operation. Cascading multi-level (or multi-layer) butterfly arithmetic structures / units can achieve FFT calculations with even larger point counts. In FFT calculations, the radix N of the butterfly arithmetic structure / unit determines the divide-and-conquer strategy and computational efficiency of the algorithm. A larger radix for each butterfly reduces the number of multiplications and computational stages, meeting the needs of high-performance radar signal systems. However, implementing larger radix butterfly arithmetic structures / units increases complexity. Therefore, in practical hardware design, it is necessary to consider factors such as the number of points to be processed, hardware resources, and real-time requirements.
[0022] In an FMCW (Frequency-Modulated Continuous Wave) radar, for the intermediate frequency (IF) signal obtained through a mixer, range information (i.e., the location of the peak corresponds to the target's distance) can be extracted by performing a one-dimensional FFT (range FFT) on the ADC sampling data of a single chirp (i.e., the same frequency point after range FFT); velocity information (i.e., the location of the peak corresponds to the target's relative velocity) can be extracted by performing a second-dimensional FFT (Doppler FFT) on the same range cell of multiple chirs (i.e., the same frequency point after range FFT); and the target's angle can be estimated by performing a third-dimensional FFT (beamforming or angle FFT) on the same range-Doppler cell of multiple receiving antennas (Rx) / receiving channels. In other words, an FMCW radar signal processing link requires a large number of continuous, high-throughput FFT operations.
[0023] In the application-layer pipelined FFT architecture, there are mainly serial FFT architectures and parallel FFT architectures. Serial structures mainly include single-path delay feedback (SDF), single-path delay commutator (SDC), and single-stream feedforward (SFF), which have advantages such as high hardware resource utilization, small area, and low power consumption, but their throughput is low, typically processing only one sample per clock cycle (efficiency of 1 sample / cycle). This can become a bottleneck, especially when processing high data rates in radar, particularly when high-point multidimensional FFT is required. Parallel structures mainly include multi-path delay feedback (MDF), serial commutator (SC), and multi-path delay commutator (MDF), which have advantages such as high throughput and powerful performance, but their hardware resource consumption (such as logic units, memory, wiring, etc.) is extremely high, and their power consumption is also high. This is often unacceptable for radar chips that require dozens or even hundreds of receiving channels and are extremely sensitive to cost and power consumption. In other words, how to achieve high-throughput FFT operations sufficient to meet the real-time processing requirements of radar within a limited chip area and power budget has become a current technical challenge.
[0024] Taking radar chips as an example, the number of FFT calculation points varies depending on the detection accuracy and the computational needs of different functions. How a hardware-implemented FFT processor can flexibly meet the needs of various applications has a significant impact on the versatility of radar chips. Designing and implementing an FFT processor that balances execution efficiency and hardware cost, is flexible and configurable, and better meets the needs of various radar applications is crucial.
[0025] Taking the implementation of FFT computation using a radix-2 based butterfly operation unit as an example, to implement a radix-2 FFT at point X, a total of Y levels of butterfly operation structures are required. There are a total of G butterfly operation units. Each level has X / 2 butterfly units. For example, to implement an 8-point FFT using a radix-2 butterfly arithmetic unit, a 3-level butterfly arithmetic structure is needed, with 4 units per level, for a total of 12 units. For example, Figure 1a As shown, this is the architecture for implementing 8-point FFT computation; the four origins connected by each intersecting arrow are as follows: Figure 1b The diagram shows a pair of addition and subtraction operations, also known as a butterfly operation unit (butterfly structure). The W parameter indicated by each arrow represents a multiplication operation; the parameter is called the twiddle factor. This operation is called twiddle multiplication. As you can see, Figure 1a The FFT computation architecture shown includes a total of 12 butterfly structures, i.e., 24 addition and subtraction operations, and 12 twiddle multiplications. = 24 is directly proportional.
[0026] Taking the above 8-point FFT operation structure as an example, Figure 1a Of the three-level butterfly structure included, only the upper right part of the butterfly structure is used to perform operations, such as... Figure 1c The portion shown can perform 4-point FFT calculations, including 4 butterfly structures and 4 twiddle multiplications. = Proportional to 8; only the butterfly structure in the upper right corner is used to perform the calculation, such as Figure 1b The portion shown can perform 2-point FFT calculations. In other words, for a 3-level butterfly structure that performs 8-point FFT calculations, by controlling the use of the relevant butterfly operation units, 4-point and 2-point FFT calculations can also be achieved.
[0027] As can be seen, based on multiple radix-2 butterfly arithmetic units and using a multi-level hardware cascade approach, the needs of large-point FFT calculations can be met. If we consider implementing 8-point FFT calculations... Figure 1a The three-level operation structure shown is considered as a whole. Based on this whole structure, further FFT calculations with more points can be achieved through iterative operation. In this case, the three-level FFT butterfly structure consisting of 12 butterfly operation units that performs 8-point FFT calculations can be considered as a whole, called a radix-8 butterfly operation structure / unit. Taking range-dimensional FFT calculations in radar signal processing as an example, 512 / 1024-point FFT calculations are often involved. Using this multi-level cascading approach to implement the FFT processor, corresponding to 5 / 10 levels of cascading, requires 1280 / 5120 radix-2 butterfly operation units, posing a significant challenge to hardware cost and chip space.
[0028] By employing a time-for-space / hardware tradeoff, a 3-level (or 3-layer) radix-8 butterfly operation structure / unit can be used to achieve FFT calculations with a larger number of points through iteration. Using a radix-N butterfly operation structure / unit for k iterations, an N^k-point FFT calculation can be achieved. For example, using a radix-8 butterfly operation structure / unit for 3 iterations can achieve an 8^3 = 512-point FFT calculation. Compared to the basic Radix-N structure FFT, iterative structure FFT calculations require an additional set of twiddle multipliers.
[0029] To improve throughput, multiple 8-point FFT cores can be configured for pipelined collaboration to achieve FFT operations with more points. That is, based on the Radix-N FFT computation engine's ability to perform N-point FFT operations, this application proposes using a single Radix-N FFT computation engine to iteratively implement L-point FFT operations. If L is a power of N, then only k iterations are needed, with each iteration performing L / N Radix-N operations.
[0030] As can be seen, in order to meet the needs of FFT calculations with larger number of points, the physical cascade series of the butterfly operation structure / unit can be expanded, or iterative methods can be used in time based on a finite butterfly operation structure / unit.
[0031] It should be noted that, in the embodiments of this application, iteration can be defined as using a single FFT computing engine (such as the FFT processor in the embodiments of this application) to complete the entire large-point FFT calculation step by step through multiple loop iterations, thereby effectively improving flexibility and area efficiency. That is, the iterative architecture can greatly save chip area. The single FFT computing engine can be a single-core FFT processor, with a single core implementing radix-N FFT calculations, i.e., using a single-core FFT processor for loop iterations to achieve large-point FFT calculations; the single FFT computing engine can also be a multi-core cascaded FFT processor, with each core implementing radix-N. The FFT computation is implemented using a multi-core cascaded FFT processor. Point-wise FFT computation employs a multi-core cascaded FFT processor for iterative iteration, enabling the computation of FFTs with a larger number of points. For example, a radix-4 + radix-4 dual-core FFT processor can complete a 16^m point FFT through m iterations. In contrast, a pipelined architecture would require m radix-4 + radix-4 dual-core FFT processors connected in series to achieve the same 16^m point FFT. Therefore, compared to a pipelined architecture, the iterative architecture in this embodiment significantly improves area efficiency. Furthermore, since the core of the iterative architecture is a reusable computation unit (i.e., a butterfly arithmetic structure) and the corresponding control logic and memory scheduling scheme, the configurability of the FFT processor is achieved, thereby enhancing flexibility.
[0032] This disclosure provides an FFT processor, which may include two cascaded FFT cores. The FFT processor may have at least two modes selected from large-point FFT mode, pipelined mode, and independent parallel mode. When the FFT processor is in the large-point FFT mode, the two FFT cores (i.e., dual cores) are configured to decompose an N-point FFT into two FFTs. When the FFT processor is in the pipelined mode, the cascaded FFT core is configured to perform a first FFT, and the cascaded FFT core is configured to perform a second FFT based on the result of the first FFT. And when the FFT processor is in the independent parallel mode, the two FFT cores are configured to independently and in parallel process data from different channels. Among them, the aforementioned large-point FFT mode can be equivalently implemented as Radix-N^2, that is, through the close cascading of two cores, it can collaboratively complete the processing of ultra-large points (such as 4096 points), thereby achieving high-resolution FFT; the aforementioned pipeline mode can be configured to implement the sequential pipelined range FFT and velocity (Doppler) FFT of radar signal processing, that is, the output of the range FFT serves as the basis for Doppler FFT processing, such as... Figure 10a As shown, core 1 can be dedicated to implementing the first FFT, while core 2 can be dedicated to implementing the second FFT, forming a highly efficient processing pipeline that matches the radar frame structure. In the aforementioned independent parallel mode, the two cores work independently, processing data from different channels simultaneously, thereby enhancing the system's parallel capabilities.
