Fixed-point reversible FFT (Fast Fourier Transform) hardware accelerator design implementation method based on Vitis platform
By combining Split-radix structure and upgrade structure on the Vitis platform, high-level comprehensive design is carried out, which solves the irreversibility and development bottleneck problems in the hardware development of traditional FFT algorithms, and realizes the reversibility and efficient hardware acceleration performance of fixed-point FFT algorithms. The generated IP is platform reusable and portable.
Patent Information
- Application Number
- CN202310664409.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-31
- Publication Date
- 2024-12-03
AI Technical Summary
The hardware development of traditional FFT algorithms has problems such as irreversibility, long development processes, and inability to port IP, making it difficult to achieve efficient hardware acceleration performance.
Adopt a high-level comprehensive design based on the Vitis platform, combined with Split-radix structure and upgrade structure, a hardware accelerator is built through the Vitis platform, and a high-level comprehensive tool is used to optimize the kernel code to realize the hardware accelerator design of fixed-point reversible FFT algorithm.
The reversibility of the fixed-point FFT algorithm is realized, quantization error is reduced, hardware acceleration performance is improved, and the bottleneck of traditional FFT hardware design is broken. The generated IP is platform reusable and portable.
Smart Images

Figure CN119066308A_ABST
Abstract
Description
Technical Field
[0001] The present invention takes FFT, a commonly used algorithm in the field of digital signal processing, as the object, and combines cutting-edge hardware design technologies such as heterogeneous development of Vitis accelerator cards and high-level integrated development, to disclose a high-level integrated hardware accelerator design and implementation of a split-radix structure fixed-point reversible FFT algorithm based on the Vitis platform. The obtained FFT algorithm IP core can realize the reversibility of the fixed-point FFT algorithm. At the same time, the implementation idea can be used to design hardware accelerators with higher development efficiency to meet the requirements under different objectives, pursue more ideal performance realization, and break through the bottleneck of traditional FFT algorithm hardware development and implementation. Background Art
[0002] Today's society is still in the information age. From study and work to food, clothing, housing and transportation, massive amounts of information exist in all aspects of people's daily lives. With the rapid development of 5G-related wireless communication technologies, people have a huge demand for data, which poses a huge challenge to the capabilities of digital processing technology.
[0003] Fast Fourier Transform (FFT), as an efficient and fast implementation mode of Discrete Fourier Transform (DiscreteFourier Transform), has promoted the development of digital signal processing technology. Various forms of FFT structures such as radix-2, radix-4, split-radix FFT and other structures have also been proposed by scholars to meet the data processing needs in different scenarios. Among them, the split-radix algorithm has the least amount of computation, and as the sequence length increases, its computational advantage becomes more obvious. In terms of structure, split-radix is slightly more complex, but it still retains a regular iterative structure.
[0004] In response to the demand for high-speed real-time signal processing, compared with software implementation, FPGA-based design has great competitive advantages in the field of digital signal processing due to its high performance, high flexibility, and online programmability. In engineering practice, floating-point numbers are relatively expensive, and the numbers in the calculation process are often quantized to fixed bits, and the accuracy of the data changes accordingly, resulting in the inability to pursue perfect reconstruction PR (Perfect Reconstruction) of the signal. When fast Fourier transform and fast Fourier transform pair are required as part of the calculation, the calculation error introduced may be large. For some applications that pursue high-precision restoration, this loss is unacceptable. In 2002, Oraintara S et al. proposed the concept of IntFFT, which can be used to construct FFT with integer coefficients and provide a new method to approximate DFT without any multiplication. It has been verified that this transform is completely reconstructed, that is, it can realize the reversible FFT operation of integer to integer mapping; in addition, compared with the traditional fixed-point FFT algorithm (Fxpfft), the number of additions in this algorithm is reduced by 23% to 27%, and the number of shifts is reduced by 41% to 42%, which can achieve higher performance with fewer resources. Method.
