Wavefield data reverse-time migration calculation method and system
By constructing ultra-large memory for computing nodes and employing hybrid memory technology to achieve high-speed access to wavefield data, the massive storage and computational requirements for large-scale seismic data reverse time migration calculations were solved, improving computational efficiency and reducing costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA PETROLEUM & CHEMICAL CORP
- Filing Date
- 2021-09-13
- Publication Date
- 2026-06-09
AI Technical Summary
The calculation of large-scale earthquake data in reverse time migration requires massive computation and storage, leading to bottlenecks in computer hardware technology. Existing technologies, through methods such as reconstruction calculation and lossless compression, have failed to effectively improve computational efficiency, and hard disk read and write times have a serious impact.
The computing nodes are built with ultra-large memory, and hybrid memory technology is adopted, using phase-change memory as main memory and dynamic random access memory as cache to achieve high-speed access to wave field data and reduce the amount of computation and storage requirements.
It improves the efficiency of reverse time migration calculation for large-scale seismic data by at least 46%, saves computing time and energy consumption, and reduces service costs.
Smart Images
Figure CN115808715B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial seismic reflection wave imaging, specifically relating to a method and system for calculating the reverse time migration of wavefield data, which can be applied to geological exploration fields such as oil exploration and mineral exploration. Background Technology
[0002] Wave equation reverse time migration offers the highest imaging accuracy, the strongest adaptability to complex structures, and is not limited by imaging dip angle. It can be used for imaging complex areas and steep-dip structures, making it a widely adopted imaging method. As oil and gas exploration gradually develops towards "broad, wide, and high" approaches, acquiring broadband, wide-azimuth, and high-density seismic data improves more accurate imaging and oil and gas prediction, leading to a continuous increase in the scale of seismic data.
[0003] As the scale of seismic data continues to increase, the spatial scale of single-shot data imaging in reverse time migration calculations also increases, leading to a continuous increase in the computational workload and wavefield data volume for wavefield extrapolation. Therefore, large-scale seismic data reverse time migration places enormous demands on computation, storage, and memory, exhibiting massive computational and storage requirements. In practical applications, it encounters technical bottlenecks in computer hardware.
[0004] Current technologies for addressing these bottlenecks include: using coprocessors to implement finite-difference wavefield extensions, in-node CPU and coprocessor collaborative parallel and decoupled algorithms, multi-node parallelism to improve computational efficiency, and optimizing checkpoint methods and lossless compression algorithms to reduce storage requirements. Although existing technologies have shortened the processing time for large-scale seismic data reverse time migration to one to two weeks, it remains the algorithm with the longest processing time and requires further efficiency improvements.
[0005] Due to manufacturing process limitations, the density and power consumption of DRAM memory are generally limited to no more than 256GB of memory in computing nodes. Faced with the massive storage requirements of large-scale seismic reverse time migration wavefield data, high-speed access to wavefield data has always been a technical bottleneck in reverse time migration calculation. Existing technologies usually use wavefield data storage strategies that trade computation for storage and time for space, and reduce the amount of local hard disk storage of wavefield data through source wavefield reconstruction calculation methods. However, the time spent on reconstruction calculation and the long local hard disk read and write time seriously affect the improvement of reverse time migration calculation efficiency.
[0006] In 2012, Wang Baoli et al. introduced seven wavefield data storage strategies in the paper (Wang Baoli, Gao Jinghuai, Chen Wenchao, et al. Effective boundary storage strategies for pre-stack reverse time migration of earthquakes [J]. Chinese Journal of Geophysics, 2012, 55(7): 2412-2421). These strategies included the rarely used strategy of storing the source wavefield on a full hard disk with the least computational cost (2Nt), the strategy of not storing the source wavefield reconstructed from time 0 with a significantly increased computational cost (approximately Nt*(Nt+3) / 2), the strategy of storing the wavefield value every N steps for source wavefield reconstruction, and the strategy of using the last few The literature discusses various boundary storage strategies, including backpropagation using the adjoint wave equation for time-series source wavefield data, boundary storage strategies that only store source wavefields within the artificial boundary layer for source wavefield reconstruction, checkpoint storage strategies that better balance computational and storage costs, and random boundary storage strategies (3Nt computational cost, 0 storage cost). Further optimization of these boundary storage strategies has led to the proposal of an effective boundary storage strategy and an effective boundary storage strategy based on checkpoint technology (with computational costs decreasing from 3Nt to no more than 4Nt), significantly reducing the local hard drive storage requirements for wavefield data.
