Strong real-time signal processing method based on multi-stream processing

Abstract

The invention discloses a strong real-time signal processing method based on multi-stream processing, and the method comprises the steps: carrying out the remainder of a preset stream number of received coherent processing interval data through a CPU according to a CPI number, obtaining a stream number, distributing the coherent processing interval data to each stream distributed calculation program corresponding to the stream number, each stream distributed computing program comprises a CUDA stream and a CPU processing thread which are executed in sequence; the GPU and the CPU execute each stream distributed calculation program to overlap the calculation time and the data copying time of the CUDA stream and the CPU processing thread to obtain each trace point condensation data with different completion time; the CPU obtains the corresponding report to be sent according to the trace point condensation data, and orderly sends the report to be sent to the target program based on a report waiting synchronization mechanism, and the application can complete front-end downloading data receiving, detection processing and report sending in strong real time, thereby improving the development efficiency, shortening the product development period, and reducing the project cost.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and more specifically to a strong real-time signal processing method based on multi-stream processing. Background Art

[0002] The CPU+GPU architecture is currently being explored and expanded in certain signal processing engineering applications. This is due to the CPU's capabilities in general computing and complex logic control, the GPU's advantages in high-speed parallel processing, the excellent development ecosystem and portability of C / C++ and CUDA, and the fact that both have been iteratively developed in the market for many years, resulting in a large market share and a mature product portfolio. Compared to the traditional DSP+FPGA architecture, which suffers from long development cycles and poor portability, the CPU+GPU architecture is becoming a significant development trend in signal processing engineering applications. Different engineering applications and varying data volumes place varying demands on the computing power of GPU boards. While increasing data volumes can be addressed by adding CPU and GPU boards for distributed computing, this also comes with increased hardware costs and space constraints. Furthermore, in real-world engineering applications, copying data from the CPU to the GPU and vice versa accounts for a significant portion of the processing cycle. Summary of the Invention

[0003] The present application aims to at least solve the technical problem in the prior art that it is difficult to give full play to the computing power of the GPU due to unreasonable architecture under limited computing power. To this end, the first aspect of the present application proposes a strong real-time signal processing method based on multi-stream processing, comprising: the CPU modulo the received coherent processing interval data by the preset number of streams according to the CPI number to obtain a stream number, and distributes each coherent processing interval data to each stream distributed computing program corresponding to the stream number, wherein each stream distributed computing program includes a CUDA stream and a CPU processing thread executed sequentially; the GPU and the CPU execute each stream distributed computing program to overlap the computing time and data copy time of the CUDA stream and the CPU processing thread to obtain each point trace condensation data with different completion time; the CPU obtains the corresponding to-be-sent report based on each point trace condensation data, and sends each to-be-sent report to the target program in an orderly manner based on the report waiting synchronization mechanism.

[0004] Optionally, the GPU and CPU execute each stream distributed computing program to overlap the computing time and data copying time of the CUDA stream and the CPU processing thread to obtain point trace condensation data, including: the CPU processing thread sets a first parameter array and a third parameter array based on a preset stream number; the CPU processing thread parses each coherent processing interval data to obtain corresponding control parameters, and stores each control parameter in the first parameter array; each CUDA stream executes a point trace detection task based on the control parameters in the first parameter array to obtain each point trace detection data with different completion time, and copies the control parameters in the first parameter array to the second parameter array; the CPU processing thread copies the control parameters in the second parameter array to the third parameter array, and after copying each point trace detection data with different completion time to the CPU, performs point trace condensation processing to obtain each point trace condensation data with different completion time.

[0005] Optionally, before the CPU processing thread copies the control parameters in the second parameter array to the third parameter array and copies the point trace detection data with different completion times to the CPU, the method further includes: triggering a CPU-GPU synchronization instruction after the GPU completes each point trace detection task; in response to the CPU-GPU synchronization instruction, the CPU processing thread synchronizes the execution results of each GPU point trace detection task to the CPU.