[0033] In this embodiment, the cascaded FFT cores included in the FFT processing can operate in various ways. They can work collaboratively to achieve a larger FFT size as a whole, or they can perform smaller FFTs independently. Alternatively, the output of the cascaded first FFT core can be used as the input of the cascaded second FFT core, forming a pipelined operation. In other words, the two FFT cores can run in parallel, partially in parallel, or sequentially to form an FFT processor that meets the requirements of the corresponding scenario.
[0034] Optionally, when two cascaded FFT cores operate as a single unit, range-dimensional FFT and Doppler-dimensional FFT of radar signals can be implemented. For example, the current iteration is used for range-dimensional FFT, while the next iteration is used for the corresponding Doppler-dimensional FFT, thus implementing range-dimensional FFT and Doppler-dimensional FFT separately through two iterations. Furthermore, if the FFT core's computational power is sufficient, the cascaded first FFT core can also perform range-dimensional FFT in radar signal processing, while the cascaded second FFT core can be configured to perform Doppler-dimensional FFT based on the result of the preceding range-dimensional FFT.
[0035] In some exemplary embodiments, the FFT kernel includes at least two layers (levels) of computational structure; between adjacent layers, the cardinality of the smallest computational unit (such as the cardinality of each butterfly operation) increases or decreases sequentially along the data flow direction. For example, for an FFT kernel for DIF, the cardinality of the smallest computational unit between adjacent layers decreases along the data input / output dimension, while for an FFT kernel for DIT, the cardinality of the smallest computational unit between adjacent layers increases along the data input / output dimension (such as the cardinality of the first-level butterfly operation being 2, the cardinality of the second-level butterfly operation being 4, and the cardinality of the third-level butterfly operation being 8). The FFT kernel can be a parallel structure, meaning the cardinality of the FFT kernel is greater than or equal to 2.
[0036] It should be noted that, in the embodiments of this application, FFT cores can be cascaded through storage-write-read transpose and twiddle factor multiplication, thereby achieving serial processing of the data stream. Furthermore, different cores can also be time-division multiplexed (iterated) to achieve FFT processing with a larger number of points based on fewer physical cores. Specifically, an L-point Radix-N FFT core can include a structure formed by multiple Radix-N butterfly basic units, control logic, memory, and twiddle factor multipliers, etc., to implement an L-point FFT, where L is greater than N.
[0037] In some exemplary embodiments, at least one of the FFT cores includes at least two computational layers, some of which are configured to be skipped. That is, when some of the computational layers are skipped, the core's cardinality changes. For example, for an FFT core with a maximum cardinality of N, when some of its layers are configured to be skipped, it can function as a Radix MFFT core with a different cardinality, where M can be a factor of N, i.e., there exists an integer ratio r such that M... r = N.
[0038] In some exemplary embodiments, the FFT processor may further include a buffer located between the two FFT cores with the largest cardinality; when the FFT processor is in the large-point FFT mode or the pipelined mode, the buffer is configured to transpose the output data of the cascaded FFT core and then use it as the input to the cascaded FFT core; when the FFT processor is in the independent parallel mode, the buffer is configured to include two simplex memories so that the two FFT cores can work simultaneously.
[0039] For example, to enable two FFT cores to cooperate serially to complete an N^2-point FFT, a buffer can be set between the two cores to transpose the data from (N1, N0) order to (N0, N1) order. This buffer can be used to cache N cycles of output data from one core, with N data points per cycle, for a total of N^2 output data points. Subsequently, according to the extracted order, the N^2 data points can be divided into N cycles, with N data points per cycle passed to the other core as input. Simultaneously, to allow the two cores to work concurrently, the aforementioned memory can also be configured to operate in full-duplex mode, simultaneously accepting write requests from one core and read requests from the other core, and supporting the aforementioned extraction requests; or the memory can have two simplex memories operating in ping-pong mode, so that at any given time, one simplex memory accepts write requests from one core, and the other simplex memory provides read requests from the other core. Furthermore, tasks can be exchanged every N cycles after N^2 data points have been written / read simultaneously.
[0040] This disclosure provides an FFT processor, such as... Figure 2 As shown, it includes: at least two FFT cores 210 cascaded in series, wherein the maximum number of bases supported by the FFT cores 210 is [missing information]. FFT calculation; The FFT core includes at least one butterfly operation structure, and the FFT core 210 is configured to operate according to a set core operation radix. Determine the number of butterfly operation structures in this FFT kernel that participate in the FFT calculation and the cardinality of each butterfly operation structure; <= , where i represents the i-th FFT kernel.
[0041] It should be noted that, in this embodiment of the disclosure, the FFT kernel is a maximum support... The butterfly computation unit for point FFT computation includes at least one level of butterfly operation structure, and supports dynamic configuration of the core operation radix within the maximum supported FFT point range. This allows for FFT computation with a target number of points. The FFT processor includes at least two FFT cores connected serially. By cascading multiple FFT cores, it overcomes the limitations of a single FFT core in processing points, enabling FFT computation with a large number of points. Each of the multiple FFT cores has an independently configured core arithmetic radix, and the multiple FFT cores work collaboratively to meet the requirements of the target number of points for FFT computation. An FFT processor consisting of at least one FFT core serves as the basic unit for iterative computation. Iterative computation based on this multi-core FFT processor enables FFT computation with even larger numbers of points or multi-dimensional FFT computation.
[0042] It should be noted that the butterfly operation structure in each stage of the FFT operation core in this disclosure embodiment is a general term for the basic Radix-N operation structure / unit. This includes structures like radix-2 (2 inputs, 2 outputs, and a twitch factor multiplication), resembling a butterfly; structures like radix-3 (3 inputs, 3 outputs, involving more complex twitch factor combinations, and not exhibiting a "butterfly" shape); and structures like radix-4 (4 inputs, 4 outputs, referred to as "Radix-butterfly"), or other structures, not limited to radix-2 butterfly structures. These butterfly operation structures can be used individually or cascaded to construct the overall computational structure of the FFT core.
[0043] In some exemplary embodiments, the FFT kernel supports a maximum cardinality. : ; Where J is the total number of butterfly operation levels included in the i-th FFT kernel. Let be the cardinality of the j-th level butterfly operation structure in the i-th FFT kernel; exist If the FFT core is equal to the base of the first-level butterfly operation structure in this FFT core, or the product of the bases of the multi-level butterfly operation structures, the FFT core is set to determine that the first-level or multi-level butterfly operation structure participates in the execution of this FFT calculation, and bypasses other levels of butterfly operation structures other than the first-level or multi-level. exist If the cardinality of the FFT core is not equal to the cardinality of the first-level butterfly operation structure in this FFT core, nor equal to the product of the cardinality of the multi-level butterfly operation structures, then the FFT core is configured to reconfigure the data flow path of at least one level butterfly operation structure in this FFT core, so that the cardinality of this FFT core is... FFT calculation.
[0044] in, The case where the number of operations equals the radix of a single-level butterfly operation structure in this FFT kernel, or the product of the radixes of multiple-level butterfly operation structures, is understood as the kernel's operational radix. The kernel arithmetic radix is obtained by multiplying the radix of one or more levels corresponding to the original butterfly arithmetic structure at least one level. Also known as factor base. The case where the base number is not equal to the base number of a single-level butterfly operation structure in this FFT kernel, nor equal to the product of the base numbers of multiple-level butterfly operation structures, is understood as the kernel operation base number. It cannot be obtained by multiplying the radix of one or more levels corresponding to the original butterfly operation structure at least one level; the radix of this kernel operation... Also known as a non-factory basis. In the case of a non-factory basis, it is necessary to reconfigure the data flow path of at least one level of butterfly operation structure in this FFT core to achieve the effect of changing the radix of one or more levels of butterfly operation structure.
