CN103268342A - CUDA-based DEM dynamic visualization acceleration system and method - Google Patents

Abstract

The invention provides a DEM (Digital Elevation Model) dynamic visualization accelerating method based on a CUDA (Computer Unified Device Architecture) technology. The DEM dynamic visualization accelerating method comprises a data preprocessing method, a dynamic visualization method and a parallel co-scheduling method, wherein the data preprocessing method achieves that tile data capable of dynamically loading are generated under a parallel environment, the dynamic visualization method comprises visualization acceleration and computation acceleration, namely LOD (Level of Detail) is used for accelerating visualization, the CUDA is used for accelerating topographic changes so as to achieve computation, and the parallel co-scheduling method achieves parallel and co-scheduling of four operations of external storage-internal storage data exchange, internal storage-video memory data exchange, CPU computation and CUDA (GPU) computation. The dynamic visualization accelerating method can meet the requirement for displaying the large-scale dynamic DEM.

Description

CUDA-based DEM dynamic visualization acceleration system and method

technical field

The invention belongs to the field of high-performance geographic information visualization. It specifically relates to a DEM (Digital Elevation Model, digital elevation model) dynamic visualization acceleration system and method based on CUDA (Compute Unified Device Architecture, general parallel computing architecture) technology.

Background technique

Geomorphic process and evolution have always been an important aspect in geoscience research. Before the 1960s, most studies on landform evolution focused on the theoretical model of landform evolution. In the past three decades, the application of new technologies such as simulation experiments, precision surveys, geographic information systems and remote sensing has laid a solid foundation for quantitative research in geomorphology. Geomorphological statistical methods have been perfected day by day. The current research hotspots and difficulties are based on dynamics to establish a geomorphic process-response mathematical model to simulate the evolution process of geomorphology.

Digital elevation model (Digital Elevation Model, DEM) is an important data for studying surface processes/tectonic landforms. Dynamic visualization of DEM is an effective method for landform evolution simulation. The process can be divided into two steps:

First, the terrain display based on DEM data. DEM directly describes the most basic information of the terrain—elevation, which can be easily converted into the necessary polygon vertices for display, and combined with texture maps to improve the visualization effect, it can display realistic landforms.

Second, the calculation of landform changes combined with DEM neighborhood analysis. The neighborhood analysis of DEM (such as slope analysis, slope aspect analysis, edge detection, filter change, etc.) can extract a lot of basic data. In landform simulation, the results of neighborhood analysis of DEM (such as slope and aspect) are often used to connect with thematic data (such as rainfall and soil information) to calculate the elevation change of terrain. Before each frame is rendered, on the basis of the original landform elevation, the newly calculated elevation change can be superimposed to realize dynamic display.

In the field of Geographical Information System (GIS), the DEM data that needs to be displayed is huge, and it is difficult to realize the above two steps. For the first step—the topography display based on DEM data, the vertices of polygons that need to be displayed simultaneously in the intuitive algorithm exceed the capabilities of existing hardware and cannot meet the needs of real-time display. In this regard, there have been many fruitful researches, forming a method system that comprehensively utilizes various technologies such as grid simplification, level-of-detail model (Level-Of-Detail, LOD), and internal and external memory scheduling optimization. With these, the problem of display efficiency can be solved. For the second step, the DEM neighborhood analysis, the calculation of the change in terrain elevation, and the superposition of the original landform and elevation change all have the characteristics of parallelization. Combined with the general purpose GPU (GPGPU) method of GPU represented by CUDA (Compute Unified Device Architecture, general purpose parallel computing architecture), large-scale calculations can be significantly accelerated to solve the problem of computational efficiency.

Although there are mature technologies for accelerating display and computing efficiency alone, it is not easy to combine the two technologies. It is also necessary to solve how to embed CUDA computing in the rendering process, and reasonably arrange the execution of CPU and GPU to reduce waiting. Unified scheduling of storage, memory, and video memory.

Contents of the invention

The present invention proposes a CUDA-based DEM dynamic visualization acceleration system and method. The purpose is to provide a general acceleration method, which can not only use the existing DEM high-efficiency visualization technology to display large-scale DEM data, but also combine CUDA to efficiently complete Calculation of dynamic changes in terrain.