[0007] In 2013, Liu Shouwei proposed a boundary wavefield reconstruction strategy for NPML boundary conditions suitable for GPU computing in the literature (Liu Shouwei, Wang Huazhong, Chen Shengchang, et al. Research on the implementation scheme of three-dimensional reverse time migration GPU / CPU cluster. Chinese Journal of Geophysics, 2013, 56(10): 3487-3496). By adding a simulation calculation of the source wavefield, the local hard disk storage scale of the wavefield data was reduced.
[0008] In 2015, Shi Ying et al. in the paper (Shi Ying, Ke Xuan, Zhang Yingying et al. Analysis of reverse time migration boundary conditions and storage strategies [J]. Progress in Geophysics, 2015, 30(2): 581-585) focused on comparing the computational and storage costs of commonly used random boundary storage strategies and absorbing boundary storage strategies such as full-wave field storage, checkpoint storage, and effective boundary storage. They selected two storage strategies, random boundary and effective boundary storage, and conducted GPU field tests on the same model. They also conducted comparative analysis from three aspects: computation time, storage time, and imaging accuracy.
[0009] In the literature (Wang Baoli, Gao Jinghuai, Chen Wenchao, et al. Effective boundary storage strategy for pre-stack reverse time migration of earthquakes [J]. Chinese Journal of Geophysics, 2012, 55(7): 2412-2421) and (Shi Ying, Ke Xuan, Zhang Yingying, et al. Analysis of boundary conditions and storage strategies for reverse time migration [J]. Progress in Geophysics, 2015, 30(2): 581-585), by comparing the imaging effects of random boundary and absorbing boundary storage strategies, it was found that the random boundary reverse time migration with the least increase in computation and no wave field data storage produced more obvious noise interference at the shallow boundary.
[0010] Chinese patent publication CN105717539A discloses a three-dimensional TTI medium reverse time migration imaging method based on multi-GPU computing, and Chinese patent publication CN104133987A discloses a reverse time migration method for carbonate reservoirs. These patent publications both focus on implementation methods for wavefield data storage strategies in reverse time migration algorithms. They employ reconstruction calculation methods (increasing computational load) to reduce the local hard drive storage of wavefield data or eliminate local hard drive storage of wavefield data. However, they have the following shortcomings:
[0011] Factors such as the time spent on reconstruction calculations and the long local hard drive read / write time significantly impact the efficiency improvement of time-offset calculations.
[0012] Taking the commonly used checkpoint storage strategy as an example, a reconstruction calculation method is used to reduce the size of the wavefield data. Lossless compression is used to further reduce the size of the wavefield data. The compressed data is stored on a slow local hard drive. This requires an additional forward calculation, increasing the computational load by 1 / 3, thus increasing both computational load and time; it also results in a longer local hard drive read / write time. Summary of the Invention
[0013] The purpose of this invention is to solve the problems existing in the prior art and provide a method and system for calculating reverse time migration of wavefield data. By constructing a computing node with ultra-large memory, a novel wavefield data storage method is proposed, and a reverse time migration algorithm for accessing and storing wavefield data with ultra-large memory is designed. This efficiently solves the problem of storing large amounts of wavefield data and improves the efficiency of reverse time migration calculation of large-scale seismic data by reducing the amount of computation and using high-speed memory access.
[0014] This invention is achieved through the following technical solution:
[0015] In a first aspect, the present invention provides a method for calculating the reverse time migration of wavefield data, wherein the method enables high-speed access to wavefield data and reverse time migration calculation by constructing a computing node with a very large memory.
[0016] A further improvement of the present invention is that:
[0017] The method includes:
[0018] (1) Steps for building a large memory for computing nodes;
[0019] (2) Steps for calculating the reverse time migration of wavefield data.
[0020] A further improvement of the present invention is that:
[0021] Step (1) is implemented using hybrid memory.
[0022] A further improvement of the present invention is that:
[0023] The operation of step (1) includes:
[0024] Use phase-change memory with memory slot interfaces as main memory;
[0025] Use dynamic random access memory as a cache for main memory;
[0026] The CPU performs data read and write operations in the cache and completes data migration between the cache and main memory.
