GPU-based parallel computing method, system, device and medium for flow-induced noise prediction

Abstract

The present application belongs to the technical field of flow-induced noise prediction, and discloses a flow-induced noise prediction method, system, device and medium based on GPU parallel computing, comprising: obtaining unsteady flow field data; calculating the sound source term of each integral surface unit at each time on the CPU end; based on the video memory capacity of the GPU end, the integral surface unit is divided into several data blocks, and a linear storage layout corresponding to the data block is constructed; inputting the current data block into the GPU end to calculate the acoustic contribution, obtaining the corresponding local calculation result, using asynchronous data transmission and multi-flow mechanism, the local calculation result of the current data block is transmitted back to the CPU end, and the next data block is transmitted into the GPU for parallel calculation; the results of each data block are sequentially spliced and aligned on the CPU end, and the noise prediction results of each observation point are output. The present application can significantly improve the efficiency of the integral stage without relying on multiple nodes, thereby realizing efficient prediction of flow-induced noise.

Description

Technical Field

[0001] This invention belongs to the field of flow-induced noise prediction technology, and in particular relates to a flow-induced noise prediction method, system, device and medium based on GPU parallel computing. Background Technology

[0002] Flow-induced noise is widely present in engineering scenarios such as aircraft takeoff and landing, propeller propulsion, underwater vehicles, and drones. Its accurate prediction is a crucial foundation for aircraft aerodynamic optimization, propeller noise reduction design, and environmental noise assessment. In practical engineering, flow-induced noise prediction often employs a hybrid approach of "CFD (Computational Fluid Dynamics) and acoustic analogy," followed by prediction of far-field noise using acoustic analogy theories such as FW-H (Fowcs Williams-Hawkings). Because this method requires integrating the acoustic source term over a large number of integral surface elements and a long time history, and solving for the acoustic response at multiple observation points, its computational complexity increases approximately multiplicatively with "time × integral surface × observation point," resulting in a significant demand for computational resources.

[0003] Traditional solutions primarily rely on the parallel computing capabilities of multi-core CPUs or CPU clusters, improving efficiency by dividing the integral surface or observation points. However, limited by the floating-point throughput and memory bandwidth of a single CPU core, even with tens or hundreds of cores running in parallel, the FW-H integration stage remains a bottleneck in the flow-induced noise prediction process, making it difficult to meet the needs of engineering-grade ultra-large-scale computational examples.

[0004] In recent years, with the continuous development of graphics processing units (GPUs), GPUs possess massive thread parallelism and high memory bandwidth, making them suitable for parallel computation of flow-induced noise (FW-H) integration. However, achieving high efficiency under conditions of limited single-card memory and bandwidth sensitivity requires targeted design in areas such as data layout, thread mapping, parallel reduction, and asynchronous pipelining. To address this, this invention proposes a FW-H parallel computation method for single-machine, single-GPU systems, significantly improving the efficiency of the integration stage without relying on multiple nodes, thereby achieving efficient prediction of flow-induced noise. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device, and medium for predicting flow-induced noise based on GPU parallel computing, so as to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a flow-induced noise prediction method based on GPU parallel computing, comprising: S1. Data Acquisition and Source Term Construction: Obtain unsteady flow field data from CFD and / or experiments, calculate the thickness source and load source in the FW-H acoustic analogy, and establish a unified index for time steps, integral surface elements and observation points; obtain the three-dimensional index space. S2. Blocking and Data Preparation: Based on the GPU memory capacity and bandwidth, the index space is divided into blocks in the time dimension and / or the integral surface dimension, and a linear storage layout corresponding to the blocks is constructed. S3, GPU Thread Mapping and Parallel Integration: Map two or three of the time step index, integration surface cell index and observation point index to the GPU's three-dimensional thread grid, enabling a single thread to calculate the acoustic contribution of an integration surface cell at a time step to an observation point. S4. Intra-block parallel accumulation and delay time alignment: On the GPU side, the local contributions of the same observation point or the same integral surface unit are reduced in parallel, and the delay time is discretized and aligned based on the propagation distance and sound speed to obtain the local result of the data block. S5. Asynchronous transmission and pipelined processing: Utilizing asynchronous data transmission and multi-streaming mechanisms, the next data block is passed to the GPU for parallel computation, while the partial results of the previous data block are returned to the CPU. S6. CPU-side splicing and output: The CPU performs sequence splicing and alignment correction on the results of each data block, and outputs the time-domain sound pressure history and / or frequency-domain sound pressure level spectrum of each observation point.