In order to achieve the above object, technical scheme of the present invention is as follows:

A CUDA-based DEM dynamic visualization acceleration system, comprising: a data preprocessing module, a dynamic visualization module and a parallel collaborative scheduling module, characterized in that:

The data preprocessing module is used to cut the data used for display and calculation into tiles (that is, divide a large file into multiple small files with independent functions), and prepare data for dynamic loading during display; It can be completed in a standard environment, which can be significantly accelerated compared with a stand-alone environment. The implementation method is as follows: the cut raw data is divided into DEM data and thematic data; the cut results are LOD, tile DEM, and tile thematic data (the latter two are collectively called tile data). There is a hierarchical structure between LODs. Different levels of LODs correspond to different display resolutions. Each level of LODs is divided into multiple small files, and each small file represents the same number of grids. The tile data does not have a hierarchical structure, and only divides large files into files of the same size. The geographical range corresponding to each small file is the same as the range corresponding to a file in the bottom LOD. In distributed data preprocessing, a computer is used as a server to assign tasks to multiple clients.

The dynamic visualization module includes a visualization acceleration unit and a dynamic calculation acceleration unit.

Among them, the purpose of the visualization acceleration unit is to reduce the number of polygon vertices displayed on the same screen, and use LOD to achieve this goal. At the same time, the scene graph is used to organize the LOD to make the hierarchy clearer and accelerate the rendering process. The implementation method is as follows: the scene graph organizes LOD hierarchically in the form of a quadtree. That is, a low-resolution LOD corresponds to 4 slightly higher-resolution LODs of the same geographical range, and these 4 LODs are used as child nodes of the low-resolution LOD; and so on, these 4 LODs will also have own child nodes. Through this organization method, some unnecessary calculations can be reduced and accelerated in the scene graph rendering process (update-picking-rendering).

The purpose of the dynamic calculation acceleration unit is to quickly calculate the change of DEM between two frames and apply the change to LOD, use CUDA instead of CPU calculation to accomplish this goal, and combine CUDA calculation with scene graph organization to make the display process correct , efficient. The implementation method is as follows: based on the tile DEM and tile thematic data, use CUDA to calculate the elevation change of the DEM; add a calculation module to the node of the scene graph, and use CUDA to update the calculation module during the update traversal of the scene graph. Elevation changes overlayed to LOD. The final rendering will show the dynamic DEM.

The parallel cooperative scheduling module enables four simultaneous operations of external memory (mainly hard disk) memory data exchange, memory-video memory data exchange, CPU calculation, and CUDA (GPU) calculation to be performed in parallel without affecting the display results. processing to improve display efficiency. The implementation method is as follows: the optimization of internal and external memory data exchange is mainly realized through pre-reading, there is no effective optimization method for memory-video memory data exchange, the optimization of CPU computing scheduling is realized by first calculating dynamic LOD, and the optimization of CUDA computing scheduling is achieved by combining CUDA computing The different steps of the distribution are implemented in different steps of rendering.

The present invention also includes a CUDA-based DEM dynamic visualization acceleration method, comprising the steps of:

1) Generate the LOD, tile DEM, and tile thematic data required for display through data preprocessing, as the data basis for subsequent display.

2) The display of dynamic DEM is divided into 4 steps:

2.1) DEM elevation change calculation:

Use CUDA to start computing DEM elevation changes for areas that need to be displayed dynamically. Since CUDA calculations can be completely parallelized with CPU calculations, in order to speed up the rendering process, many calculations in the DEM elevation change are performed in parallel in the subsequent three steps.

2.2) Update traversal of the scene graph:

Traverse the scene graph, select the LOD to be displayed; use CUDA to superimpose the DEM elevation change on the LOD that needs to be displayed dynamically.

The update traversal of the scene graph consists of two steps:

Step 1: Use the hierarchical structure of the scene graph to efficiently select the LODs that need to be displayed. The implementation method is: the scene graph organizes the LODs hierarchically in the form of a quadtree, and a low-resolution LOD corresponds to 4 identical geographical locations. Slightly higher resolution LODs of the range, make these 4 LODs children of the lower resolution LOD; and so on, these 4 LODs will also have their own child nodes; if the lower resolution LOD is not in the display range , then the corresponding high-resolution LOD must not be in the display range, so there is no need to make repeated judgments;

Step 2: Use CUDA to superimpose the DEM elevation change on the LOD that needs to be dynamically displayed. The implementation method is: realize each node of the scene graph as a LOD node that can call CUDA functions, hereinafter referred to as CUDA-LOD node; CUDA-LOD The node adds a computing module to the LOD node, which contains computing functions and references to computing resources; the CUDA-LOD node allows the computing module to access the resources of the node and all its child nodes, and will execute when the node is updated and traversed The calculation function in the calculation module, the calculation function calls the CUDA kernel function to complete the parallel calculation; due to the different resolutions of CUDA-LOD at different levels, the grid points and the elevation change data of the tile DEM do not correspond one-to-one, but exist A mapping relationship, which needs to be implemented in the calculation function.