[0045] by Figure 1a Taking the FFT kernel shown as an example, it includes a J=3-level butterfly operation structure, with each level of the butterfly operation structure having a base of 2, i.e. Figure 1a The FFT kernel shown supports a maximum of 2×2×2=8 point FFT computation. When this FFT kernel is regarded as a whole, it is called a radix-8 FFT kernel.
[0046] Number of kernel operations per FFT kernel This is a configurable parameter, within the range of the maximum supported base. Figure 1a For example, The case where it equals the radix of the first-level butterfly operation structure in this FFT kernel, or the product of the radixes of multiple-level butterfly operation structures, i.e. It is equal to the radix 2 of a level 1 butterfly operation structure, or equal to the radix product 2×2=4 of a level 2 butterfly operation structure, or equal to the radix product 2×2×2=8 of a level 3 butterfly operation structure. Correspondingly, the other two levels are bypassed, and the level 1 butterfly operation structure is retained as valid, which is equivalent to Figure 3a The FFT kernel in the smallest bounding box; bypassing other level 1 operations while retaining the effective level 2 butterfly operation structure, is equivalent to Figure 3a The FFT kernel in the middle frame; no bypassing, retaining the effective 3-level butterfly operation structure, equivalent to Figure 3a The FFT kernel in the largest bounding box.
[0047] It should be noted that, in the embodiments of this application, "configurable" can refer to at least some or similar parameters among the parameters of the FFT, such as the number of points (N), radix, data bit width, window function, and scaling scheme, which can be configured through registers to achieve system adaptability and chip reuse. For example, when a radar chip or a radar system faces multiple scenarios (such as highway and parking scenarios for autonomous driving), a configurable architecture enables the radar to dynamically adjust parameters such as FFT to achieve different resolutions and operating distances to adapt to different modes or scenarios, thereby achieving a balance between performance and power consumption. Similarly, when the same chip can be used in different radar product lines through configurable parameters, redundant development is avoided, effectively reducing R&D costs and complexity.
[0048] The FFT algorithm generally includes two basic decomposition approaches in its architecture: Decimation-In-Time (DIT) and Decimation-In-Frequency (DIF). These two methods differ in the order of operations, the timing of twiddle factor multiplication, and the order of the input / output sequences, thus leading to different hardware implementation architectures. Correspondingly, from a hardware implementation perspective, DIT-based FFT and DIF-based FFT, these two different basic architectures, will also differ in data flow direction, memory access patterns, and control logic complexity.
[0049] For example, the input / output order of the DIF-based FFT architecture is: input in natural order, output in reverse order. Simultaneously, the multiplication (rotation factor) needs to be set after the butterfly operation, meaning the data path is addition followed by multiplication. In logic control, address generation is relatively regular and easy to control. Since direct input of ADC data can be achieved through backend processing and corresponding address mapping, it has a high degree of compatibility with radar data streams. Therefore, this architecture is well-suited for ASIC / dedicated chip scenarios, especially for streaming data processing applications such as FMCW radar where the input is in natural order.
[0050] Figure 3a This illustrates an 8-point radix-2 DIF FFT architecture. Figure 3b An 8-point radix-2 DITFFT architecture is illustrated. The FFT core in this embodiment is illustrated using the DIF FFT architecture as an example, but it can be converted to a DIT FFT architecture with some adjustments based on actual application requirements; for example... Figure 3a The DIF FFT shown here performs the transformation at the output, while... Figure 3b The DIT FFT shown here, since its transformation is at the input end, can also realize the FFT of DIT by changing the position of the data transformation based on the DIF FFT core in the embodiment of this application, so as to be used for general processing or DSP FFT operation in radar or other fields where the output order is strictly required.
[0051] The FFT processor in this embodiment has a cascaded dual-core / multi-core configurable hybrid radix iterative FFT architecture. Its algorithm can be based on decimation-by-frequency (DIF) to efficiently process the natural sequential data stream input from the front-end ADC (analog-to-digital converter), simplifying data interface control. It can also process the output sorting through back-end address mapping to achieve better overall hardware efficiency.
[0052] Still with Figure 1a For example, The case where the base is not equal to the base of the first-level butterfly operation structure in this FFT kernel, nor equal to the product of the bases of multiple-level butterfly operation structures, for example... If it equals 3, or equals 6, then it needs to be adjusted. Figure 1a The data flow path of the illustrated 3-level butterfly operation structure is reconfigured, for example, reconfigured into a radix-2 and radix-3 two-level butterfly operation structure, thereby enabling 3-point FFT computation or 6-point FFT computation. In this case, For example, the numerical values that can be achieved by changing the data flow path based on an initial butterfly operation structure of at least one level, such as... Figure 4a The inner part of the right side 420 is reconnected, that is, the data flow path is reconfigured, resulting in... Figure 4b The portion within the right side 430 remains unchanged, while the portion within the left side 410 remains unchanged, thus achieving one = 6 FFT kernel.
[0053] Changing the data flow path involves reorganizing the computational resources within the FFT kernel, altering the data order and computation process to achieve this type of non-factorial FFT kernel. These computational resources include a butterfly operation structure and its corresponding twiddle factor. More detailed reconfiguration schemes are not illustrated here.
[0054] Each FFT core is configured with a core operand radix. Then, the FFT processor can be used to achieve... Point FFT calculation, ; I represents the number of FFT cores.
[0055] As can be seen, by serially connecting at least two FFT cores, the number of FFT points that an FFT processor can handle can be increased. For example, it includes two FFT cores, each with the following configuration: Figure 1a The radix-8 FFT kernel shown can achieve a maximum of 8×8=64 points of FFT computation with two serial connections; and can achieve a maximum of 8×8×8=512 points of FFT computation with three serial connections.
[0056] In some exemplary embodiments, the FFT processor further includes at least two in-core twitch factor storage units corresponding one-to-one with the at least two FFT cores, which respectively store the twitch factors required for performing FFT calculations within each FFT core.
[0057] It is understandable that an FFT kernel implementation... Point FFT calculations require corresponding twiddle factors, known as kernel twiddle factors. Therefore, corresponding storage units are configured to store this data. Each FFT kernel corresponds to its own storage unit and the required twiddle factor data.
[0058] In some exemplary embodiments, the FFT processor further includes: at least one inter-core twitch factor storage unit, each inter-core twitch factor storage unit storing the twitch factor required for two adjacent FFT cores to perform FFT calculations serially; And at least one set of inter-kernel multipliers, the set of inter-kernel multipliers being used to perform multiplication operations of the FFT calculation result of the previous FFT kernel and the corresponding inter-kernel twiddle factor to obtain the data to be processed in the current FFT kernel.
[0059] Each set of inter-core multipliers performs the inter-core twiddle factor multiplication operation required when two adjacent FFT cores are cascaded.
[0060] It is understandable that two FFT cores are connected in series to achieve this. Point FFT calculations also require a corresponding twiddle factor, known as the inter-core twiddle factor. Therefore, a corresponding storage unit is configured to store this data. Each pair of adjacent FFT cores corresponds to one storage unit and the required inter-core twiddle factor data.
[0061] To leverage the existing butterfly arithmetic architecture and further achieve FFT calculations with more points, iterative calculations can be performed based on the FFT processor, i.e., the FFT processor is triggered multiple times to execute calculations. For example, an FFT processor with two FFT cores, each core supporting a maximum of 4-point FFT calculations, can achieve 4×4=16-point FFT calculations. Using this FFT processor for two iterations can achieve 16×16=256-point FFT calculations.
[0062] In some exemplary embodiments, the FFT processor further includes: at least one iterative rotation factor storage unit, each iterative rotation factor storage unit storing the rotation factor required for two consecutive iterations of FFT calculation performed by the multi-core FFT processor; And at least one set of iterative multipliers, which are used to perform multiplication operations on the FFT calculation result of the previous iteration and the corresponding iterative twitch factor to obtain the data to be processed in the current iteration.
[0063] Each set of multipliers corresponds to the iterative twiddle factor multiplication operation required for two consecutive iterations.
[0064] It's understandable that using an FFT processor for iterative computation to achieve large-point FFT calculations also requires a corresponding rotation factor, known as the iterative rotation factor. Therefore, a corresponding storage unit is configured to store this data. Each two iterations correspond to one storage unit and the required rotation factor data.
[0065] The intra-kernel twiddle factor, inter-kernel twiddle factor, and iterative twiddle factor involved in the embodiments of this disclosure are predetermined and stored in corresponding allocated storage units according to different FFT calculation needs. It should be noted that in the scheme of the embodiments of this disclosure, the FFT processor involves multiplying the corresponding twiddle factors intra-kernel, inter-kernel, and between adjacent iterations, which can be implemented using the corresponding configured multipliers. The specific aspects of the multipliers are not discussed in detail here.