[0007] Optionally, step S1 includes: (1) Calculate the normal of the integral surface element, the geometric relationship between the area and the observation point, the density of the medium and the speed of sound; (2) Calculate the delay time based on the propagation distance and the speed of sound and then discretize and map it to the source time step index to achieve delay time alignment.

[0008] Optionally, the linear storage layout in step S2 adopts a time-continuous, integral surface / observation point block approach, so that the index mapped to the thread grid time direction corresponds to the continuous access direction of the video memory, thereby achieving merged memory access.

[0009] Optionally, when constructing the three-dimensional thread mesh in step S3, the mesh dimensions satisfy: gridDim.x=Nt / nx, gridDim.y=Ne / ny, gridDim.z=No / nz; In the formula, gridDim.x represents the dimension of the thread grid in the x-direction, gridDim.y represents the dimension of the thread grid in the y-direction, gridDim.z represents the dimension of the thread grid in the z-direction, Nt, Ne, and No represent the time steps, number of integral surface units, and number of observation points of this data block, respectively, and nx, ny, and nz represent the number of threads in the three-dimensional direction within the thread block, with nx being an integer multiple of the width of the underlying parallel execution unit.

[0010] Optionally, the computation of acoustic contribution includes mapping the time step, integral surface unit, and observation point index to the thread grid of the GPU, so that a single thread calculates the acoustic contribution of an integral surface unit to an observation point at a time step, and performs parallel reduction on the local acoustic contribution of the same observation point or the same integral surface unit on the GPU to obtain the local computation result corresponding to the current data block. Using asynchronous data transmission and multi-stream mechanism, the local computation result of the current data block is sent back to the CPU, and the next data block is sent to the GPU for parallel computation. Optionally, in step S4, the parallel reduction adopts a combination of shared memory caching and intra-block tree / segmented reduction, and the numerical precision is controlled during the accumulation process, using a double-precision accumulation and / or compensation summation strategy.

[0011] Optionally, step S5 employs multi-stream (multi-queue) asynchronous pipelined and page-locked host memory to achieve overlap between host-device data transmission and GPU kernel function computation.

[0012] Optionally, for cases where the video memory is insufficient to accommodate the complete index space, the process is divided into multiple time windows in the time dimension, and time delay alignment and cross-window accumulation are performed at the window boundaries to ensure that the result is equivalent to the overall integral value.

[0013] Optionally, for the three-dimensional flow-induced noise prediction task, FW-H time-domain integration is used; while for the two-dimensional flow-induced noise prediction task, FW-H frequency-domain integration is used, and the frequency-domain sound pressure level spectrum is output in step S6.

[0014] Optionally, constant or slowly varying observation point coordinates and medium parameters can be stored in the GPU constant cache or read-only cache to reduce global memory access overhead.

[0015] Optionally, the thread block size and register usage limit can be automatically selected based on GPU architecture parameters to improve utilization, and the block size can be adaptively adjusted to maximize throughput and bandwidth utilization. If necessary, fused multiplication can be enabled for critical kernels to reduce rounding errors.