[0005] Traditional FPGA hardware acceleration usually requires the use of hardware description languages to design and implement algorithms. The FFT algorithm has a complex calculation process and a large amount of calculation. It is not convenient to rewrite it directly using hardware description languages. The introduction of high-level synthesis tools has largely solved this problem. The emergence and development of the Vivado HLS tool has broken through the implementation bottleneck of previous hardware language design. Compared with the traditional FPGA development process, the development process based on high-level synthesis only requires users to write the corresponding C++ code and add guidance and optimization strategies to the code. After that, with the assistance of the development tool, the hardware circuit described by the code can be implemented. For developers, it saves a lot of time for hardware language development and manual timing design. It also provides more efficient optimization ideas for the implementation of some complex structures, greatly improving the efficiency of FPGA design. Combining this design method with the split-radix algorithm can well make up for the structural disadvantages of the algorithm.
[0006] The Vitis unified software platform launched by Xilinx is a development environment for heterogeneous applications. It helps promote concurrent development and testing of hardware and software elements of heterogeneous applications. The platform has basic hardware and software architecture and application environment, as well as development kits such as Vitis HLS for building applications, analyzing performance bottlenecks, and debugging acceleration algorithms, as well as more than 400 optimized open source acceleration libraries. In addition, the deployed Vitis AI development environment can accelerate AI inference on Alveo accelerator cards, providing comprehensive APIs for pruning, quantization, optimization, and compiling trained networks, thereby achieving the highest AI inference performance for applications deployed by developers and improving development efficiency from multiple angles. Replacing the traditional FPGA development method with the Vitis platform can pursue better hardware performance with a more efficient development method.
[0007] This paper fully explores the structural characteristics of the split-radix integer reversible FFT algorithm, innovatively proposes an accelerator card based on the Vitis platform to perform high-level synthesis design of the algorithm in pursuit of better hardware acceleration performance, and based on this idea proposes a kernel code reconstruction idea, and gives the full process of hardware accelerator design of the algorithm based on the Vitis platform. Summary of the invention
[0008] The technical objectives of the present invention are:
[0009] In order to design a fixed-point FFT algorithm IP core with good performance and reversibility, this paper provides an implementation idea for hardware accelerator design of FFT algorithms with split-radix structure and lifting structure based on Vitis platform using high-level synthesis tools, so as to solve the problems of irreversibility of fixed-point FFT, long design and development process, non-portability of IP, and break through the development bottleneck of traditional FFT algorithm hardware design.
[0010] The present invention realizes the following contents:
[0011] The fixed-point reversible FFT algorithm design based on the Vitis platform designed by the present invention mainly includes three parts. The overall design and development process is as follows: Figure 1 As shown. The first part is to determine the algorithm structure, and select the split-radix structure with less complexity and the lifting structure that can achieve reversibility and reduce the complexity of the algorithm to implement the FFT algorithm. The second part is to build the development environment, and select the Vitis unified platform for heterogeneous development for efficient and high-performance development; the third is the application of high-level comprehensive optimization methods, optimizing the kernel at the software level, and dividing the kernel into blocks for optimization of different performance targets, making development more efficient and convenient, and shortening the development cycle.
[0012] The FFT algorithm structure of the present invention adopts a Split-radix structure and a lifting structure to ensure reversibility. The Split-radix FFT algorithm combines the advantages of the radix-2 FFT algorithm and the radix-4 FFT algorithm. Its basic butterfly structure is as follows Figure 2 As shown, it has the least complexity, and the larger the sequence, the more obvious the advantage. The structure has a regular iterative structure, and the advantages of the structure are magnified through high-level comprehensive and reasonable optimization; the form of the improved structure is as follows Figure 3 As shown in , this structure ensures full reconfigurability in the fixed-point FFT calculation implementation process, and on the other hand, it does not use any multiplication to approximate DFT. The hardware implementation structure can further reduce the implementation resources by replacing the complex multiplication of the rotation factor part of the FFT algorithm with a lifting structure, such as Figure 4 As shown in the figure, IntFFT can realize the mapping from integer to integer. Considering the shift relationship between fixed-point numbers and integers, the mapping from fixed-point to fixed-point numbers can be realized by adding corresponding shift functions before and after the input data. The algorithm is extended to the FFT calculation of fixed-point numbers. Combined with the above discussion, the software implementation idea of the fixed-point reversible FFT algorithm is as follows: Figure 5 shown.