[0006] Optionally, before the CPU modulo the preset number of streams for each received coherent processing interval data according to the CPI number, the method further includes: the CPU processing thread sets the parameter space and storage space of the CPU and GPU respectively, and the parameter space and storage space are physically isolated; wherein the parameter space includes a first parameter array, a second parameter array and a third parameter array, and the storage space includes the CPU's memory and the GPU's video memory; the CPU processing thread stores the first parameter array and the third parameter array in the CPU's memory, and stores the second parameter array in the GPU's video memory.

[0007] Optionally, the first parameter array is PARA_CPU_0[n], the second parameter array is PARA_GPU[n], and the third parameter array is PARA_CPU_1[n], where n represents the number of preset streams.

[0008] Optionally, the CPU obtains corresponding reports to be sent based on each dot trace condensation data, and sends each report to be sent to the target program in an orderly manner based on the report waiting synchronization mechanism, including: the CPU processing thread packages each dot trace condensation data with different processing completion time to generate a report, obtains a packaged report corresponding to the CPI number, and sends each report to be sent to the target program based on the CPI number.

[0009] Optionally, the sending of each report to be sent to the target program based on the CPI number includes: the CPU processing thread determines the target CPI number corresponding to the report to be sent currently based on a preset report sending queue; determines whether the CPI number of the currently received report is equal to the target CPI number, and if so, sends the currently received report; otherwise, stores the currently received report in the report sending queue until the report corresponding to the target CPI number arrives, and then sends all arrived reports in the order of report sending.

[0010] Optionally, before the CPU modulo the preset number of streams according to the CPI number of each received coherent processing interval data, the method further includes: the CPU processing thread responds to the received digital beamforming data and parses each coherent processing interval data in the digital beamforming data.

[0011] An embodiment of the present application provides a strong real-time signal processing method based on multi-stream processing. Compared with the existing technology, its beneficial effects are as follows: the CPU takes the modulus of the preset number of streams for each received coherent processing interval data according to the CPI number to obtain a stream number, and distributes each coherent processing interval data to each stream distributed computing program corresponding to the stream number, wherein each stream distributed computing program includes a CUDA stream and a CPU processing thread executed sequentially; the GPU and the CPU execute each stream distributed computing program to overlap the computing time and data copy time of the CUDA stream and the CPU processing thread to obtain each point trace condensation data with different completion time; the CPU obtains the corresponding to-be-sent report based on each point trace condensation data, and sends each to-be-sent report to the target program in an orderly manner based on the report waiting synchronization mechanism. The present application can complete the front-end downlink data reception, detection processing and report sending in strong real-time, thereby improving development efficiency, shortening product development cycle and reducing project cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To more clearly illustrate the technical solution of this application, the following briefly introduces the drawings required for use in the embodiments or prior art descriptions. Obviously, the drawings described below are merely some embodiments of this application, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0013] Figure 1 A flow chart showing the method of the present application;

[0014] Figure 2 A diagram showing the logical relationship of processing threads of the present application method;

[0015] Figure 3 This is a diagram showing an example of the report sorting waiting mechanism of the method of the present application. DETAILED DESCRIPTION

[0016] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0017] This specification provides method operation steps such as embodiments or flowcharts, but may include more or fewer operation steps based on routine or non-inventive work. When implemented in an actual system or server product, the methods shown in the embodiments or figures may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0018] Signal processing in certain fields requires numerous beams, large amounts of data, and high real-time performance. Using default stream processing, the time required to process multiple beams is insufficient. Traditional solutions involve adding processing boards, distributing the beams across multiple boards, and then synchronizing them, which increases hardware costs.

[0019] To address this problem, this method adopts a strong real-time signal processing method based on multi-stream processing in the field of signal processing, so that different CPIs are distributed on different CUDA streams for calculation, overlapping the data transmission time and calculation time. By designing a parameter transfer mechanism, stream synchronization and a report waiting synchronization mechanism, it is possible to use a single CPU board and a single GPU to complete strong real-time processing of multi-beam and large data volumes.