[0016] This invention also provides a flow-induced noise prediction system based on GPU parallel computing, comprising: The data acquisition module is used to acquire unsteady flow field data; it defines integral surface elements, calculates the sound source term of each integral surface element at each time step on the CPU, and uniformly numbers and establishes a unified index for time steps, integral surface elements, and observation points; wherein, the sound source term includes thickness source and load source; it calculates the propagation distance between the integral surface element and the observation point, calculates the delay time based on the propagation distance and sound speed, and discretizes the delay time to the source time step index to achieve delay time alignment; The block partitioning module is used to divide the integral surface unit or time step into several data blocks according to the GPU's memory capacity, and construct a linear storage layout corresponding to the data blocks. The in-block data processing module is used to input the current data block into the GPU to calculate the acoustic contribution. The calculation of the acoustic contribution includes mapping the time step, integral surface unit, and observation point index to the GPU's thread grid, so that a single thread calculates the acoustic contribution of an integral surface unit to an observation point at a time step, and performs parallel reduction on the local acoustic contribution of the same observation point or the same integral surface unit on the GPU to obtain the local calculation result corresponding to the current data block. Using asynchronous data transmission and multi-stream mechanism, the local calculation result of the current data block is transmitted back to the CPU, and the next data block is input into the GPU for parallel calculation. The data output module is used to perform sequence splicing of the results of each data block on the CPU based on the unified index of observation points and time steps, and to perform alignment correction of the results of each data block according to the delay time alignment rules, and output the noise prediction results of each observation point.

[0017] The present invention also provides an electronic device, comprising: CPU and GPU that are interconnected at high speed with said CPU; The memory is used to store flow field data, acoustic source terms, and program code; The CPU and GPU work together to execute the stored program code to implement the aforementioned flow-induced noise prediction method based on GPU parallel computing.

[0018] The present invention also provides a computer-readable storage medium storing a computer program that, when executed collaboratively by a CPU and a GPU, implements the aforementioned method for predicting flow-induced noise based on GPU parallel computing.

[0019] The technical effects of this invention are as follows: This invention first obtains unsteady flow field data from CFD or experiments, calculates acoustic source terms such as thickness and load sources in the FW-H acoustic analogy on the CPU, and establishes a unified index for time, integral surface, and observation points. Then, based on memory capacity and bandwidth constraints, the index space is block-based, and the "time step—integral surface unit—observation point" is mapped to GPU threads using a fixed or adaptive thread grid. Merged memory access, shared memory caching, and intra-block reduction are used to calculate the acoustic contribution of a single data block. After delay time alignment and parallel accumulation at the block level, asynchronous data transmission and a multi-stream mechanism are used to overlap between the host and GPU until all data blocks are covered. Finally, sequence alignment and splicing are performed on the CPU to obtain the time-domain or frequency-domain sound pressure level results at the observation points. This invention fully utilizes the parallel capabilities of a single GPU, significantly improving the computational efficiency and memory bandwidth utilization of the FW-H integration stage, and has the advantages of clear structure, simple implementation, and easy engineering integration. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram illustrating the data interaction and processing between the CPU and GPU in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the implementation of an embodiment of the present invention. Detailed Implementation

[0022] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as a limitation of the present invention, but rather as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0023] It should be understood that the terminology used in this invention is merely for describing particular embodiments and is not intended to limit the invention. Furthermore, with respect to numerical ranges in this invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Every smaller range between any stated value or intermediate value within a stated range, and any other stated value or intermediate value within said range, is also included in this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0024] Various modifications and variations can be made to the specific embodiments described in this specification without departing from the scope or spirit of the invention, as will be apparent to those skilled in the art. Other embodiments derived from this specification will also be obvious to those skilled in the art. This application specification and embodiments are merely exemplary.

[0025] The terms “include,” “including,” “have,” “contain,” etc., used in this article are all open-ended terms, meaning that they include but are not limited to.

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments. Example

[0027] like Figure 1 - Figure 2 As shown, this embodiment provides a flow-induced noise prediction method based on GPU parallel computing, including: acquiring unsteady flow field data; defining integral surface units, calculating the sound source term of each integral surface unit at each time step on the CPU, and uniformly numbering and indexing the time step, integral surface units, and observation points; wherein, the sound source term includes thickness sources and load sources; dividing the integral surface units or time steps into several data blocks based on the GPU's memory capacity, and constructing a linear storage layout corresponding to the data blocks; inputting the current data block into the GPU to calculate the acoustic contribution, wherein calculating the acoustic contribution includes mapping the indices of the time step, integral surface units, and observation points. The threaded grid of the GPU enables a single thread to calculate the acoustic contribution of an integral surface cell to an observation point at a time step. On the GPU, the local acoustic contributions of the same observation point or the same integral surface cell are reduced in parallel to obtain the local calculation result corresponding to the current data block. Using asynchronous data transmission and multi-streaming mechanisms, the local calculation result of the current data block is sent back to the CPU, and the next data block is sent to the GPU for parallel calculation. Based on the unified index of the observation point and time step, the results of each data block are sequentially concatenated on the CPU, and the results of each data block are aligned and corrected according to the delay time alignment rules to output the noise prediction result of each observation point.