2.3) Selection traversal of the scene graph:

Picking and traversal is a necessary step for scene graph rendering, but for the present invention, the scene graph must be selected first, and the superposition of DEM elevation change to LOD can be carried out; therefore, the picking of LOD has been completed in step 2.2). This step performs almost no calculations.

2.4) Rendering traversal of the scene graph:

Use the GPU to complete the final rendering of the LOD that needs to be rendered.

Beneficial effects of the present invention: the technical scheme described in the present invention can use the parallel computing power of GPU for the calculation of dynamic evolution when large-scale DEM is displayed, and realize efficient external memory-internal memory-video memory data exchange, CPU, GPU Parallel computing to meet the needs of large-scale dynamic DEM display.

Description of drawings

Fig. 1 method flowchart of the present invention

Figure 2 embodiment drawings

Detailed ways

Data preprocessing. In GIS, the storage space of large-scale DEM data often exceeds the capacity of memory, so it is necessary to preprocess DEM and related thematic data (mainly cut into small pieces of files) to realize dynamic loading. The data preprocessing is a kind of parallel data preprocessing, which is completed in a distributed environment. When processing in parallel, the granularity of tasks needs to be set reasonably. When a single file is small, the generation of multiple files should be bound into one task.

Preprocessing requires the completion of two parts:

1. Generate LOD for display through DEM to realize dynamic loading during display. LOD is a geometric grid of different resolutions generated by the original DEM. When displaying, the high/low resolution LOD is called according to the distance/nearness of the viewpoint, which improves the display efficiency without losing details. LOD is hierarchically organized in a quadtree structure. The first layer of LOD represents the entire DEM area, and the subsequent layers of LOD are refined in turn, and each LOD node is saved as a file. For example, the first layer of LOD has 1 node and presents the entire area with a resolution of 64*64; the second layer of LOD has 4 nodes and presents the entire area with a resolution of 128*128, and the number of grids for each node remains the same. It is 64*64, corresponding to 1/4 of the entire area; the resolution of the third layer is 256*256, the number of nodes is 16, and the grid number of each node is still 64*64...until the bottom LOD reaches The resolution of the original DEM data.

2. Cut DEM and related thematic data to realize dynamic loading during calculation. Both DEM data and related thematic data need to be cut into blocks corresponding to the lowest level LOD. For example, for a 1024*1024 DEM data, the resolution of each node in the generated LOD is 64*64, and the bottom layer has 16*16=256 blocks, then the DEM and thematic data should also be cut into corresponding 256 blocks, each block The represented geospatial corresponds to the LOD, known as tile DEM and tile thematic data. According to the size of the DEM neighborhood operator, there is a certain overlapping coverage area between the tile DEMs.

Visual acceleration: use scene graph and LOD. LOD has an obvious hierarchical relationship when it is generated, but this cannot be reflected when it is organized as a file on the hard disk. To use this hierarchical relationship in the display, you need to use a scene graph. The scene graph hierarchically organizes multi-layer LODs in the form of quadtrees. This structure will greatly help dynamic loading and selection in the future display process. Since the LOD is a roughening of the original DEM, each LOD has a certain geometric error. When displaying, set a threshold for screen error. In the scene graph, judge from the first layer of LOD. Each LOD needs to calculate the screen error according to the position of the viewpoint and its own geometric error. If the screen error is less than the threshold, the LOD will be displayed, otherwise, go to the next layer of LOD and repeat. the above process. On the one hand, the scene graph completes the organization of display resources, and on the other hand, it also corresponds to an efficient rendering process. The rendering process of the scene graph performs three traversal operations in sequence:

1. Update: In the update traversal, the scene graph library ensures that all display resources are ready to complete rendering, including changing the scene graph structure (adding, deleting, changing nodes), updating node data (using CUDA to complete LOD and DEM elevation changes overlay). The update traversal allows the program to modify the scene graph and is the key to implementing dynamic scenes.

2. Culling: In a culling traversal, the scene graph library examines the bounding volumes of all nodes in the scene. If a node is within the viewport, the scenegraph library will add a reference to that node in the final render list.

3. Drawing: In the drawing traversal, the scene graph will traverse the rendering list generated by the picking traversal process, call the underlying API, and use the GPU to render the geometry.