[0027] A further improvement of the present invention is that:
[0028] The operation of step (2) includes:
[0029] (21) GPU source wavefield forward propagation calculation, obtain the forward propagation wavefield data required for relevant imaging, and store it in hybrid memory after lossless compression;
[0030] (22) GPU detector wavefield backpropagation calculation to obtain backpropagation wavefield data required for relevant imaging;
[0031] (23) Decompress and read the forward propagation wavefield data and reverse propagation wavefield data of the corresponding time step in the mixed memory, perform related imaging calculations, obtain related imaging results, and superimpose the related imaging results with the previous related imaging results.
[0032] (24) Repeat steps (22) and (23) to finally obtain the offset imaging results.
[0033] A further improvement of the present invention is that:
[0034] The operation of step (21) includes:
[0035] The GPU performs forward propagation calculations to obtain the forward propagation wavefield data required for relevant imaging calculations.
[0036] The CPU performs lossless compression on the forward propagation wave field data and stores it in hybrid memory: The CPU receives the forward propagation wave field data sent by the GPU, performs lossless compression on the forward propagation wave field data in the cache to obtain compressed wave field data, and stores the compressed wave field data in the cache.
[0037] The compressed wavefield data is migrated from the cache to main memory, and the compressed wavefield data for the last time step will be stored in the cache.
[0038] A further improvement of the present invention is that:
[0039] The operation of step (23) includes:
[0040] Relevant imaging calculations are performed from the last time step of the propagation towards the zero time step: First, the propagation compressed wavefield data of the last part of the time step stored in the cache is used, and then the compressed wavefield data of the corresponding time step is migrated from main memory to the cache.
[0041] The CPU decompresses the compressed wave field data in the cache to obtain the decompressed forward wave field data, and then transmits the decompressed forward wave field data to the GPU.
[0042] The GPU uses the forward propagation wavefield data and the reverse propagation wavefield data obtained in step (22) to perform correlation imaging calculations, obtain correlation imaging results, and superimpose the correlation imaging results with the previous correlation imaging results.
[0043] A second aspect of the present invention provides a wavefield data reverse time migration calculation system, the system comprising:
[0044] Compute nodes are built-in units with extremely large memory.
[0045] Offset calculation unit.
[0046] A further improvement of the present invention is that:
[0047] The computing node's ultra-large memory building unit adopts hybrid memory;
[0048] The computing node's ultra-large memory building unit includes:
[0049] Main memory building block: Uses phase-change memory with memory slot interfaces as main memory;
[0050] Cache building block: Uses dynamic random access memory as a cache for main memory.
[0051] A further improvement of the present invention is that:
[0052] The offset calculation unit includes:
[0053] Forward propagation calculation module: Utilizes GPU to perform forward propagation calculation of the source wavefield, obtains the forward propagation wavefield data required for relevant imaging, and stores it in hybrid memory after lossless compression;
[0054] Backpropagation calculation module: Utilizes GPU to perform backpropagation calculation of the wavefield at the detector point to obtain the backpropagation wavefield data required for relevant imaging;
[0055] Correlation imaging calculation module: Connected to the reverse propagation calculation module, it decompresses and reads the forward propagation wavefield data and reverse propagation wavefield data of the corresponding time step in the hybrid memory to complete the correlation imaging calculation, obtain the correlation imaging result, and superimposes the correlation imaging result with the previous correlation imaging result to finally obtain the migration imaging result.
[0056] Compared with existing technologies, the beneficial effects of this invention are: compared with traditional reverse time migration algorithms that employ checkpoint storage strategies, the method of this invention can improve the computational efficiency of large-scale seismic data reverse time migration by at least 46%. When applied to large-scale seismic data reverse time migration processing in the oil exploration field, this invention saves computation time and machine time costs, reduces computational energy consumption, decreases service costs, and improves the computational efficiency of reverse time migration. Attached Figure Description
[0057] Figure 1 Hybrid memory architecture diagram
[0058] Figure 2 Flowchart of traditional time offset calculation using checkpoint storage strategy;
[0059] Figure 3 A flowchart illustrating the steps of the method of this invention. Detailed Implementation
[0060] The present invention will now be described in further detail with reference to the accompanying drawings:
[0061] Reverse time migration (RTM) is a high-precision reflected wave imaging algorithm applicable to resource fields such as oil exploration, gas field development, and mineral resource detection. In recent years, RTM has become a widely adopted imaging method, utilizing high-speed coprocessors such as GPUs, MACs, and FPGAs, along with CPU-coprocessor parallelism and decoupling algorithms to improve computational efficiency and shorten the cycle time. However, due to limitations in computing node memory capacity, the need for large-scale wavefield data storage has not been effectively addressed. Commonly, wavefield storage strategies are employed to reduce local hard drive storage or to avoid local hard drive storage altogether. The time spent on repetitive calculations and the long, slow local hard drive read / write times severely impact the efficiency of RTM calculations.