[0020] This method distributes the data of multiple CPIs (Coherent Processing Intervals) to different CUDA streams for calculation, and solves the problem of single-stream overall processing time timeout through multi-stream distributed computing, overlapping data transmission and computing time. Through the parameter passing mechanism, it is guaranteed that the CPI processed by each stream corresponds to the control parameters it receives; through the stream synchronization mechanism, it is guaranteed that the CPU starts point trace processing only after the result processed by the GPU is sent to the CPU, thereby realizing orderly collaboration between the CPU and GPU; through the report waiting synchronization mechanism, it is solved that the calculation result of the short CPI with a later timing arrives at the report packaging node before the long CPI with an earlier timing, thereby causing the problem of report sending timing error, thereby realizing the simultaneous calculation of multiple CPI data and the orderly reporting of the calculation results. The method proposed in this application has the advantages of fast development process, strong program portability, and low cost.

[0021] refer to Figure 1 , Figure 1This is a flowchart of a strong real-time signal processing method based on multi-stream processing provided in an embodiment of the present application. This method can be jointly completed by a system composed of a CPU and a GPU. The point cloud and image feature fusion method based on bird's-eye view hybrid coding may include the following execution process:

[0022] S10. The CPU modulo the received coherent processing interval data by the preset number of streams according to the CPI number to obtain a stream number, and distributes the coherent processing interval data to each stream distributed computing program corresponding to the stream number, wherein each stream distributed computing program includes a CUDA stream and a CPU processing thread that are executed sequentially;

[0023] Among them, each stream has its own dedicated data storage space and parameter storage space, so that each stream can calculate independently to avoid mutual interference. When setting the number of streams, it should be considered that the next time data is allocated to the stream, the data of the current stream has been calculated to avoid being overwritten.

[0024] Specifically, the CPU sets up n CUDA streams and calculates the stream number corresponding to the current data by taking the modulo n of the received CPI data based on the CPI number. This distributes the multiple CPI data across different streams for computation, thereby overlapping computation time and data copy time. To enable independent operation of multiple streams and decouple data reception and processing, the CPU sets up n streams and launches n threads. Each thread controls a stream and determines which thread to start based on the assigned stream number.

[0025] For example, in order to overlap the transmission time and calculation time of data with different CPIs, so that the data of the i+1th CPI can be copied while the current i-th CPI data is being calculated, multiple streams are set (taking 4 as an example, in the actual implementation process, the number of streams needs to be determined according to actual needs and hardware equipment), and different CPI data are distributed to the 4 streams for calculation. When each frame of CPI data arrives, the stream number that should be assigned to the current frame can be obtained by taking the modulo 4 of the CPI count. Each stream has its own unique stream number (0 to 3). In order to enable the 4 streams to perform calculations independently without coupling with each other, each stream is controlled by its own independent thread.

[0026] In one embodiment of the present application, before step S10, the strong real-time signal processing method based on multi-stream processing may further include the following execution process:

[0027] The CPU processing thread sets the parameter space and storage space of the CPU and GPU respectively, and the parameter space and storage space are physically isolated;

[0028] The parameter space includes a first parameter array, a second parameter array, and a third parameter array, and the storage space includes a CPU memory and a GPU video memory;

[0029] The CPU processing thread stores the first parameter array and the third parameter array in the CPU memory, and stores the second parameter array in the GPU memory.

[0030] In one embodiment of the present application, the first parameter array is PARA_CPU_0[n], the second parameter array is PARA_GPU[n], and the third parameter array is PARA_CPU_1[n], where n represents the number of preset streams.

[0031] In one embodiment of the present application, before step S10, the strong real-time signal processing method based on multi-stream processing may further include the following execution process:

[0032] The CPU processing thread parses the received digital beamforming data to obtain coherent processing interval data in the digital beamforming data in response to the received digital beamforming data.

[0033] S20, the GPU and the CPU execute the distributed computing program of each stream to overlap the computing time and data copy time of the CUDA stream and the CPU processing thread, and obtain condensed data of each point trace with different completion time;

[0034] In an embodiment of the present application, step S20 may include the following execution process:

[0035] S201, the CPU processing thread sets a first parameter array and a third parameter array based on a preset stream number;

[0036] S202, the CPU processing thread parses each coherent processing interval data, obtains corresponding control parameters, and stores each control parameter in a first parameter array;

[0037] S203, each CUDA stream performs a point trace detection task based on the control parameters in the first parameter array, obtains each point trace detection data with different completion times, and copies the control parameters in the first parameter array to the second parameter array;

[0038] S204 , the CPU processing thread copies the control parameters in the second parameter array to the third parameter array, and copies the point trace detection data with different completion times to the CPU, and then performs point trace condensation processing to obtain point trace condensation data with different completion times.