[0066] The FFT processor provided in this disclosure, through a combination of in-core butterfly architecture configuration, inter-core serial computation, and multi-round iteration, flexibly meets various FFT computation needs while utilizing a limited number of hardware resources (butterfly arithmetic structure), under the basic premise of meeting the number of FFT computation points. Radar chips implemented based on this FFT processor can meet the FFT computation needs of multiple stages in radar signal processing. In some exemplary embodiments, considering the FFT computation needs of mainstream radar applications, the FFT processor includes two FFT cores, also referred to as a serial dual-core configurable FFT processor.
[0067] In the cascaded (or serial) dual-core / multi-core architecture of this application embodiment, FFT processing can be implemented using two or more FFT cores based on a "serial" data flow, thereby improving performance and enabling flexible task allocation. Since two or more FFT cores (i.e., dual-core / multi-core architecture) enable parallel data processing, the throughput of a dual-core architecture can be increased by up to two times compared to a serial architecture. Simultaneously, the dual-core / multi-core architecture capable of parallel processing can implement various collaborative working modes. For example, a dual-core architecture can at least achieve: Mode 1, using a decomposition algorithm to collaboratively compute a large FFT; Mode 2, each core independently computes two relatively smaller FFTs to meet the needs of radar simultaneously processing range and velocity dimensions, or different range accuracy / resolution for different range ranges; Mode 3, one core is dedicated to range FFT, and the other to Doppler FFT, forming an efficient pipeline to further improve overall throughput.
[0068] Using the FFT processor provided in the disclosed embodiments, this disclosure also provides an FFT calculation method, such as... Figure 5 As shown, the method includes: Step 510: Based on the number of FFT points to be processed and the computing power parameters of the FFT processor, determine the number of iterations for performing the FFT calculation, and the actual computing configuration parameters of the FFT processor in each iteration; wherein, the FFT processor includes at least two FFT cores cascaded in series, and the maximum supported cardinality of the FFT core is [missing information]. FFT calculation; Step 520: Based on the number of iterations, perform the following steps in each iteration: Step 521: Obtain the actual computational configuration parameters for this iteration and determine the core operation radix for each FFT core. ; Step 522, based on the core operation radix Configure the corresponding FFT core, start the FFT processor, and use the at least two FFT cores to serially execute the FFT calculation for this iteration; Wherein, the kernel operation radix of the FFT kernel Used to determine the number of butterfly operation structures participating in this FFT calculation in this FFT kernel and the cardinality of each butterfly operation structure; The computing power parameters include: maximum supported FFT points. The actual calculation configuration parameters include: the actual number of FFT points calculated. ; ; ; I represents the number of FFT cores. <= , where i represents the i-th FFT kernel.
[0069] The FFT calculation method provided in this disclosure, based on the aforementioned flexible and configurable FFT processor, fulfills various FFT point count requirements. Before executing the FFT calculation, it is determined whether iterative calculation is needed, the number of iterations, and the actual number of calculation points for each FFT core in each iteration, based on the calculation requirements and the hardware capability parameters of the FFT processor. Further, if the actual number of calculation points is less than the maximum supported number of points for the FFT core, bypass processing of one or more levels of butterfly operation structures, or data flow path reconfiguration processing, is performed. Finally, iterative calculation is performed according to the number of iterations. Specifically, iterative execution achieves a larger number of FFT points. The total achievable FFT points are equal to the product of the number of FFT points achieved in each iteration. The butterfly operation structure included in the FFT processor can remain unchanged in each iteration, enabling the same number of points for each iteration. Alternatively, the butterfly operation structure in at least two FFT cores can be reconfigured as needed in each iteration, enabling different number of points for each iteration.
[0070] In some exemplary embodiments, the step of basing operations on the kernel is... Configure the corresponding FFT core, including: Based on the kernel operation radix of each FFT kernel The corresponding number of butterfly operation structure levels and the cardinality of each butterfly operation structure involved in this FFT calculation are determined, and the effective first-level or multi-level butterfly operation structure for executing this FFT calculation is configured in each FFT core, thereby determining the effective serial butterfly operation structure for the FFT processor to execute this FFT calculation.
[0071] It is understandable that this is based on the core operation radix of each FFT core. The corresponding number of butterfly operation structure levels and the cardinality of each butterfly operation structure involved in this FFT calculation are determined to determine the effective first-level or multi-level butterfly operation structure in the FFT core for this FFT calculation; that is, first determine the multi-level structure within each core. Based on the serial order of the at least two FFT cores, and based on the effective multi-level butterfly operation structure for executing the current FFT calculation in each FFT core, the effective serial butterfly operation structure for the FFT processor to execute the current FFT calculation is determined; that is, the overall structure of the multi-core serial operation is then determined.
[0072] As can be seen, in the overall FFT calculation, based on the total number of FFT computation points (target FFT computation points), the number of iterations is first determined, decomposing the total number of FFT computation points into the number of FFT computation points required for each iteration. The product of the number of FFT computation points achieved in multiple iterations is the total number of FFT computation points. Then, for at least two FFT cores in each iteration, the number of FFT computation points for that iteration is determined, decomposing the number of FFT computation points required for that iteration into each FFT core, i.e., the core operation base of each FFT core. The product of the number of FFT computation points achieved by at least two FFT cores is the number of FFT points required for this iteration. Furthermore, the core operation radix for each core... Determine the number of levels of the butterfly operation structure actually involved in this iteration and the cardinality of each level.
[0073] It should be noted that, compared to the initial full butterfly operation structure in the FFT kernel, the kernel operation radix... The butterfly operation structure configured to perform this calculation is called the effective first-level or multi-level butterfly operation structure of the FFT core. Compared with the full butterfly operation structure initially set for each FFT core in the FFT processor, the butterfly operation structure configured to perform this calculation after being decomposed into each FFT core according to the total number of FFT calculation points is called the effective serial butterfly operation structure of the FFT processor.
[0074] Using a defined and effective serial butterfly operation structure, perform FFT calculations to calculate the total number of FFT points and output the calculation results.
[0075] In some exemplary embodiments, starting the FFT processor and serially executing the FFT calculation for the current iteration using the at least two FFT cores includes: Obtain the kernel rotation factor corresponding one-to-one with the at least two FFT kernels, and use it for FFT calculation within each FFT kernel; Obtain the inter-core rotation factor between adjacent FFT cores in the at least two FFT cores, and use it for serial cascading of the corresponding adjacent FFT cores; The FFT processor is started, the data to be processed in this iteration is obtained, the valid serial butterfly operation structure is input, the FFT calculation is performed according to the intra-core twitch factor and the inter-core twitch factor, and the corresponding FFT calculation result is output.
[0076] As can be seen, FFT computation requires that the FFT computation within each FFT core be performed based on the acquired intra-core twiddle factor, while the computation between adjacent FFT cores be performed based on the acquired inter-core twiddle factor. The specific data of these twiddle factors are predetermined and stored based on the FFT computation requirements and the determined effective multi-core serial butterfly operation structure.
[0077] In some exemplary embodiments, the two FFT cores in a dual-core FFT processor support a maximum of = The core operation radix of the two FFT cores in the configured effective dual-core FFT processor is = ,or, ≠ For cases where the numbers are unequal, this is also referred to as a dual-core mixed-radix FFT processor. A typical FFT is usually radix-2 or radix-4, meaning the number of points N is a power of 2 or 4 (e.g., 1024, 4096, etc.). However, the mixed-radix FFT in this embodiment allows the number of points N to be the product of multiple different radixes (e.g., N = 256 = 4x4x4x4 or N = 360 = 5x8x9), thus giving the FFT processor great flexibility in point configuration. For FMCW radar, this makes its standards and modes more diverse, enabling the combination of different FFT point numbers to achieve optimal resolution in different applications / scenarios (short-range, medium-range, long-range, in-cabin, etc.). It also avoids the waste of resources caused by padding with zeros to make the power of "2". For example, for calculating a 300-point FFT, the mixed basis can be calculated directly and efficiently, while the zero-padding method based on "2" needs to pad with zeros to 512 points, wasting 44% of the computation. That is, while avoiding the waste of resources, it can also save computation time and power consumption.