[0028] To address the problems existing in the prior art, this embodiment provides a GPU parallel computing method for flow-induced noise prediction, solving the problems of high computational load, low parallel efficiency, and data communication between the CPU and GPU in the integration stage of noise prediction. While ensuring computational accuracy, it reduces the number of communications between the CPU and GPU, lowers communication costs, and improves parallel computing efficiency and data throughput. The specific implementation process of this embodiment includes: (1) Read the pressure, velocity, density and other data of the integral surface output by CFD over time at the CPU end, and calculate the FW-H thickness source and load source; (2) Number the time step, integration surface unit and observation point uniformly, calculate the geometric relationship, propagation distance and sound speed between the observation point and integration surface unit, generate a mapping table from delay time to source time step index, and generate a mapping table from delay time to time step index. (3) Set the time dimension window size W and the integral surface size E according to the GPU memory capacity to form a data block {W×E×No}; (4) Construct the thread block size (nx, ny, nz), where nx is an integer multiple of the width of the underlying parallel execution unit, and calculate the grid dimension accordingly; (5) In the kernel function, each thread reads the source terms and geometric quantities at the corresponding time step and integral surface unit from the global memory in the merge memory access method, caches the intermediate quantities of the same observation point or the same unit in the shared memory, and completes the block accumulation by tree reduction; (6) Perform intra-block alignment and local result output based on the delay time mapping table; (7) Use two or more CUDA streams: Stream A transmits the next block of input data, Stream B performs the integration of the current block, and Stream C returns the result of the previous block; (8) The CPU end splices all the blocks according to the observation point number and time series to obtain the sound pressure time history of each observation point.

[0029] It is feasible to achieve overlap between data transfer and kernel function computation through multi-stream and page-locked memory.

[0030] The feasible frequency domain calculation process is as follows: after step (1), perform a fast Fourier transform on the source time signal, change the kernel function to frequency domain contribution calculation, and keep the block reduction and alignment consistent; the final output is the sound pressure level spectrum at the observation point.

[0031] Time windowing and cross-window alignment: For extremely large time histories, windowing is used, and cross-window alignment and accumulation are performed to ensure equivalence with the overall integral value. Specifically, this includes: When the time step is too large to be loaded into the video memory at once, the time step is divided into multiple windows {W1, W2, ...}. Steps (3) to (6) are executed independently in each window. The overlapping area is trimmed / compensated and accumulated at the window boundary according to the delay time alignment rule to ensure consistency with the overall integral.

[0032] Compared with existing CPU implementations, this embodiment significantly shortens the computation time of the FW-H integration stage under single-card conditions; alleviates memory limitations and improves bandwidth utilization through blockization and asynchronous pipelining; achieves stable and efficient solutions for large-scale examples without using multiple nodes; and uses unified indexing and latency alignment to ensure that the numerical results are consistent with the overall integration.

[0033] The specific application example of this embodiment is as follows: Based on the scheme in this embodiment, a flow around a cylinder was selected as a case study, and calculations were performed for the number of integral surface elements of 5000, 10000, 20000, and 40000, respectively. The acceleration effect under different mesh sizes was compared. Table 1 below compares the computation time under different mesh sizes.

[0034] As shown in Table 1, in the case with 5000 grids, the computation time of a single-core CPU is 32107.03s, while the computation time using a single GPU is 54.37s, resulting in a speedup of 590.52. In the case with 40000 grids, the computation time of a single-core CPU is 257990.52s, while the computation time using a single GPU is 420.26s, resulting in a speedup of 613.88.