[0013] The present invention uses the heterogeneous development environment Vitis platform provided by Xilinx to build a hardware acceleration design environment, and the construction method is as follows Figure 6 The build program consists of host and kernel files. The host program uses OpenCL API to build a heterogeneous system, including device selection, kernel parameter setting, data transmission, testing, etc. The kernel file is divided into fft.cpp and ifft.cpp, which respectively implement the FFT transform and its inverse transform functions; Based on the Alveo U200 data center accelerator card, the hardware acceleration environment is built. The process of using Vitis for accelerator design and development is as follows Figure 7 shown.
[0014] The present invention performs optimization design of hardware accelerators based on high-level synthesis. First, the kernel code is reconstructed in combination with algorithm theory and the requirements of high-level synthesis on software algorithms. The code style can be optimized from the perspectives of minimizing circuit mapping resources, the requirements of high-level synthesis instructions on code structure, and parallelization of reconstructed code. The optimized code is optimized using the HLS tool provided by the Vitis development core suite. With latency, throughput, and resources as the goals, pragma (hardware optimization instructions provided by HLS) is optimized for implementation structure, function, loop, data type, etc. The implementation results are compared and analyzed through operations such as compilation, simulation, and running applications. According to the optimization requirements, the target module is continuously optimized with high-level synthesis. Finally, the designed hardware accelerator is generated into a portable, platform-reusable fixed-point reversible FFT algorithm IP core. The high-level synthesis design process is as follows: Figure 8 shown.
[0015] Compared with the existing FPGA-based FFT hardware implementation, the present invention has the following advantages:
[0016] 1. The implemented IP can realize the reversibility of the fixed-point FFT algorithm
[0017] The application of the lifting structure makes the quantization in the fixed-point calculation process occur in the lifting coefficients instead of directly quantizing the node data. For IFFT transformation, the corresponding steps are still operated with the same coefficients but opposite signs. Therefore, the previous input data can be completely reconstructed without reconstruction errors due to quantization. It can replace other fixed-point FFT IPs in application scenarios with higher reconstruction accuracy.
[0018] 2. The design optimization ideas provided break the bottleneck of traditional FFT FPGA design methods and have greater optimization space.
[0019] First, in terms of the choice of algorithm structure, the split-radix structure has fewer multiplications, but its structure is relatively complex. Combining high-level synthesis tools to reconstruct the kernel at the software level can complement each other's strengths and weaknesses and achieve better performance with fewer resources. The high-performance accelerator card and Vitis AI development environment provided by the Vitis platform also provide greater optimization space for the design of hardware accelerators.
[0020] 3. This method is more efficient in accelerator hardware design and the generated IP is more applicable.
[0021] The Vitis platform provides a development environment for software and hardware collaboration. It has a framework specifically designed for hardware accelerators and provides core development kits, acceleration libraries, etc. to improve development efficiency. In addition, hardware optimization using high-level synthesis can reconstruct the kernel and add control instructions in a more convenient way at the software level. The IP designed based on the Vitis platform has platform reconfigurability and portability. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 Design and implementation ideas of overall reversible fixed-point FFT accelerator
[0023] Figure 2 The basic L-type structure of the split-radix FFT
[0024] Figure 3 Lifting structure that satisfies reversibility
[0025] Figure 4 Illustration of a butterfly structure that replaces complex multiplication with a lifting structure
[0026] Figure 5 Fixed-point reversible FFT software algorithm implementation ideas
[0027] Figure 6 Hardware accelerator design environment construction process based on Vitis platform
[0028] Figure 7 Hardware accelerator design process based on Vitis platform
[0029] Figure 8 High-Level Synthesis-Based Optimization Flow DETAILED DESCRIPTION
[0030] The following describes in detail the fixed-point reversible FFT hardware accelerator design and implementation method based on the Vitis platform provided by the present invention in conjunction with the accompanying drawings, but does not constitute a limitation to the present invention.