[0078] like Figure 2As shown, at least one of the cascaded FFT core 1 and FFT core 2 has Figure 3a The layered structure shown can be used as a Radix M-point FFT with different bases when some computational layers are skipped (i.e., invalidated), where M is a factor of N, i.e., there exists an integer ratio r such that M × r = N. For example, consider a three-layer Radix8 FFT computational structure (e.g., Figure 3a As shown, if it is configured to skip one layer of computation (i.e., invalidate that layer), it can function as a Radix 4 FFT; if it skips two layers of computation, it can function as a Radix 2 FFT. Correspondingly, if the base of FFT core 1 is configured as M, the buffer in this embodiment will be transposed from (N, M) row-column order to (M, N) row-column order according to its configuration change. The corresponding FFT core 2 can then select the corresponding rotation factor from the read-only rotation factor structure based on the ratio of N to M. That is, in this embodiment, for dual-core or multi-core FFT processing, multiple point-count FFTs can be achieved by changing the configuration of at least one FFT core, i.e., an M*N point FFT can be achieved through dual-core cooperation. Furthermore, an M0×N×M1×N point FFT can be further achieved through iteration, where M0 and M1 represent configurable base cores that can be configured with different bases in the two iterations.
[0079] In some exemplary embodiments, based on the above-mentioned approach of flexibly setting different layers, the computational resources (such as butterflies, twiddle multipliers, etc.) in the iterative Radix structure can be reorganized (e.g., rewiring / rerouting operations) to change the data order and computation process. This allows for the implementation of non-factorial FFT bases, i.e., without being restricted by the condition that the integer ratio r makes M×r = N, thus enabling arbitrary M... An N-point FFT, for example, M = 2, N = 6, can be used to calculate a 12-point FFT.
[0080] In some exemplary embodiments, the radar chip utilizes a dual-core serial FFT processor, which balances requirements such as the number of FFT computation points, FFT processor hardware cost, and FFT computation efficiency. Specifically, the FFT processor includes two serially connected FFT cores: FFT core 1 and FFT core 2. = In the case of FFT, in each iteration, the FFT calculation is performed according to the following formula: ; Where N= , The nuclear rotation factor of FFT kernel 1, The nuclear rotation factor of FFT kernel 2. The internuclear rotation factor between FFT kernel 1 and FFT kernel 2; X represents the FFT data to be processed, N represents the total number of FFT data points to be processed in this iteration by the FFT processor, and k represents the index of the FFT data points to be processed in this iteration. This is the grouping index for the FFT data points to be processed in this iteration. This is the index of the FFT data point to be processed in this iteration within the current group. = 0, 1,…, – 1; = 0, 1, …, – 1.
[0081] Where N is the number of FFT computation points executed in one iteration by a dual-core FFT processor. , This refers to the radix of the kernel operations for each of the two kernels. If the operation is performed in multiple iterations, each iteration corresponds to the same or different N. , Configuration.
[0082] In some exemplary embodiments, in each iteration, activating the FFT processor and serially executing the FFT calculation for this iteration using the at least two FFT cores further includes: According to the set data transposition rules, transposed data is obtained from the calculation result cache as the data to be processed in the current iteration of the FFT processor; wherein, the calculation result cache stores the FFT calculation result of the previous iteration of the FFT processor; and / or, according to the set data transposition rules, transposed data is obtained from the calculation result cache as the data to be processed in the current iteration of the non-first FFT core among the at least two FFT cores; wherein, the calculation result cache stores the FFT calculation result of the previous FFT core connected in series in the current iteration.
[0083] When FFT cores are serially cascaded to expand the number of FFT computation points, each subsequent FFT core needs to extract or transpose the result data from the previous FFT core's FFT computation as needed before processing, depending on the FFT computation requirements. Data connections between cores are implemented using shared memory. For example... Figure 10a As shown, the data to be processed in FFT core 2 is obtained by transposing the result data after FFT calculation performed by computing core 1.
[0084] like Figure 10a As shown, the FFT processor in this embodiment may include two FFT cores arranged in a serial relationship, namely FFT core 1 and FFT core 2. The output of FFT core 1, after preprocessing operations such as transposition (e.g., twiddle multiplication), is used as the input of FFT core 2, thereby realizing operations such as FFT decomposition and linear processing of the data stream. The preprocessing factor parameters, such as the transposition factor, can be stored in a read-only memory such as a dedicated twiddle. Furthermore, FFT core 1 and FFT core 2 can also run in parallel and independently to perform independent FFT processing, in which case the aforementioned preprocessing-related devices are not required or needed.
[0085] When an FFT processor expands the number of computation points through multiple iterations, the result data from the previous iteration needs to be extracted or transposed as required by the FFT computation in each iteration before further processing, depending on the computational needs. Data transfer between iterations is achieved using shared memory. For example... Figure 10b As shown, the data to be processed in the second iteration is obtained by transposing the result data of the FFT calculation performed in the first iteration; for example, the first iteration performs distance FFT calculation, and the second iteration performs Doppler FFT calculation.
[0086] The data transpose rules set can be the same or different in different calculation stages. In some exemplary embodiments, they also change according to the changes in the actual kernel operation base of the FFT calculation. Not limited to specific aspects, more detailed transpose rules and their effects are not discussed in detail here.
[0087] An FFT processor includes at least two FFT cores. In one iteration, the FFT calculation result of the previous iteration is transposed and used as the data to be processed in the current iteration, indicating that this data is the data to be processed by the first FFT core. The data to be processed by non-first FFT cores is obtained by transposing the FFT calculation result of the previous FFT core connected in series in the current iteration.
[0088] In some exemplary embodiments, the calculation result cache includes at least two iterative data partitions, a first iterative data partition for storing the FFT calculation result of the current iteration, and a second iterative data partition for storing the FFT calculation result of the previous iteration; in two adjacent iterations, the physical cache areas corresponding to the first iterative data partition and the second iterative data partition are swapped.
[0089] As can be seen, the iterative data partition is a shared data storage area between the last FFT core and the first FFT core in two adjacent iterations. The current iteration retrieves the FFT calculation result from the previous iteration from the iterative data area, transposes it, and then executes the FFT calculation for the current iteration. To continuously execute FFT calculations, these two iterative data partitions work in a ping-pong manner. In the current iteration, the FFT processor reads data from the first iterative data partition, transposes it, executes the FFT calculation, and writes the calculation result to the second iterative data partition; in the next iteration, the FFT processor reads data from the second iterative data partition, transposes it, executes the FFT calculation, and writes the calculation result to the first iterative data partition. Based on this, pipelined processing of multiple batches of input data can be achieved, improving overall operating efficiency. When applied to radar data processing chips, it can effectively improve processing speed and meet the needs of real-time data processing.
[0090] In some exemplary embodiments, the computation result cache includes at least one pair of inter-core data partitions, each pair of inter-core data partitions being a shared data storage area for two connected FFT cores. The pair of inter-core data partitions includes: a first inter-core data partition and a second inter-core data partition. The first inter-core data partition is used to store the FFT computation result of the preceding FFT core in the current iteration, and the second inter-core data partition is used to store the FFT computation result of the preceding FFT core in the previous iteration. In two adjacent iterations, the physical cache areas corresponding to the first inter-core data partition and the second inter-core data partition are interchanged.
[0091] As can be seen, the FFT processor includes I cascaded FFT cores, and the computation result cache includes I-1 pairs of inter-core data partitions. Each pair of inter-core data partitions is shared by the corresponding two connected FFT cores, and each pair includes two inter-core data partitions that operate in a ping-pong manner. Taking a dual-core FFT processor as an example, the result cache includes one pair of inter-core data partitions. In the p-th iteration, core 1 writes the FFT computation result to the first inter-core data partition; in the p+1-th iteration, core 1 writes the FFT computation result to the second inter-core data partition; in the p+2-th iteration, core 1 writes the FFT computation result to the first inter-core data partition; and so on. Correspondingly, when core 2 reads data from the first or second inter-core data partition, performs transposition processing, and executes the FFT computation, core 1 and core 2 do not access the same inter-core data partition at the same time point or within the same time interval. Figure 7As shown, to enable two FFT cores (core 1 and core 2) to cooperate serially to complete an N^2-point FFT, a buffer is set between the two cores to transpose the data from (N1, N0) row-column order to (N0, N1) row-column order. This buffers N output data from core 1, with N data points per output, for a total of N^2 output data points. Following the extracted order, the N^2 data points are divided into N outputs, with N data points per output passed to core 2 as its input. Simultaneously, to allow both cores to work concurrently... Figure 7 The memory shown can also have a full-duplex operating mode, such as being able to simultaneously receive write requests from core 1 and read requests from core 2, and also simultaneously support the above-mentioned extraction requests. Alternatively, it can also use two simplex memories to operate in a ping-pong manner, that is, as long as at any given time, one simplex memory can accept write requests from core 1 and the other simplex memory can provide read requests from core 2. At the same time, it can also switch tasks to run every N cycles after simultaneously writing / reading N^2 data.