[0062] The present invention provides a method for memory access of wavefield data with reverse time offset. Specifically, the method involves constructing a large memory for computing nodes to achieve high-speed memory access of wavefield data.
[0063] This invention reduces the amount of computation and significantly improves the efficiency of large-scale seismic data migration calculation by constructing a computing node with ultra-large memory and adopting a high-speed wavefield data memory access method.
[0064] This invention is a reverse-time offset calculation scheme for ultra-large memory wave field data storage. An embodiment of the method of this invention is as follows:
[0065] Example 1
[0066] The method includes:
[0067] Steps for constructing ultra-large memory for computing nodes and steps for calculating the reverse time offset of ultra-large memory wave field data storage.
[0068] (1) Steps for building a large memory for compute nodes
[0069] Due to the limitations of manufacturing processes, the density and power consumption of Dynamic Random Access Memory (DRAM) are constrained by its low capacity, high price, and high power consumption. As a result, the memory capacity of server configurations is generally no more than 256GB, which is difficult to meet the needs of many large memory applications.
[0070] In recent years, new non-volatile random access memory (NVM) media have driven innovation in server memory technology. Phase-change memory (PCM), used as RAM, possesses characteristics such as non-volatility, large capacity, and low power consumption. However, it suffers from relatively lower performance compared to DRAM and a limited lifespan. Hybrid memory technology, which leverages the performance advantages of both DRAM and PCM, can significantly increase server memory capacity and reduce costs, meeting the demands of ultra-large memory applications. Using server hybrid memory technology, such as... Figure 1 As shown, a large-capacity phase-change memory (PCM) with memory slot interfaces is used as main memory, while a high-speed, small-capacity dynamic random access memory (DRAM) is used as a cache for the main memory. The CPU performs data read and write operations in the DRAM cache, and the system manages data migration between the DRAM cache and the PCM main memory according to a hybrid memory data management strategy. For example, DDR4 can be used as the cache, and Optane memory as the main memory. With hybrid memory technology, a single dual-CPU compute node can be configured with up to several terabytes of memory.
[0071] (2) Calculation steps for reverse time offset of ultra-large memory wave field data storage
[0072] The conventional reverse time offset algorithm, which employs a checkpoint storage strategy, reduces the amount of local hard drive storage required for wavefield data by adding one forward pass calculation and lossless compression. Its calculation process is as follows: Figure 2 As shown:
[0073] (1) First, perform GPU forward calculation and then compress and store the source checkpoint wavefield data on the local hard drive without loss;
[0074] (2) Before the reverse pass calculation, read the decompressed checkpoint wave field data from the local hard drive for GPU forward pass calculation (i.e., Figure 2 The "GPU source wavefield reconstruction calculation" (which adds computational load) obtains the wavefield data required for imaging and compresses and stores it in memory.
[0075] (3) GPU reverse transmission calculation: The CPU reads the compressed wavefield data corresponding to the time from memory, decompresses it, and transmits it to the GPU for relevant imaging calculations. Starting from the last checkpoint Nc to the first checkpoint N0, steps (2) and (3) are repeated to finally obtain the offset imaging result.
[0076] Because wavefield data is massive, high-speed, low-capacity memory cannot store it. Storing it entirely on disk not only requires significant disk capacity but also incurs substantial time overhead due to slow disk read / write speeds. Given the powerful computing capabilities of GPUs, the industry has adopted a compromise: trading computation for storage. Only a small amount of checkpoint data is stored on the hard drive, reducing storage requirements and mitigating the impact of slow disk I / O bottlenecks. Leveraging the high performance of GPUs, a second forward pass is performed using checkpoints to obtain the wavefield data needed for imaging calculations, thus resolving the problems associated with full disk storage.
[0077] However, the time-offset algorithm of the checkpoint storage strategy adds one forward pass calculation, and the increased calculation time and longer local hard disk read / write time seriously affect the efficiency improvement of time-offset calculation.