[0031] The present invention is an optimized design and implementation of a fixed-point reversible FFT hardware accelerator based on the Vitis platform. The entire implementation process is as follows: Figure 1 As shown below, combined Figure 1 The detailed implementation methods of each module in the process are introduced in detail.
[0032] Step 1: Software code composition of the algorithm
[0033] To implement the FFT algorithm, the basic structure split-radix FFT structure is selected. This structure is equivalent to radix-2-FFT and radix-4FFT and is a type of Cooley-Turkey FFT algorithm (also known as butterfly algorithm). Split-radixFFT divides X(k) (input sequence) into four subsequences X(4r), X(4r+1), X(4r+2) and X(4r+3), merges the first and third subsequences and calculates them using radix-2FFT, and the second and fourth subsequences using radix-4FFT. The formula for the split-radix FFT algorithm is derived as follows: Decompose X(k) using the DIF method to divide it into four subsequences X(4r), X(4r+1), X(4r+2) and X(4r+3), merge X(4r) and X(4r+2), and get the value from the radix-2DIF-FFT calculation formula.
[0034]
[0035] in From the base-4DIF-FFT calculation formula, we get
[0036]
[0037] In the above formula From (1), we can see that X(2r) can be obtained by the sequence The N / 2 point DFT is obtained, where The definition is as follows
[0038]
[0039] Where n = 0, 1, ..., N / 2, let
[0040]
[0041] Similarly, X(4r+1) and X(4r+3) in the sequence can be obtained by and The N / 4-point DFT of the two sequences is obtained, where and The definition is as follows
[0042]
[0043] The basic calculation unit of split-radix FFT calculation is obtained as follows Figure 2As shown in the figure. The Split-radix algorithm cleverly combines the advantages of the radix-2FFT and radix-4FFT algorithms. Compared with the two, it has the least multiplication complexity, and the larger the sequence length, the more obvious the advantage. The structure also retains the regular iterative structure. Using split-radix as the basic structure of the hardware accelerator design for subsequent optimization design can pursue a larger performance optimization space.
[0044] In addition, in order to achieve the reversibility of the algorithm function, the lifting structure is used to replace the rotation factor in the FFT calculation process. For example, Figure 3 As shown, in the dual-channel system in the figure, the two inputs x 0 and x 1 is fed to the inclusion operator A 0 and A 1 In the two-port network, with A 0 The first branch of the operation is called dual-lifting, and the second branch is called dual-lifting. 1 The operation is called lifting. It can be seen that for any A 0 and A 1 This system is fully reconfigurable. No matter they are nonlinear, have rounding operations, or round up or down, they will not destroy the full reconfigurability of the system, that is, reversibility. Assume that a is a complex number with an amplitude of 1, that is, a = c + js, and assume that a complex number x = x r +jx i ,but
[0045] y=ax=(cx r -sx i )+j(cx i +sx r ) (6)
[0046] It can be expressed in phasor matrix form as follows
[0047]
[0048] in
[0049] The magnitude of a is 1 if and only if R is an orthogonal matrix. Assume that c 2 +s 2 = 1, where s ≠ 0 (if s = 0, then a = 1 or a = -1, in which case no quantization is required), the matrix R can be decomposed into three lifting steps
[0050]
[0051] like Figure 4As shown, the complex multiplication structure of the rotation factors in the FFT operation process is decomposed into a lifting structure. Such a structure can bring reversibility to the operation. In addition, the number of decomposition multiplications is reduced from the original four to three. Although the number of real additions increases from two to three, the reduction in the number of multiplications has a greater impact on resources in hardware design. Taking such a structure as the hardware design object will also pursue better hardware performance; such a structure can also realize integer-to-integer mapping. The input fixed-point numbers and rotation factors in the calculation process can be shifted and quantized into integers, and the reversible FFT algorithm can be applied to fixed-point calculations. The software implementation process of the algorithm is as follows Figure 5 shown.