[0039] It should be noted that in order to make the CPI data on each stream correspond to its control parameters one by one without misalignment between parameters and data, the CPU allocates a parameter array PARA_CPU_0[n], where n is the number of streams set. Taking 4 streams as an example, n=4. When each frame of CPI data arrives, its control parameters can be parsed into the corresponding array space according to its assigned stream number. At the same time, in order to ensure that the control parameters of each stage in the processing cycle can correspond to the data without misalignment, the control parameters in the first parameter array need to be copied into the corresponding second parameter array PARA_GPU[n] in the GPU. PARA_GPU[n] in the GPU processing stage only has the function of storing control parameters (because if the parameters are not passed down with the data, when the CPU performs report processing later, the parameters in the first parameter array space may be refreshed by the new CPI parameters, resulting in mismatch between data and parameters). The GPU processing stage still uses the control parameters of PARA_CPU_0[n] (because the CPU executes The execution of the line and the GPU are asynchronous. After the CPU issues the CUDA kernel function start command, it will immediately execute the next instruction, while the actual execution of the GPU will be slower than the CPU issuing the instruction. Therefore, the second parameter array will not be modified by the next CPI at this time. After the GPU completes the processing, the control parameters of the second parameter array in PARA_GPU[n] are copied to the CPU along with the data. The control parameters of the second parameter array are copied to the third parameter array PARA_CPU_1[n]. After the GPU stage processing is completed, the subsequent CPU processing uses the corresponding parameters in PARA_CPU_1[n]. It should be noted that PARA_CPU_0[n] and PARA_CPU_1[n] are arrays located in the CPU memory, and PARA_GPU[n] is an array located in the GPU video memory.

[0040] Before step S204, the strong real-time signal processing method based on multi-stream processing may further include the following execution process:

[0041] After the GPU completes each point detection task, it triggers the CPU-GPU synchronization instruction;

[0042] In response to the CPU-GPU synchronization instruction, the CPU processing thread synchronizes the execution results of each point trace detection task of the GPU to the CPU.

[0043] Due to the asynchronous execution of the CPU and GPU mentioned above, the subsequent point trace aggregation and packaged report processing must be carried out after the GPU completes the detection processing and sends the detection results and control parameters back to the CPU before the next step of processing can be carried out. Therefore, before the CPU point trace processing is carried out, the CPU and GPU must be synchronized so that the CPU waits for the GPU to complete the processing and send the data and parameters back before proceeding to the next step of processing. Before this, the CPU needs to wait all the time.

[0044] S30. The CPU obtains corresponding to-be-sent reports according to the condensed data of each point trace, and sends each to-be-sent report to the target program in an orderly manner based on the report waiting synchronization mechanism.

[0045] In an embodiment of the present application, step S30 may include the following execution process:

[0046] S301. The CPU processing thread packages the trace condensation data with different processing completion times to generate a report, obtains the packaged report corresponding to the CPI number, and sends each to-be-sent report to the target program based on the CPI number.

[0047] Specifically, sending each to-be-sent report to the target program based on the CPI number may include the following execution process:

[0048] The CPU processing thread determines the target CPI number corresponding to the current report to be sent based on the preset report sending queue;

[0049] Determine whether the CPI number of the currently received report is equal to the target CPI number. If they are equal, send the currently received report;

[0050] Otherwise, the currently received report is stored in the report sending queue until the report corresponding to the target CPI number arrives, and then all arrived reports are sent in the order in which the reports are sent.

[0051] Among them, in order to solve the problem that the order of sending reports may be inconsistent with the CPI count due to differences in flow calculation speed caused by different data sizes, this application sets up a report sending sorting waiting mechanism in the report sending link, and sets up a report sending queue. The queue calculates the CPI number that should be sent each time. If the CPI arrives, it will be sent out. If not, it will wait and queue future reports and sort them. When the report to be sent arrives, it will be sent out and the next frame report will be sent instead.