[0035] In summary, this embodiment provides a GPU parallel computing method for flow-induced noise prediction, belonging to the fields of computational fluid dynamics and computational acoustics. The method first obtains unsteady flow field data from CFD or experiments, calculates acoustic source terms such as thickness sources and load sources in the FW-H acoustic analogy on the CPU, and establishes a unified index for time, integral surface, and observation points. Then, based on memory capacity and bandwidth constraints, the index space is block-based, and a fixed or adaptive thread grid is used to map "time step—integral surface unit—observation point" to GPU threads. Merged memory access, shared memory caching, and intra-block reduction are used to calculate the acoustic contribution of a single data block. After delay time alignment and parallel accumulation at the block level, asynchronous data transmission and a multi-stream mechanism are used to overlap between the host and GPU until all data blocks are covered. Finally, sequence alignment and splicing are performed on the CPU to obtain the time-domain sound pressure level or frequency-domain sound pressure level at the observation point. This embodiment fully leverages the parallel capabilities of a single GPU, significantly improving computational efficiency and memory bandwidth utilization during the FW-H integration stage. It boasts advantages such as clear structure, simple implementation, and ease of engineering integration.

[0036] This embodiment also provides a flow-induced noise prediction system based on GPU parallel computing, comprising: The data acquisition module is used to acquire unsteady flow field data; it defines integral surface elements, calculates the sound source terms of each integral surface element at each time step on the CPU, and assigns unified numbers and establishes a unified index to the time steps, integral surface elements and observation points; wherein, the sound source terms include thickness sources and load sources; The block partitioning module is used to divide the integral surface unit or time step into several data blocks according to the GPU's memory capacity, and construct a linear storage layout corresponding to the data blocks. The in-block data processing module is used to input the current data block into the GPU to calculate the acoustic contribution. The calculation of the acoustic contribution includes mapping the time step, integral surface unit, and observation point index to the GPU's thread grid, so that a single thread calculates the acoustic contribution of an integral surface unit to an observation point at a time step, and performs parallel reduction on the local acoustic contribution of the same observation point or the same integral surface unit on the GPU to obtain the local calculation result corresponding to the current data block. Using asynchronous data transmission and multi-stream mechanism, the local calculation result of the current data block is transmitted back to the CPU, and the next data block is input into the GPU for parallel calculation. The data output module is used to perform sequence splicing of the results of each data block on the CPU based on the unified index of observation points and time steps, and to perform alignment correction of the results of each data block according to the delay time alignment rules, and output the noise prediction results of each observation point.

[0037] Alternatively, this embodiment also provides an electronic device, including: CPU and GPU that are interconnected at high speed with said CPU; The memory is used to store flow field data, acoustic source terms, and program code; The CPU and GPU work together to execute the stored program code to implement the aforementioned flow-induced noise prediction method based on GPU parallel computing.

[0038] In practice, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed collaboratively by a CPU and a GPU, implements the aforementioned flow-induced noise prediction method based on GPU parallel computing.

[0039] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting flow-induced noise based on GPU parallel computing, characterized in that, include: S1. Obtain unsteady flow field data; Define an integral surface element, calculate the sound source term of each integral surface element at each time step on the CPU, and assign unified numbering and indexing to the time step, integral surface element, and observation point; wherein, the sound source term includes thickness source and load source; Calculate the propagation distance between the integral surface element and the observation point, calculate the delay time based on the propagation distance and the speed of sound, and discretize and map the delay time to the source time step index to achieve delay time alignment; S2. Based on the GPU's memory capacity, the integral surface unit or time step is divided into several data blocks, and a linear storage layout corresponding to the data blocks is constructed. S3. Input the current data block into the GPU to calculate the acoustic contribution. The calculation of the acoustic contribution includes mapping the time step, integral surface unit, and observation point index to the GPU's thread grid, so that a single thread calculates the acoustic contribution of an integral surface unit to an observation point at a time step, and performs parallel reduction on the local acoustic contribution of the same observation point or the same integral surface unit on the GPU to obtain the local calculation result corresponding to the current data block. Using asynchronous data transmission and multi-stream mechanism, the local calculation result of the current data block is sent back to the CPU, and the next data block is input into the GPU for parallel calculation. S4. Based on the unified index of observation points and time steps, the results of each data block are sequentially spliced ​​on the CPU side, and the results of each data block are aligned and corrected according to the delay time alignment rule, and the noise prediction results of each observation point are output.