[0052] Step 2: Build a hardware acceleration design implementation environment based on the Vitis platform
[0053] Based on the Vitis platform to achieve application acceleration, the software program is split into a host program running on the CPU and a computing function or kernel running on the Alveo data center accelerator card. The overall design environment is constructed as follows: Figure 6 The host program controls the hardware acceleration platform as a whole. You can use Xilinx's OpenCL API extension to write the host program in native C++ language. The overall idea is as follows Figure 6Left half: In the host program, you can set the acceleration device ID, kernel parameters, etc., and transfer data with the kernel through PCLe and global memory. In addition, the host also plays an important role in verifying the correctness of kernel functions. First, it is necessary to allocate memory space in the host memory for the prepared input and test data, and complete the preparation of the data required by the corresponding testbench; the next step is to set up the OpenCL environment, specify the accelerator device ID and load the device binary .xclbin file, which contains the bitstream and other metadata required for FPGA programming, and is used by the XRT API to actually complete the FPGA programming. It also includes obtaining platform information in the system, specifying the accelerator device ID, loading the device binary file, creating OpenCL context, program, command queue, etc., to complete the corresponding heterogeneous design environment design; the next step is to allocate cache in the global memory, establish a data transmission bridge associated with the kernel, and achieve efficient cache and communication from the host to the device; set the kernel parameters. After the software application identifies the kernel executed on the device and allocates the cache, it needs to define the actual parameters to establish an association with the kernel. These parameters will also be integrated into the corresponding interface in the subsequent process; after completing these preparations, the kernel and the hardware accelerator can be transmitted, including the input of kernel data, the control of kernel startup, and the reading back of kernel data after the calculation is completed; finally, the host completes the verification of the read-back data through the designed testbench test platform, thereby verifying the kernel function.
[0054] For kernel files, by analyzing the integer reversible FFT algorithm, the iterative loop of the L-shaped structure occupies most of the resources and computing time, which is the main bottleneck of performance optimization. This part can be designed as a kernel with the rest of the algorithm for acceleration. At the same time, considering the functions implemented by the algorithm, two kernels, intfft.cpp and intifft.cpp, are created. The kernel is essentially a customized data path (optimized for the desired function) and the associated data storage and mobile network. The kernel code is required to be constructed in the "load-compute-store" mode. The overall idea is as follows Figure 6 The right half, where fft.cpp and ifft.cpp are Figure 5 Software implementation of the algorithm construction idea. After creating the host and kernel programs, the overall acceleration environment can be built and the subsequent optimization process can be carried out.
[0055] Step 3: Refactoring of kernel code
[0056] For accelerator cores, to achieve acceptable quality of results, the original software code still needs to be rewritten to help the Vitis HLS tool achieve the desired performance goals. Generally speaking, the structure produced by the original code of the software algorithm is less efficient than that of an ordinary CPU program. Even if pipelining and expansion are used, it will not play a big role. Therefore, it is necessary to reconstruct an equivalent algorithm based on hardware principles but suitable for high-level synthesis. The reconstructed code needs to correspond well with the hardware. While analyzing the principle of the reversible FFT algorithm, the implementation form of the hardware structure should be considered. Figure 5 Taking the L-type structure in as an example, multiple nested loops are usually used in the implementation of software algorithms, and FFT calculations are performed in situ, that is, the data between each level uses the same memory and is stored in the form of memory overlay. This implementation method has neat code and strong readability in the software program. At the same time, the in-situ calculation method can also save a lot of storage space; one of the very important hardware optimization methods in the subsequent Vitis HLS optimization is task pipelining, which divides the algorithm into multiple stages and uses the HLS tool to generate an implementation method that allows different stages of the algorithm to run on different data sets. It can be seen that the previously implemented L-type module method is not suitable for this optimization, so it is necessary to reconstruct the algorithm of the calculation module and split the M-1 level of the L-type structure of the FFT algorithm, where M = log 2 N, N is the number of FFT points. Each level of calculation is packaged into a function separately, that is, each level of L-type calculation is divided into a separate task. The refactored code allows the execution of task pipeline operations here, that is, it supports the second task to be started after the first task completes only the first level of L-type calculation, instead of waiting for the first task to complete all functions in the code before starting the second task. In addition, the lifting coefficients can also be optimized. In the reversible FFT algorithm implemented by split-radix, the multiplication factors of the butterfly orthogonal transform are 1, -1, j, and -j. They are all simple sign transformations in the calculation process, and the lifting structure is the main multiplication operation. The complex lifting coefficients consume more resources in terms of complexity and structural implementation, but these lifting coefficients are regular and can be calculated in advance using MATLAB software. The functional modules can be replaced by lookup tables, which can save a lot of computing resources and greatly improve latency. When calculating the number of FFT points, there is a dependency between the lifting coefficient calculation and the DFT calculation, including the calculation of rotation factors and the acquisition of lifting coefficients through rotation factors. If a lot of lifting factor calculation time can be saved before implementing DFT calculations, the hardware computing speed can be greatly improved.