[0078] The innovation of this invention is that it uses the ultra-large memory of computing nodes to achieve high-speed access to all wave field data in memory.
[0079] The calculation process for step (2) is as follows: Figure 3 As shown, Figure 3 The flowchart shown is for all calculation points. This is the calculation process for a single shot offset. If there are 100 calculation nodes, and each calculation node calculates 1 shot, then the 100 nodes can calculate 100 shots simultaneously in parallel, as detailed below:
[0080] (21) The GPU calculates the forward propagation of the source wave field to obtain the forward propagation wave field data required for related imaging. The CPU performs lossless compression on the forward propagation wave field data required for related imaging calculation and stores it in the hybrid memory.
[0081] Specifically, in step (21), the CPU receives the forward wavefield data obtained from the forward calculation sent by the GPU, performs lossless compression in the DRAM cache to obtain compressed wavefield data, and stores the compressed wavefield data in the cache. Then, according to the hybrid memory data management strategy (a mature existing technology, which will not be elaborated here), the compressed wavefield data is migrated from the cache to the large-capacity main memory. The compressed wavefield data of the last part of the time step will be stored in the cache. The specific number of time steps included in the "last part of the time step" is determined by the hybrid memory data management strategy, which can be determined according to the specific management strategy, and will not be elaborated here.
[0082] The forward propagation calculation obtains the forward propagation wavefield data required for correlation imaging calculation. If the forward propagation calculation performs 10,000 time steps in the forward propagation direction, since a correlation imaging calculation will be performed every set time step (e.g., 10 time steps), the wavefield data for these 1,000 time steps needs to be stored for the correlation imaging calculation.
[0083] (22) GPU detector wavefield backpropagation calculation to obtain backpropagation wavefield data required for relevant imaging; the backpropagation calculation is a calculation of 10,000 time steps opposite to the forward propagation direction.
[0084] (23) Perform relevant imaging calculations using forward propagation wavefield data and reverse propagation wavefield data:
[0085] Because the relevant imaging calculations are performed from the last time step of the forward transmission to the zero time step, and the compressed wavefield data of the last part of the forward transmission time step used first is already stored in the cache, the calculations can be performed directly using the data in the cache. After the calculations for the last part of the time step are completed, the compressed wavefield data of the corresponding time step will be migrated from main memory to DRAM cache according to the hybrid memory data management strategy (a mature technology that will not be elaborated here). The CPU decompresses the compressed wavefield data in the cache to obtain the decompressed forward transmission wavefield data, and then transmits the decompressed forward transmission wavefield data to the GPU.
[0086] The GPU uses the back-propagating wavefield data and forward-propagating wavefield data obtained in step (22) to perform correlation imaging calculations. That is, it performs correlation imaging calculations once every set time step (e.g., 10 time steps) to obtain correlation imaging results, and then superimposes the correlation imaging results with the previous correlation imaging results.
[0087] (24) Repeat steps (22) and (23) to finally obtain the offset imaging results.
[0088] The above steps employ existing forward transmission calculation methods, reverse transmission calculation methods, and offset imaging calculation methods. This invention does not improve upon these calculation methods, therefore the specific steps of these calculation methods will not be described in detail.
[0089] Figure 3 The 12GB / s speed mentioned refers to the data transfer speed of the GPU's PCIe 3.0 x16 interface, which is implemented at the system level and will not be elaborated upon here.
[0090] In summary, this invention reduces computational load and improves the efficiency of large-scale seismic data migration calculations by constructing ultra-large memory for computing nodes and employing a high-speed memory access method for wavefield data. Furthermore, it leverages the high-speed memory access characteristics to achieve high-speed reading and writing of wavefield data.
[0091] The present invention also provides a system for calculating the reverse time migration of wavefield data, and an embodiment of the system is as follows:
[0092]
Example 2
[0093] The system includes:
[0094] Compute nodes are built-in units with extremely large memory.
[0095] Offset calculation unit.
[0096] The computing node's ultra-large memory building unit adopts hybrid memory, which includes:
[0097] Main memory building block: Uses phase-change memory with memory slot interfaces as main memory;
[0098] Cache building block: Uses dynamic random access memory as a cache for main memory.