Dynamically changing computing acceleration: use CUDA computing and embed CUDA computing into the scene graph rendering process. Geomorphic processes can be calculated using geomorphic evolution process equations. A classic geomorphological evolution process equation is as follows:

h=R+C

$\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; R</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}$

$\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; C</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; \&dtri;</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; \&dtri;</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}$

h is the elevation, R is the elevation of the bedrock, C is the thickness of the regolith, U is the structural uplift rate, w is the rate of bedrock surface reduction due to weathering, E is the rate of bedrock valley river process, Q _s is the valley mud The sediment transport rate, q _m is the sediment transport rate per unit width slope. Some parameters need to use thematic data, such as w involves lithology, q _m involves precipitation, and some need to use topographic and geomorphic elements, such as q _m involves slope process, and slope and aspect need to be considered. The calculation of topographic and geomorphic elements is DEM neighborhood analysis, and CUDA can be used to obtain a good acceleration effect; the parameters related to thematic data and the superposition of multiple parameters to calculate the elevation change also have the characteristics of parallel acceleration; therefore, using CUDA to complete these calculations can Get a good acceleration effect.

The dynamic change of DEM is the expression method of the landform process. In the program, the area to be dynamically displayed is determined according to the user's choice, and the remaining area is still displayed statically.

To display dynamic scenes, before the rendering process of the scene graph, it is necessary to calculate the elevation change of the DEM in the selected area. Since the DEM is cut into DEM tiles and loaded dynamically, it is inconvenient to directly record the selected area in subsequent calculations. According to the selected area, the program records the loaded tile DEM, thematic data and the corresponding dynamic range of each tile. Then, according to the tile DEM, tile thematic data, and dynamic area range, use CUDA to calculate the elevation change of all points in each tile DEM (that is, the elevation change of this frame relative to the previous frame), and the elevation that is not in the dynamic range change to 0. Afterwards, the elevation change is superimposed on the DEM tiles, and the changed DEM elevation is calculated to prepare data for the calculation of the next frame. It is also necessary to record the total elevation change from the original DEM to the latest DEM, this data will be used when the LOD is reloaded (for example, when the resolution of the LOD that needs to be displayed in the same area changes, the corresponding LOD needs to be reloaded). In short, each frame needs to calculate the new DEM elevation, DEM elevation change, and DEM total elevation change. The DEM elevation change and DEM total elevation are also stored in the form of tile data corresponding to the DEM.

After the calculation of the DEM elevation change is completed, the DEM elevation change data can be superimposed on the display resource (that is, LOD), and the dynamic display of the DEM can be realized. To effectively complete this step in the scene graph structure, each node of the scene graph needs to be implemented as a LOD node that can call CUDA functions (hereinafter referred to as CUDA-LOD node). The CUDA-LOD node adds a computing module to the LOD node, which contains computing functions and references to computing resources. The CUDA-LOD node allows the computing module to access the resources of the node and all its child nodes. When the node is updated and traversed, the computing function in the computing module will be executed, and the computing function calls the CUDA kernel function to complete parallel computing. As mentioned earlier, each bottom-level LOD corresponds to a tile DEM. The elevation change data corresponding to the tile DEM is used as the resource of the bottom node of CUDA-LOD (initially empty). The calculation module of the CUDA-LOD node accesses the elevation change resources of its bottom-level child nodes. If they are all empty, no calculation is performed. If they are not empty, the calculation function is called to change the polygon vertices of the CUDA-LOD node according to the elevation change. . Due to the different resolutions of CUDA-LOD at different levels, there is not a one-to-one correspondence with the elevation change data of the tile DEM, but there is a mapping relationship (maybe one-to-one or one-to-many), which needs to be implemented in the calculation function A mapping relationship. When the user roams the scene and the LOD resolution in the dynamic display area changes, the elevation of the newly loaded LOD is the original elevation of the data, and it is necessary to superimpose the total change data of the DEM elevation for the new LOD, not the DEM elevation change data between two frames .

Parallel and collaborative scheduling. The rendering of each frame is divided into four steps: DEM elevation change calculation and scene graph update, selection, and drawing traversal. When the program is running, it is necessary to consider the parallel and cooperative scheduling of external memory-memory data exchange, memory-video memory data exchange, CPU calculation, and CUDA (GPU) calculation in the above four steps. The so-called parallel scheduling means that at the right time, two or more of the above four can run at the same time to improve efficiency; Make sure that the data is completely copied.