[0092] For example, an FFT processor includes two FFT cores, each supporting a maximum radix of 4. A dual-core FFT processor can perform 4×4=16-point FFT calculations. The first-level radix-4 butterfly arithmetic structure included in the FFT core can also be as follows: Figure 6a The Radix-4 direct structure shown is compared to the two-level radix-2 butterfly structure, as shown in the example. Figure 1c The diagram shown is a radix-4 structure treated as a whole, with an additional 4-1=3 multipliers on the left side for calculating inter-kernel twiddle factors. Its simplified representation is as follows: Figure 6b As shown. A dual-core serial connection implements 16-point FFT computation, with each FFT core performing 16 / 4 = 4 radix-4 FFT computations. The processing steps for 16-point FFT computation are as follows: Figure 8 As shown, the multiplication operations related to the internuclear rotation factor are located between the two vertical lines on the right.
[0093] In some alternative embodiments, such as Figure 8 As shown, a 4^2 = 16-point FFT kernel is calculated using a Radix-4 FFT in two iterations, meaning each iteration performs four Radix-4 operations. The result of the first round of radix calculation is extracted (data exchanged) and used as the input for the four radix operations in the second round. Figure 8As shown, the three layers of read nodes (the solid black dots in the diagram) along the direction from the input end (left side of the diagram) to the output end (right side of the diagram) correspond sequentially to the data read behavior (input end), data write-desorption (read) behavior (middle position), and data write behavior (output end). Since the Radix-4 iterative structure requires simultaneous reading / writing of 4 data points, the corresponding memory bandwidth is at least 4 (i.e., greater than or equal to the radix N of the FFT core).
[0094] In some exemplary embodiments, in a dual-core FFT processor, core 1 is the base core. FFT kernel, kernel 2 as the base The FFT kernel. Implementing N= Point FFT calculation. That is, decomposing the N-point sequence X[n] into... Segment, length of each segment At this time, n = n1 + n2, n1 = 0, 1, …, – 1; n² = 0, 1, …, – 1. From another perspective, X[k] is... For interval sampling, draw Points, total draw Here, k is the frequency index, k1 is the number of frequency extractions, and k2 is the number of points extracted in each extraction (decomposing an N-point sequence into k1 and k2). In this case, k = k2 + k1, k1 = 0, 1, …, – 1; k2 = 0, 1, …, – 1. For example Figure 9 As shown.
[0095] Taking a dual-core radix-4 FFT processor as an example, each Radix-4 operation of core 1 uses its four results as inputs for four Radix-4 operations of core 2; core 2's four Radix-4 operations use the results of the four Radix-4 operations of core 1 as inputs. This process is called data transposition, also known as data decimation or commute. Decimation is a computational requirement for cascaded or iterative FFT structures. In some exemplary embodiments, the eight Radix-4 operations shown in Figure 6 can be completed in time-sharing by fewer Radix-4 structures to reduce the number of butterfly operation structures. The computation results of core 1 are output to memory, and core 2 reads them back in the required order to achieve data group exchange.
[0096] The FFT processor built with the aforementioned dual-core radix-4 FFT cores can perform 16-point FFT calculations in a single iteration, 16^k-point FFT calculations in k iterations, and 256-point FFT calculations in two iterations. When this dual-core FFT processor is considered as a single FFT calculation unit, it can also be called a radix-16 FFT calculation unit / structure. With appropriate configuration, this FFT processor can support various FFT point counts, including 2, 4, 8, 16, and 256 points. When applied to radar chips, it can fully meet diverse FFT calculation needs.
[0097] In some exemplary embodiments, a dual-core FFT processor is configured to perform two iterations. In the first iteration, the core operation radix of core 1 is... The number of kernel operations in core 2 In the second iteration, the kernel operation radix of kernel 1... The number of kernel operations in core 2 Then, by using this FFT processor, it can be achieved through two iterations. Point FFT calculation.
[0098] It can be known that the maximum supported cardinality based on each FFT core is The maximum number of parallel inputs required To process each piece of data to be computed into the corresponding FFT core, the chip needs to support simultaneous read / write operations to meet this parallel processing requirement. Therefore, the maximum supported cardinality of the designed and implemented FFT kernel is [number of data points]. The memory access bandwidth within the chip needs to be fully considered. Therefore, based on this constraint, the maximum supported cardinality for each computing core is... However, it cannot be too large. To achieve large-point FFT calculations while maintaining appropriate computational efficiency, a multi-core cascading approach is used for expansion. Each additional FFT core requires a corresponding additional hardware butterfly arithmetic structure. More cores in the FFT processor significantly increase hardware costs and occupy more chip physical space, failing to meet the requirements for chip miniaturization. Taking radar chips as an example, only a few applications may require large-point FFT calculations; designing and implementing a multi-core cascaded FFT processor for this purpose would also result in a waste of hardware resources.
[0099] Based on the configurable FFT processor provided in this disclosure, and using a multi-core FFT processor with a relatively small number of cores, combined with iterative control, the number of FFT calculation points can be further expanded through multiple iterations. This approach avoids wasting hardware resources, effectively controls hardware costs and chip size, and enables large-point FFT calculations, fully meeting the diverse FFT calculation needs in radar chips.
[0100] The various functionalities of radar signal processing, such as pulse compression, range-dimensional FFT, Doppler-dimensional FFT, range-Doppler imaging, spectrum analysis, digital beamforming, constant false alarm rate detection, high-resolution angle estimation, and clutter suppression, all involve FFT calculations. FFT processors used in radar chips include two-stage FFT cores. These two cores can be two identical cores, for example, two 4×4=16 FFT cores, each including a 2-stage radix-4 butterfly arithmetic structure; or two 2×2×2=8 FFT cores, each including a 3-stage radix-2 butterfly arithmetic structure. Alternatively, they can be two different FFT cores, for example, one 4×4=16 FFT core including a 2-stage radix-4 butterfly arithmetic structure, and one 2×2×3=12 FFT core including a 3-stage radix-2, radix-2, radix-3 butterfly arithmetic structure.
[0101] As can be seen, the FFT processor, which supports flexible configuration of the number of computing cores and iterative operation, can meet the FFT calculation needs of multiple functions in the radar chip. The FFT processor is defined as a functional operator, which meets the design requirements of modular and highly parallel radar signal processing on-chip system.
[0102] This disclosure also provides a radar signal processing on-chip system, including: Controller, multiple operators, and memory; The controller is configured to activate at least one of the plurality of operators to perform an operator function. The plurality of operators include: an FFT processor as described in any embodiment of the present disclosure, comprising two cascaded FFT cores, and having at least two operating modes among large-point FFT mode, pipelined mode and independent parallel mode; The FFT processor is configured to retrieve a working mode from the FFT processor configuration data storage area set in the memory according to the operator start instruction from the controller, and perform FFT calculation according to the working mode.
[0103] This disclosure also provides a radar signal processing on-chip system, including: Controller, multiple operators, and memory; The controller is configured to activate at least one of the plurality of operators to perform an operator function. The plurality of operators include: an FFT processor comprising at least two FFT cores cascaded in series, as described in any embodiment of this disclosure; The FFT processor is configured to obtain the actual computational configuration parameters for this iteration from the FFT processor configuration data storage area set in the memory according to the operator start instruction from the controller, and execute the FFT computation for this iteration according to the actual computational configuration parameters.
[0104] In some exemplary embodiments, the FFT processor is further configured to determine the core operation radix of each FFT core in the current iteration calculation based on the actual calculation configuration parameters, and to serially execute the current iteration FFT calculation based on each FFT core with a determined core operation radix.
[0105] In some exemplary embodiments, the FFT processor is further configured to initiate the execution of the current iteration calculation according to the operator start instruction from the controller, obtain the data to be processed in the current iteration from the second iteration data partition set in the memory, and store the FFT calculation result generated in the current iteration into the first iteration data partition set in the memory; The data to be processed in this iteration includes: the initial input data for this iteration, or the data after transposing the FFT calculation results generated in the previous iteration according to the set data transpose rules.
[0106] In some exemplary embodiments, the controller is further configured to determine the number of iterations for performing FFT calculations and the actual calculation configuration parameters of the FFT processor in each iteration based on the number of FFT points to be processed and the computing power parameters of the FFT processor; and to write the actual calculation configuration parameters into the set FFT processor configuration data storage area in the memory.