[0099] The offset calculation unit includes:
[0100] Forward propagation calculation module: Utilizes GPU to perform forward propagation calculation of the source wavefield, obtains the forward propagation wavefield data required for relevant imaging, and stores it in hybrid memory without loss;
[0101] Backpropagation calculation module: Utilizes GPU to perform backpropagation calculation of the wavefield at the detector point to obtain the forward propagation wavefield data required for relevant imaging;
[0102] Correlation imaging calculation module: Connected to the reverse propagation calculation module, it decompresses and reads the forward propagation wavefield data and reverse propagation wavefield data of the corresponding time step in the hybrid memory to complete the correlation imaging calculation, obtain the correlation imaging result, and superimposes the correlation imaging result with the previous correlation imaging result to finally obtain the migration imaging result.
[0103] An embodiment of the present invention is as follows:
[0104]
Example 3
[0105] The application test of the method of the present invention using actual production data is as follows:
[0106] A server with two Intel 6226R CPUs, 1.5TB of memory (e.g., twelve 16GB (192GB) DDR4 modules as cache and twelve 128GB (1536GB) Optane memory modules as main memory) and two V100S GPUs was used as the test computing node. A large-scale seismic dataset from actual production processing was selected, consisting of 87,838 shots, 7,027 channels per shot, with an imaging calculation grid of 563×487×1501 and a total extension step count Nt of 8285. Relevant imaging calculations were performed every 10 steps, therefore the source wavefield data required storage for 830 steps. Each step of wavefield data storage was approximately 1.533GB, resulting in a total storage size of approximately 1272.5GB for the single-shot source wavefield data, and approximately 626.8GB after lossless compression. On the test computing node, the traditional checkpoint technique and the wavefield data memory storage wavefield data reverse time offset calculation method of the present invention were run sequentially to perform single shot offset calculation. The calculation time of the two offset calculation methods is shown in Table 1. Table 1 is a comparison table of the test results of high-density data single shot offset calculation of a computing node (i.e., a single node). As can be seen from Table 1, the single shot calculation time of the memory storage offset calculation method of the present invention is shortened by 279 seconds, and the calculation efficiency is improved by 46%.
[0107] Storage method Checkpointing technique Wavefield data in-memory storage Computational length 884 seconds 605 seconds
[0108] Table 1
[0109] This invention provides an efficient method for calculating reverse time migration of wavefield data with in-memory access. For large-scale seismic data reverse time migration, compared with the traditional reverse time migration algorithm that uses a checkpoint storage strategy, it saves at least 46% of the running time and machine time cost. It is an advantageous method to reduce computing energy consumption, reduce service costs, and improve the efficiency of reverse time migration calculation.
[0110] This invention is a new artificial seismic reflection wave imaging technology that can be applied to geological surveying fields such as oil exploration and mineral exploration.
[0111] Finally, it should be noted that the above technical solution is only one embodiment of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the above specific embodiments of the present invention. Therefore, the methods described above are only preferred and have no limiting significance.
Claims
1. A method for calculating the reverse time migration of wavefield data, characterized in that: The method achieves high-speed access to wavefield data and reverse-time migration calculation by constructing ultra-large memory for computing nodes, including: (1) Steps for constructing a large memory for a compute node; Step (1) is implemented using hybrid memory, and the operation includes: Use phase-change memory with memory slot interfaces as main memory; Use dynamic random access memory as a cache for main memory; The CPU performs data read and write operations in the cache and completes data migration between the cache and main memory; (2) Steps for calculating the reverse time migration of wavefield data; The calculation steps are all completed at high speed in memory using the computing node’s massive memory, and the calculations are performed sequentially according to the imaging time steps. The calculation steps include GPU source wavefield forward propagation calculation, which involves decompressing and reading forward and backward propagation wavefield data at the corresponding time step from the hybrid memory to perform correlation imaging calculations, obtaining correlation imaging results, and then superimposing these correlation imaging results with the previous correlation imaging results. The operations include: (21) GPU source wavefield forward propagation calculation, obtaining the forward propagation wavefield data required for relevant imaging, and storing it in hybrid memory after lossless compression, including: The GPU performs forward propagation calculations to obtain the forward propagation wavefield data required for relevant imaging calculations. The CPU receives the forward propagation wave field data sent by the GPU, performs lossless compression on the forward propagation wave field data in the cache to obtain compressed wave field data, and stores the compressed wave field data in the cache. The compressed wavefield data is migrated from the cache to main memory, while the compressed wavefield data of the last time step will be stored in the cache. (22) GPU detector wavefield backpropagation calculation to obtain backpropagation wavefield data required for relevant imaging; (23) Decompress and read the forward propagation wavefield data and reverse propagation wavefield data of the corresponding time step in the hybrid memory to perform related imaging calculations, obtain related imaging results, and superimpose the related imaging results with the previous related imaging results. This includes: the CPU decompresses the compressed wavefield data in the cache to obtain the decompressed forward propagation wavefield data, and transmits the decompressed forward propagation wavefield data to the GPU; the GPU uses the forward propagation wavefield data and the reverse propagation wavefield data obtained in step (22) to perform related imaging calculations, obtain related imaging results, and superimpose the related imaging results with the previous related imaging results; perform related imaging calculations from the last time step of the forward propagation to the zero time step: first use the forward propagation compressed wavefield data of the last part of the time step stored in the cache, and then migrate the compressed wavefield data of the corresponding time step from the main memory to the cache; (24) Repeat steps (22) and (23) to finally obtain the offset imaging results.