[0052] For example, in actual processing, there may be a situation where, because each stream executes independently, the i-th CPI, with its larger beam size and larger data volume, arrives first and is assigned to stream n=i%4. While the i-th CPI is executing on the GPU (with its larger data volume and longer processing time), the i+1-th CPI arrives. This CPI has fewer beams and a smaller data volume, and is assigned to stream n=(i+1)%4. Due to its small data volume, it is quickly processed on the GPU. While the i-th CPI's data is still being processed on the GPU, the processing results of the i+1-th CPI may be sent back to the CPU for trace processing and report packaging before the i-th CPI's. This is not allowed in some cases, as reports must be sent in a strict sequential order. To address this issue, this application sets up a report sending queue during the report sending phase. This queue calculates the CPI number that should be sent at each time. If the CPI number arrives, it sends it. If not, it waits and queues the next report and sorts it. When the report to be sent arrives, it sends it and then sends the next frame report. This design can solve the problem of sending reports in the order.

[0053] The following is further described with reference to specific embodiments:

[0054] This application is implemented using the CPU and GPU in a digital signal processing chassis. The main boards in the chassis are an Intel CPU and an Nvidia GPU. The GPU and CPU are Xeon E5-2658 processors, using the x86 architecture, 64-bit architecture, with 14 cores and 28 threads. The GPU is a Tesla P6 model with GDDR5 memory.

[0055] As attached Figure 2 The figure shows the logical relationship diagram of the processing threads, in which the main function performs initialization work, completes the allocation of memory space, loads and copies coefficients, and initializes threads. After the initialization is completed, the data receiving thread is started, and the data receiving thread completes the reception of data and parameters, and completes the parsing of parameters. According to the parsed CPI number, the remainder is calculated according to the number of set streams, and the parameters and data are allocated to the space of each stream. The corresponding thread is started for each stream to perform calculations. For a certain stream, after the GPU calculation is completed, the calculation result is copied back to the CPU. After the result is copied, the CPU trace report processing thread starts the calculation, and after the calculation is completed, the report is sent to the sending queue, and the report sending thread sends it.

[0056] like Figure 3The following diagram illustrates an example of a report sorting and waiting mechanism. Assume that at a certain moment, based on the CPI number calculation, the report that should be sent is Report 1. At this time, the queue is empty, and Report 1 has entered the queue, so Report 1 is sent. Report 2 should be sent next, but Report 3 is not queued. Report 3 is temporarily stored in the queue, and the queue continues to wait for the next frame report. At this time, Report 2 arrives, and there are two frames of reports in the queue, but Report 3 is before Report 2. The report sending queue is sorted based on the CPI number of the report, and Report 2 is sent. After Report 2 is sent, Report 3 should be sent. Report 3 is detected to be in the queue, so it is also sent. This mechanism achieves the orderly sending of reports.

[0057] In summary, the present application obtains a stream number by taking the modulus of the preset stream number based on the CPI number of each received coherent processing interval data by the CPU, and distributes each coherent processing interval data to each stream distributed computing program corresponding to the stream number, wherein each stream distributed computing program includes a CUDA stream and a CPU processing thread executed sequentially; the GPU and the CPU execute each stream distributed computing program to overlap the computing time and data copying time of the CUDA stream and the CPU processing thread, and obtain each point trace condensation data with different completion time; the CPU obtains the corresponding to-be-sent report based on each point trace condensation data, and sends each to-be-sent report to the target program in an orderly manner based on the report waiting synchronization mechanism. The present application can complete the front-end downlink data reception, detection processing and report sending in strong real-time, thereby improving development efficiency, shortening the product development cycle, and reducing project costs. The method proposed by the present invention has the advantages of fast development process, strong program portability, and low cost.

[0058] The above description is only a preferred embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application are included in the scope of protection of the present application.