[0057] Step 4: Optimize using the high-level synthesis tool Vitis
[0058] After building the hardware acceleration design platform, the Vitis core development kit provided by the Vitis platform is used for optimization design. The high-level synthesis tool Vitis HLS is mainly used to map the software code into RTL language. The specific optimization design process is as follows: Figure 8 As shown, the software code that needs to be accelerated is written and refactored. The specific reconstruction idea is described in step 3. C simulation is performed at the software level to verify the correctness of the software code function. If the verification fails, debugging and modification are performed until the C simulation is passed. Then, the HLS tool is used for C language synthesis, and the synthesized kernel is verified by C / RTL joint simulation. If an error occurs, the code is debugged again and the C simulation step is returned. If the joint simulation result is correct, the performance judgment link is entered. If the performance does not meet the target, the C language synthesis step is returned to the HLS optimization design again. The high-level synthesis tool HLS provides many optimization strategies. HLS optimization mainly uses the parallelism of FPGA to shorten the delay. The main optimization objects are loops, functions, data flows, arrays, and interfaces, etc., and loops are expanded (Unroll), Merge, flatten, pipeline and other strategies are used for optimization. Array optimization mainly involves array partitioning and array reshaping. These operations can be used to perform better optimization configuration on RAM to make it more efficient. Reasonable array allocation can greatly save resources and greatly improve data transmission efficiency. Function data flow, interface, data bit width, etc. are also optimized. By adjusting the strategy, we pursue a performance design that meets the target, and directly convert the designed C model description into an RTL-level description, which saves a lot of time for hardware language development and manual timing design, greatly improving the efficiency of FPGA design. The design process ends until the performance is met, and the Vitis HLS optimization design process ends. The IP blocks generated with the build target can be applied to FFT calculation scenarios that require integer mapping and reversibility, such as medical imaging, image compression, etc.
[0059] Based on the spliti-radix FFT algorithm, the present invention proposes a hardware accelerator card design based on the Vitis platform to implement the entire process, and analyzes from multiple angles such as algorithm structure, acceleration platform selection, and high-level integrated design to pursue better hardware acceleration performance. The FFTIP designed by the present invention can realize the reversibility of the algorithm and has the advantages of platform reusability and portability; the design process is more efficient and the development cycle is short, breaking the performance and implementation bottleneck of traditional FFT hardware design.
[0060] The above embodiments are only used to illustrate the method of the present invention rather than to limit it. A person skilled in the art may modify or make equivalent substitutions to the method of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention shall be subject to the claims.