2. The method for predicting flow-induced noise based on GPU parallel computing according to claim 1, characterized in that, The process of establishing the unified index specifically includes: The geometric relationship between the integral surface element and the observation point, the medium density and the sound speed are calculated. Based on the geometric relationship, the propagation distance between the integral surface element and the observation point is calculated. The delay time is calculated according to the propagation distance and the sound speed. A mapping table from the delay time to the source time step index is generated. The delay time is discretized and mapped to the source time step index to achieve delay time alignment.

3. The flow-induced noise prediction method based on GPU parallel computing according to claim 1, characterized in that, The linear storage layout adopts a time-continuous, integral surface / observation point block approach.

4. The flow-induced noise prediction method based on GPU parallel computing according to claim 1, characterized in that, The step of inputting the current data block into the GPU to calculate the acoustic contribution specifically includes: The time step, integral surface cell, and observation point index are mapped to the GPU's 3D thread grid, enabling a single thread to calculate the acoustic contribution of an integral surface cell to an observation point at a time step. Intermediate quantities are cached using shared memory on the GPU, and the local acoustic contributions of the same observation point or the same integral surface cell are reduced in a tree-like or segmented parallel manner to obtain the local calculation results corresponding to the current data block.

5. The flow-induced noise prediction method based on GPU parallel computing according to claim 4, characterized in that, If the time step of the current data block is greater than the preset threshold and cannot be loaded into the video memory, the time step is divided into multiple windows. The acoustic contribution calculation task is executed independently in each window. After the calculation of each window is completed, the overlapping area is clipped / compensated and accumulated at the window boundary according to the delay time alignment rule to obtain the local calculation result corresponding to the current data block.

6. A flow-induced noise prediction system based on GPU parallel computing, characterized in that, include: The data acquisition module is used to acquire unsteady flow field data; Define an integral surface element, calculate the sound source term of each integral surface element at each time step on the CPU, and uniformly number and index the time step, integral surface element, and observation point; wherein, the sound source term includes thickness source and load source; calculate the propagation distance between the integral surface element and the observation point, calculate the delay time based on the propagation distance and sound speed, and discretize and map the delay time to the source time step index to achieve delay time alignment; The block partitioning module is used to divide the integral surface unit or time step into several data blocks according to the GPU's memory capacity, and construct a linear storage layout corresponding to the data blocks. The in-block data processing module is used to input the current data block into the GPU to calculate the acoustic contribution. The calculation of the acoustic contribution includes mapping the time step, integral surface unit, and observation point index to the GPU's thread grid, so that a single thread calculates the acoustic contribution of an integral surface unit to an observation point at a time step, and performs parallel reduction on the local acoustic contribution of the same observation point or the same integral surface unit on the GPU to obtain the local calculation result corresponding to the current data block. Using asynchronous data transmission and multi-stream mechanism, the local calculation result of the current data block is transmitted back to the CPU, and the next data block is input into the GPU for parallel calculation. The data output module is used to perform sequence splicing of the results of each data block on the CPU based on the unified index of observation points and time steps, and to perform alignment correction of the results of each data block according to the delay time alignment rules, and output the noise prediction results of each observation point.

7. An electronic device, characterized in that, include: CPU and GPU that are interconnected at high speed with said CPU; The memory is used to store flow field data, acoustic source terms, and program code; The CPU and GPU, when cooperating in executing the stored program code, implement a flow-induced noise prediction method based on GPU parallel computing as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed collaboratively by the CPU and GPU, implements a flow-induced noise prediction method based on GPU parallel computing as described in any one of claims 1-5.

Images