Claims

1. A strong real-time signal processing method based on multi-stream processing, characterized in that: include: The CPU obtains a stream number by modulo the received coherent processing interval data by the preset stream number according to the CPI number, and distributes the coherent processing interval data to each stream distributed computing program corresponding to the stream number, wherein each stream distributed computing program includes a CUDA stream and a CPU processing thread that are executed sequentially; The GPU and CPU execute the distributed computing program of each stream to overlap the computing time and data copy time of the CUDA stream and the CPU processing thread, and obtain the condensed data of each point trace with different completion time; The CPU obtains the corresponding to-be-sent reports based on the condensed data of each trace, and sends each to-be-sent report to the target program in an orderly manner based on the report waiting synchronization mechanism.

2. The strong real-time signal processing method based on multi-stream processing according to claim 1, characterized in that: The GPU and CPU execute the distributed computing program of each stream to overlap the computing time and data copy time of the CUDA stream and the CPU processing thread, and obtain the condensed data of each point trace with different completion time, including: The CPU processing thread sets the first parameter array and the third parameter array based on the preset stream number; The CPU processing thread parses each coherent processing interval data to obtain corresponding control parameters, and stores each control parameter in a first parameter array; Each CUDA stream performs a trace detection task based on the control parameters in the first parameter array, obtains trace detection data of each point with different completion times, and copies the control parameters in the first parameter array to the second parameter array; The CPU processing thread copies the control parameters in the second parameter array to the third parameter array, and copies the point trace detection data with different completion times to the CPU, and then performs point trace condensation processing to obtain point trace condensation data with different completion times.

3. The strong real-time signal processing method based on multi-stream processing according to claim 2, characterized in that: Before the CPU processing thread copies the control parameters in the second parameter array to the third parameter array and copies each point trace detection data with different completion times to the CPU, the method further includes: After the GPU completes each point detection task, it triggers the CPU-GPU synchronization instruction; In response to the CPU-GPU synchronization instruction, the CPU processing thread synchronizes the execution results of each point trace detection task of the GPU to the CPU.

4. The strong real-time signal processing method based on multi-stream processing according to claim 1, characterized in that: Before the CPU modulo the received coherent processing interval data by the preset number of streams according to the CPI number, the method further includes: The CPU processing thread sets the parameter space and storage space of the CPU and GPU respectively, and the parameter space and storage space are physically isolated; The parameter space includes a first parameter array, a second parameter array, and a third parameter array, and the storage space includes a CPU memory and a GPU video memory; The CPU processing thread stores the first parameter array and the third parameter array in the CPU memory, and stores the second parameter array in the GPU memory.

5. The strong real-time signal processing method based on multi-stream processing according to claim 4, characterized in that: The first parameter array is PARA_CPU_0[n], the second parameter array is PARA_GPU[n], and the third parameter array is PARA_CPU_1[n], where n represents the number of preset streams.

6. The strong real-time signal processing method based on multi-stream processing according to claim 1, characterized in that: The CPU obtains corresponding to-be-sent reports based on the condensed data of each point trace, and sends each to-be-sent report to the target program in an orderly manner based on the report waiting synchronization mechanism, including: The CPU processing thread packages the condensed data of each point with different processing completion time to generate a report, obtains the packaged report corresponding to the CPI number, and sends each to-be-sent report to the target program based on the CPI number.

7. The strong real-time signal processing method based on multi-stream processing according to claim 6, characterized in that: The order of sending reports in the report sending queue is determined by the CPI number; The sending of each to-be-sent report to the target program based on the CPI number includes: The CPU processing thread determines the target CPI number corresponding to the current report to be sent based on the preset report sending queue; Determine whether the CPI number of the currently received report is equal to the target CPI number. If they are equal, send the currently received report; Otherwise, the currently received report is stored in the report sending queue until the report corresponding to the target CPI number arrives, and then all arrived reports are sent in the order in which the reports are sent.

8. The strong real-time signal processing method based on multi-stream processing according to claim 1, characterized in that: Before the CPU modulo the received coherent processing interval data by the preset number of streams according to the CPI number, the method further includes: The CPU processing thread obtains each coherent processing interval data in the digital beamforming data in response to the received digital beamforming data.

Images