2. A wavefield data reverse time migration calculation system, characterized in that: The system includes: A compute node hypermemory building unit, wherein the compute node hypermemory building unit adopts hybrid memory; The computing node's ultra-large memory building unit includes: Main memory building block: Uses phase-change memory with memory slot interfaces as main memory; Cache building block: Uses dynamic random access memory as a cache for main memory; The CPU performs data read and write operations in the cache and completes data migration between the cache and the main memory. Migration calculation unit: performs reverse time migration calculation of wavefield data; The calculations are performed at high speed in memory using the computing node’s massive memory, and the calculations are performed sequentially according to the imaging time step. The calculation steps include GPU source wavefield forward propagation calculation, which involves decompressing and reading forward and backward propagation wavefield data at the corresponding time step from the hybrid memory to perform correlation imaging calculations, obtaining correlation imaging results, and then superimposing these correlation imaging results with the previous correlation imaging results. The operations include: (21) GPU source wavefield forward propagation calculation, obtaining the forward propagation wavefield data required for relevant imaging, and storing it in hybrid memory after lossless compression, including: The GPU performs forward propagation calculations to obtain the forward propagation wavefield data required for relevant imaging calculations. The CPU receives the forward propagation wave field data sent by the GPU, performs lossless compression on the forward propagation wave field data in the cache to obtain compressed wave field data, and stores the compressed wave field data in the cache. The compressed wavefield data is migrated from the cache to main memory, while the compressed wavefield data of the last time step will be stored in the cache. (22) GPU detector wavefield backpropagation calculation to obtain backpropagation wavefield data required for relevant imaging; (23) Decompress and read the forward propagation wavefield data and reverse propagation wavefield data of the corresponding time step in the hybrid memory to perform related imaging calculations, obtain related imaging results, and superimpose the related imaging results with the previous related imaging results. This includes: the CPU decompresses the compressed wavefield data in the cache to obtain the decompressed forward propagation wavefield data, and transmits the decompressed forward propagation wavefield data to the GPU; the GPU uses the forward propagation wavefield data and the reverse propagation wavefield data obtained in step (22) to perform related imaging calculations, obtain related imaging results, and superimpose the related imaging results with the previous related imaging results; perform related imaging calculations from the last time step of the forward propagation to the zero time step: first use the forward propagation compressed wavefield data of the last part of the time step stored in the cache, and then migrate the compressed wavefield data of the corresponding time step from the main memory to the cache; (24) Repeat steps (22) and (23) to finally obtain the offset imaging results.
3. The wavefield data reverse time migration calculation system according to claim 2, characterized in that: The offset calculation unit includes: Forward propagation calculation module: Utilizes GPU to perform forward propagation calculation of the source wavefield, obtains the forward propagation wavefield data required for relevant imaging, and stores it in hybrid memory after lossless compression; Backpropagation calculation module: Utilizes GPU to perform backpropagation calculation of the wavefield at the detector point to obtain the backpropagation wavefield data required for relevant imaging; Correlation imaging calculation module: Connected to the reverse propagation calculation module, it decompresses and reads the forward propagation wavefield data and reverse propagation wavefield data of the corresponding time step in the hybrid memory to complete the correlation imaging calculation, obtain the correlation imaging result, and superimposes the correlation imaging result with the previous correlation imaging result to finally obtain the migration imaging result.