Claims
1. A design and implementation method of a fixed-point reversible FFT hardware accelerator based on the Vitis platform, including the selection and implementation of an algorithm structure, the construction of a development environment, and the application of a high-level synthesis optimization method, characterized in that: In the selection of algorithm structure, the split-radix structure with less complexity and the lifting structure that can achieve reversibility and reduce algorithm complexity are selected to implement the FFT algorithm; in the design of the hardware acceleration environment, the Vitis unified platform for heterogeneous development is selected for efficient and high-performance development; in the process of performance optimization, the high-level synthesis tool provided by Vitis is selected. First, the kernel is optimized at the software level to improve hardware performance, and then the kernel is divided into blocks for high-level comprehensive optimization with different performance goals, which makes development more efficient and convenient and shortens the development cycle.
2. The structure of selecting the split-radix structure as claimed in claim 1 as the structure for implementing the algorithm, characterized in that: The Split-radix algorithm divides the sequence into four subsequences. The first and third subsequences are combined and calculated using the radix-2 FFT method, and the second and fourth subsequences are calculated using the radix-4 FFT method. This structure cleverly combines the advantages of the radix-2 FFT and radix-4 FFT algorithms. Compared with the two, it has the least multiplication complexity, and the larger the sequence length, the more obvious the advantage. This structure also retains the regular iterative structure. Split-radix FFT is selected as the basic structure to pursue greater performance optimization space for the FFT algorithm.
3. The selective lifting structure of claim 1 is used to achieve reversibility and reduce algorithm complexity, characterized in that: The complex multiplication structure of the rotation factors in the FFT operation process is decomposed into a lifting structure. Such a structure can bring reversibility to the operation. In addition, the number of decomposition multiplications is reduced from four to three, and the number of real additions is increased from two to three. However, the reduction in the number of multiplications has a greater impact on resources in hardware design. Taking such a structure as the hardware design object will also pursue better hardware performance; such a structure can also realize integer-to-integer mapping. The input fixed-point numbers and rotation factors in the calculation process can be shifted and quantized into integers, and the reversible FFT algorithm can be applied to fixed-point calculations.
4. The Vitis unified platform for heterogeneous development as claimed in claim 1 is used to develop efficient and high-performance hardware accelerators, characterized in that: The Vitis unified software platform is designed for the development of heterogeneous applications. It decomposes software programs into host and kernel programs to promote concurrent development and testing of the hardware and software elements of heterogeneous applications. The platform has a basic hardware and software architecture and application environment, as well as development kits such as Vitis HLS for building applications, analyzing performance bottlenecks, and debugging acceleration algorithms to help optimize performance efficiently. In addition, there are more than 400 optimized open source acceleration libraries. The deployed Vitis AI development environment can achieve the highest AI inference performance for applications deployed by developers on Alveo accelerator cards, improve development efficiency from multiple angles, and pursue better hardware performance with a more efficient development method. The accelerator IP designed in this way has the characteristics of portability and platform reusability.
5. In the performance optimization process described in claim 4, a high-level synthesis tool provided by Vitis is selected, characterized in that: The high-level synthesis tool Vitis HLS optimizes the implementation structure, function, loop, data type, etc. with pragma (hardware optimization instructions provided by HLS) based on latency, throughput, and resources. It compares and analyzes the implementation results through operations such as compilation, simulation, and running applications. According to the optimization requirements, it continuously performs corresponding high-level synthesis optimization on the target module. It directly converts the C model description into an RTL-level description, saving a lot of time for hardware language development and manual timing design, greatly improving the efficiency of FPGA design.
6. The method of improving hardware performance by optimizing the kernel at the software level according to claim 1, characterized in that: The original software code is rewritten in software, and the algorithm operation mode is rewritten from the software level to help the Vitis HLS tool achieve the desired performance goals. For the L-type structure, an equivalent implementation form suitable for high-level synthesis is reconstructed based on hardware principles, and each level of L-type calculation is divided into a separate task to support the second task to start after the first task completes only the first level of L-type calculation, instead of waiting for the first task to complete all functions in the code before starting the second task; the original complex boost coefficient calculation process is replaced by a lookup table to reduce resource usage, save boost coefficient calculation time, and improve hardware calculation speed.
Citation Information
Cited By
FPGA adjoint method accelerator design method for fast reverse design
CN119623216A