[0107] In some exemplary embodiments, the controller is further configured to sequentially start the FFT processor to perform iterative calculations for the corresponding rounds based on the determined number of iterations.
[0108] As can be seen, in the on-chip system, the controller calculates the actual computational configuration parameters and writes them to the FFT processor configuration data storage area. The FFT processor reads this data, completes its own configuration, and then executes the corresponding FFT calculation. Based on the number of iterations, the controller controls the FFT processor to perform multiple iterations.
[0109] In some exemplary embodiments, the plurality of operators further includes at least one of the following radar functional operators: pulse compression operator, range-Doppler imaging operator, DBF operator, CFAR operator, synthetic radar aperture imaging operator, and data read / write operator; the number of iterations is w, where w is greater than 1; The controller is configured to sequentially start the FFT processor from the xth to the x+w-1th time period; and start the radar function operator in the yth time period; where y is greater than x+w-1. The FFT processor is configured to be activated in the xth to x+w-1th time period, acquire the data to be processed corresponding to each iteration calculation, execute the FFT calculation function, and write the calculation result into the first iteration data partition or the second iteration data partition set in the memory. The radar function operator is configured to be activated in the z-th time period, obtain the FFT calculation result from the first iteration data partition or the second iteration data partition set in the memory, and execute the subsequent operator functions of this operator; z is greater than y; The durations of each of the aforementioned time periods may be the same or different.
[0110] As can be seen, the controller starts the FFT processor multiple times in the preceding time period to complete the FFT calculation required by the radar function operator, and then controls the start of the corresponding radar function operator to obtain the calculation result data and execute the subsequent operator functions.
[0111] In an optional embodiment, in the dual-core FFT processor, each FFT core includes a twiddle factor multiplier, shifter, conjugate, etc.; the two FFT cores are serially cascaded through a cascade datatranpose and configured to support a series of FFT operations required for radar signal processing. For example, each FFT core may include a Radix 8 (which may include a three-layer Radix butterfly structure), one FFT core may be configured to support a factor-configurable basis, and the other FFT core may be configured not to support a configurable basis.
[0112] For example, this dual-core FFT processor can be configured with a dual-core cascaded iterative FFT architecture that supports factor- and non-factor-configurable radix. One FFT core can support factor-configurable radix 0 / 2 / 4 / 8, while the other FFT core has a fixed radix of 8 for 2^n operations, and a radix of 3... The radix can be fixed to 6 when performing 2^n operations.
[0113] For example, for a cascaded dual-core FFT processor, a single trigger can achieve a maximum FFT of 64pts, while two triggers can achieve a maximum FFT of 4096pts.
[0114] In this application, the twiddle multiplication in the FFT kernel can be used to implement multi-round FFT processing for operations such as windowing and downsampling in radar signal processing. The FFT kernel can also be configured with a shift adjustment factor to adjust the signal amplitude, while the conjugate can be used to implement operations such as IFFT operations.
[0115] It should be noted that, based on the FFT processor described in the embodiments of this application, when performing a hybrid base configuration, the twiddle multiplication in its circuit needs to send data in a preset order. When using an iterative FFT architecture to perform large-scale FFTs, multiple triggers are required.
[0116] In some exemplary embodiments, the radar functional operator further includes one or more of the following: The plurality of operators also includes at least one of the following radar function operators: Multiply-Accumulate Unit (MAC) operator, inversion (INV) operator, Sequential Variational Adaptive (SVA) operator, STATS operator, Constant False Alarm Rate (CFAR) operator, Detection (DET) operator, SIFT operator, Remap operator, and Move operator.
[0117] In some exemplary embodiments, the plurality of operators further include: a shared memory data access PBUF operator, used to access shared memory data to perform FFT processor input data reading and calculation result writing. This operator utilizes the parallel access mechanism of shared memory to achieve parallel read and write of multiple addresses.
[0118] As can be seen, since multiple radar function operators in the radar signal processing on-chip system all require the use of the FFT processor for data calculation, the FFT processor is set as a dedicated operator. The controller controls the startup timing and the access partitioning of the shared memory space by each operator. This can fully improve the utilization rate of the dedicated operator (FFT processor). Under the premise of meeting the radar signal processing efficiency and FFT calculation requirements of the overall on-chip system, the redundant configuration of multiple types of FFT processor hardware is avoided, and the chip cost is effectively controlled.
[0119] In some exemplary embodiments, the radar signal processing system-on-a-chip (SoC) is referred to as radar SoC, such as... Figure 11As shown, the system includes a controller 600, multiple operators: signal processing operators, including a multiply-accumulate unit 610, an FFT processor 620, an INV operator 630, and an SVA operator 640; spectrum analysis operators, including a STATS operator 650, a CFAR operator 660, and a DET operator 660; data processing operators: a SIFT operator 680, a REMAP operator 690, and a MOVE operator 6100; and a memory 700. The connections from each operator to the shared memory 700 and from each operator to the controller 600 are shown in the diagram. Figure 11 Not shown in the diagram, it can be flexibly selected according to the chip architecture. In some exemplary embodiments, the controller is used to control the start-up, stop-run time and order of each operator, also known as a scheduler.
[0120] In some exemplary embodiments, the multiply-accumulate unit has a total of 16 multiply-accumulate units, denoted as M4MAC.
[0121] In some exemplary embodiments, the FFT processor 620 is a radix-8 dual-core FFT processor, denoted as R8-8FFT, which supports a maximum of 64-point FFT calculations in a single iteration.
[0122] In some exemplary embodiments, the radar signal processing system-on-a-chip, also known as a radar SoC chip, includes, but is not limited to, radar chips of various types such as TDM (Time Division Multiplexing), Frequency Modulated Continuous Wave (FMCW), and Phase-modulated Continuous Wave (PMCW).
[0123] This application also provides an integrated circuit, including: a digital signal processing module, which performs FFT calculations based on the FFT calculation method described in any embodiment of this disclosure.
[0124] The digital signal processing module includes one or more processors and a storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the FFT calculation method as described in any embodiment of this application.
[0125] The digital signal processing module further includes: an FFT processor based on any embodiment of the present disclosure.
[0126] This application also provides an electromagnetic wave sensor, including: Carrier; The radar signal processing system-on-a-chip as described in any embodiment of this disclosure, or the integrated circuit as described in any embodiment of this disclosure, is disposed on the carrier. The antenna is mounted on the carrier and is either integrated into a single device with the radar signal processing system-on-a-chip or the integrated circuit, or is disposed separately. The antenna is connected to the radar signal processing system-on-chip or integrated circuit and is used to transmit the electromagnetic wave signal and / or receive the echo signal.
[0127] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.
Claims
1. A Fast Fourier Transform (FFT) processor, characterized in that, The FFT processor includes two cascaded FFT cores and has at least two operating modes among large-point FFT mode, pipelined mode and independent parallel mode. When the FFT processor is in the large-point FFT mode, the two FFT cores are configured to decompose an N-point FFT into two FFTs; When the FFT processor is in the pipelined mode, of the two FFT cores, the cascaded FFT core is configured to perform the first FFT, and the cascaded FFT core is configured to perform the second FFT based on the result of the first FFT; and, When the FFT processor is in the independent parallel mode, the two FFT cores are configured to process data from different channels independently and in parallel.
2. The FFT processor as described in claim 1, characterized in that, The FFT kernel includes at least two computational layers; Between two adjacent layers, the base of the smallest computing unit increases or decreases sequentially along the data.
3. The FFT processor as described in claim 2, characterized in that, At least one of the FFT cores includes at least two computational layers, with some of the computational layers set to be skipped.
4. The FFT processor as described in claim 1, characterized in that, It also includes a buffer located between the two FFT cores; When the FFT processor is in the large-point FFT mode or the pipelined mode, the buffer is configured to transpose the output data of the cascaded FFT core and use it as the input data of the cascaded FFT core. When the FFT processor is in the independent parallel mode, the cache includes two simplex memories to enable the two FFT cores to work in parallel.
5. A Fast Fourier Transform (FFT) processor, characterized in that, include: At least two FFT cores are cascaded in series, and the maximum cardinality supported by the FFT cores is . FFT calculation; The FFT kernel includes at least one butterfly operation structure, and the FFT kernel is configured to operate according to a set kernel radix. Determine the number of butterfly operation structures in this FFT kernel that participate in the FFT calculation and the cardinality of each butterfly operation structure; <= , where i represents the i-th FFT kernel.
6. The FFT processor according to claim 5, characterized in that, The maximum support radix of the FFT kernel ; Where J is the total number of butterfly operation levels included in the i-th FFT kernel. Let be the cardinality of the j-th level butterfly operation structure in the i-th FFT kernel; exist If the FFT core is equal to the base of the first-level butterfly operation structure in this FFT core, or the product of the bases of the multi-level butterfly operation structures, the FFT core is set to determine that the first-level or multi-level butterfly operation structure participates in the execution of this FFT calculation, and bypasses other levels of butterfly operation structures other than the first-level or multi-level. exist If the cardinality of the FFT core is not equal to the cardinality of the first-level butterfly operation structure in this FFT core, nor equal to the product of the cardinality of the multi-level butterfly operation structures, then the FFT core is configured to reconfigure the data flow path of at least one level butterfly operation structure in this FFT core, so that the cardinality of this FFT core is... FFT calculation.
7. The FFT processor according to claim 5 or 6, characterized in that, It also includes at least two core rotation factor storage units corresponding one-to-one with the at least two FFT cores, which respectively store the rotation factors required for performing FFT calculations within each FFT core.
8. The FFT processor according to claim 5 or 6, characterized in that, It also includes: at least one inter-core twitch factor storage unit, each inter-core twitch factor storage unit storing the twitch factor required for serial FFT computation between two adjacent FFT cores; And at least one set of inter-kernel multipliers, the set of inter-kernel multipliers being used to perform multiplication operations of the FFT calculation result of the previous FFT kernel and the corresponding inter-kernel twiddle factor to obtain the data to be processed in the current FFT kernel.
9. The FFT processor according to claim 5 or 6, characterized in that, It also includes: at least one iterative twitch factor storage unit, each iterative twitch factor storage unit storing the twitch factor required for two consecutive iterations of the multi-core FFT processor to perform FFT calculations; And at least one set of iterative multipliers, which are used to perform multiplication operations on the FFT calculation result of the previous iteration and the corresponding iterative twitch factor to obtain the data to be processed in the current iteration.
10. A method for calculating Fast Fourier Transform (FFT), characterized in that, The method includes: Based on the number of FFT points to be processed and the computing power parameters of the FFT processor, the number of iterations for performing the FFT calculation and the actual computing configuration parameters of the FFT processor in each iteration are determined; wherein, the FFT processor includes at least two FFT cores cascaded in series, and the maximum cardinality supported by the FFT core is [missing information]. FFT calculation; Based on the number of iterations, perform the following steps in each iteration: Obtain the actual computational configuration parameters for this iteration and determine the core operation radix for each FFT core. ; According to the nuclear operation base Configure the corresponding FFT core, start the FFT processor, and use the at least two FFT cores to serially execute the FFT calculation for this iteration; Wherein, the kernel operation radix of the FFT kernel Used to determine the number of butterfly operation structures participating in this FFT calculation in this FFT kernel and the cardinality of each butterfly operation structure; The computing power parameters include: maximum supported FFT points. The actual calculation configuration parameters include: the actual number of FFT points calculated. ; ; ; I represents the number of FFT cores. <= , where i represents the i-th FFT kernel.
11. The FFT calculation method according to claim 10, characterized in that, In each iteration, the operation is based on the kernel radix. Configure the corresponding FFT core, including: Based on the kernel operation radix of each FFT kernel The corresponding number of butterfly operation structure levels and the cardinality of each butterfly operation structure involved in this FFT calculation are determined, and the effective first-level or multi-level butterfly operation structure for executing this FFT calculation is configured in each FFT core, thereby determining the effective serial butterfly operation structure for the FFT processor to execute this FFT calculation.
12. The FFT calculation method according to claim 11, characterized in that, In each iteration, activating the FFT processor and serially executing the FFT calculation for that iteration using at least two FFT cores includes: Obtain the kernel rotation factor corresponding one-to-one with the at least two FFT kernels, and use it for FFT calculation within each FFT kernel; Obtain the inter-core rotation factor between adjacent FFT cores in the at least two FFT cores, and use it for serial cascading of the corresponding adjacent FFT cores; The FFT processor is started, the data to be processed in this iteration is obtained, the valid serial butterfly operation structure is input, the FFT calculation is performed according to the intra-core twitch factor and the inter-core twitch factor, and the corresponding FFT calculation result is output.
13. The FFT calculation method according to claim 12, characterized in that, The FFT processor includes two serially cascaded FFT cores: FFT core 1 and FFT core 2. = In the case of FFT, in each iteration, the FFT calculation is performed according to the following formula: ; Where N= , The nuclear rotation factor of FFT kernel 1, The nuclear rotation factor of FFT kernel 2. The internuclear rotation factor between FFT kernel 1 and FFT kernel 2; X represents the FFT data to be processed, N represents the total number of FFT data points to be processed in this iteration by the FFT processor, and k represents the index of the FFT data points to be processed in this iteration. This is the grouping index for the FFT data points to be processed in this iteration. This is the index of the FFT data point to be processed in this iteration within the current group. = 0, 1, …, – 1; = 0, 1, …, – 1.
14. The FFT calculation method according to claim 12 or 13, characterized in that, In each iteration, starting the FFT processor and serially executing the FFT calculation for this iteration using the at least two FFT cores further includes: According to the set data transposition rules, the transposed data is obtained from the calculation result cache as the data to be processed in the current iteration of the FFT processor; wherein, the calculation result cache stores the FFT calculation result of the previous iteration of the FFT processor; And / or, According to the set data transposition rules, the transposed data is obtained from the calculation result cache as the data to be processed in the current iteration of the non-first FFT core among the at least two FFT cores; wherein, the calculation result cache stores the FFT calculation results of the previous FFT core in the current iteration of the serially cascaded FFT core.
15. The FFT calculation method according to claim 14, characterized in that, The calculation result cache includes at least two iteration data partitions. The first iteration data partition is used to store the FFT calculation result of the current iteration of the FFT processor, and the second iteration data partition is used to store the FFT calculation result of the previous iteration of the FFT processor. In two adjacent iterations, the physical cache areas corresponding to the first iteration data partition and the second iteration data partition are swapped. And / or, The calculation result cache includes at least one pair of inter-core data partitions, each pair of inter-core data partitions being a shared data storage area for two connected FFT cores, and the pair of inter-core data partitions includes: a first inter-core data partition and a second inter-core data partition; The first inter-core data partition is used to store the FFT calculation result of the first FFT core in the current iteration of the two FFT cores, and the second inter-core data partition is used to store the FFT calculation result of the first FFT core in the previous iteration of the two FFT cores; in two adjacent iterations, the physical cache regions corresponding to the first inter-core data partition and the second inter-core data partition are swapped.
16. A radar signal processing on-chip system, characterized in that, include: Controller, multiple operators, and memory; The controller is configured to activate at least one of the plurality of operators to perform an operator function. The plurality of operators include: an FFT processor as described in any one of claims 1-4; The FFT processor is configured to retrieve a working mode from the FFT processor configuration data storage area set in the memory according to the operator start instruction from the controller, and perform FFT calculation according to the working mode; or, The plurality of operators include: an FFT processor as described in any one of claims 5-9; The FFT processor is configured to obtain the actual computational configuration parameters for this iteration from the FFT processor configuration data storage area set in the memory according to the operator start instruction from the controller, and execute the FFT computation for this iteration according to the actual computational configuration parameters.
17. The radar signal processing system-on-a-chip according to claim 16, characterized in that, The FFT processor is further configured to initiate the execution of the current iteration calculation according to the operator start instruction from the controller, obtain the data to be processed in the current iteration from the second iteration data partition set in the memory, and store the FFT calculation result generated in the current iteration into the first iteration data partition set in the memory. The data to be processed in this iteration includes: the initial input data for this iteration, or the data after transposing the FFT calculation results generated in the previous iteration according to the set data transpose rules.
18. An integrated circuit, characterized in that, include: A digital signal processing module performs FFT calculations based on the method described in any one of claims 10-15.
19. An electromagnetic wave sensor, characterized in that, include: Carrier; The radar signal processing system-on-a-chip as described in any one of claims 16-17, or the integrated circuit as described in claim 18, is disposed on the carrier. The antenna is mounted on the carrier and is either integrated into a single device with the radar signal processing system-on-a-chip or the integrated circuit, or is disposed separately. The antenna is connected to the radar signal processing system-on-chip or integrated circuit and is used to transmit the electromagnetic wave signal and / or receive the echo signal.