Internal and external memory data exchange, including: 1. Data reading 2. Data pre-reading 3. Data release. There are two types of data exchanged between internal and external memory: LOD and tile data (tile DEM and tile thematic data). The tile data is used for the calculation of elevation change, which is related to the dynamic display area selected by the user. The reading and release of the data is simple. It will be read after the user selects the area, and released when the user gives up the dynamic display of a certain area. There is no problem of data pre-reading. The reading of LOD is implemented during the update traversal. When it is detected that the LOD to be displayed in the current viewport does not exist in the memory, the data is read from the external memory. If the amount of data to be read at the same time is too large, the display efficiency cannot be guaranteed, which requires data pre-reading. Due to the continuity of viewpoint changes, it is possible to read in advance the LODs that are one layer higher and the first layer resolution of the currently displayed LOD, as well as the LODs that are not in the current viewport but are adjacent to or closer to the viewport; specific pre-reading The number of LODs depends on the amount of memory available to the program. The release of LOD, that is, the release of those LODs that are least likely to be reused, needs to consider the viewport position and the time when the LOD leaves the viewport; the dynamically displayed LOD needs to be released immediately when it is no longer displayed, because when the LOD is displayed again, The display results need to be calculated based on the original LOD and total elevation change data. The previous dynamic display results are intermediate results that cannot be reused. The read-ahead and release of LOD can be implemented at any step of rendering.

Memory-Video memory data exchange, including: 1. Memory-Video memory copy 2. Video memory-Memory copy. All data used in CUDA calculations are stored in the video memory, including LOD polygon vertex data and tile data (tile DEM, tile thematic data, tile elevation change, and tile elevation total change). The memory-video memory copy only occurs when the data is used for the first time. The copy of the tile DEM and the tile theme data is carried out after the user selects the dynamic area. The copy of the LOD polygon vertex data is implemented in the update traversal. When detected The dynamic LOD that needs to be displayed in the current viewport will run when it does not exist in video memory. The video memory-memory copy is performed every frame, only involving the LOD polygon vertex data. During the update traversal, when it is detected that an LOD to be displayed is dynamic, the LOD and the tile elevation change (or the total elevation change) are performed. Overlay calculation, execute copy after the calculation is completed. When using CUDA to calculate the elevation change, the pre-copy of video memory-memory can also be performed, that is, assuming that the LOD displayed in this frame is the same as that of the previous frame, subsequent calculation and copying are performed in advance, but generally no pre-copying is required.

CUDA calculation, including: 1.DEM elevation change calculation 2.LOD superimposed elevation change. DEM elevation change calculation is the first step of rendering, and LOD overlay elevation change is completed in the second step update traversal. The following three steps are carried out in order for DEM elevation change: 1). Elevation change calculation 2). Elevation total change calculation 3). New DEM elevation calculation. CUDA calculations can be executed completely in parallel with CPU calculations, that is, update traversal can be started without waiting for any calculation to complete. However, in order to ensure the correctness of the calculation, before performing the LOD superposition elevation change in the update traversal, it is necessary to complete the "1).DEM elevation change calculation" of the tile corresponding to the LOD, and sometimes to complete the "2).Elevation total change calculation". Therefore, when the update traversal is completed, all the elevation change calculations must have been completed, and some or all of the elevation total change calculations may have been completed. The new DEM elevation calculation is performed in conjunction with the picking and drawing passes after the update pass, but before the next frame's calculation starts.

CPU computing, involving updating, picking, and drawing traversal. In the drawing traversal, the dynamic LOD is drawn first, and then the static LOD is drawn, so that the four steps of rendering the next frame can be started before the drawing traversal is completed.

Example:

The implementation of the present invention will be described in detail below by taking a DEM data with a grid number of 256*256 (the number of grid points is 257*257) as an example of dynamic visualization of surface runoff erosion.

1. Data preprocessing:

Raw data include DEM data and rainfall data. In preprocessing, select 64*64 as the size of each LOD, then the 256*256 DEM should correspond to 3 layers of LOD, the first layer is marked as 0, the second layer is marked as 00, 01, 02, 03, and the second layer is marked as 00, 01, 02, 03. The three layers are labeled 000, 001, 002, 003, 010, 011, 012, 013, 020, 021, 022, 023, 030, 031, 032, 033; the DEM and rainfall data are also cut into the corresponding 16 tiles , marked as 000, 001, 002, 003, 010, 011, 012, 013, 020, 021, 022, 023, 030, 031, 032, 033.

2. Display of static scenes:

The user selects and displays the DEM in the program. Then the program generates a scene graph structure composed of CUDA-LOD nodes according to the configuration file, but each node does not actually load data. According to the default viewpoint (generally set to be far away), calculate the screen error of the first layer of CUDA-LOD, meet the requirements, read LOD-0, and display. The user roams in the program and reaches a certain viewpoint. At this time, the screen error corresponding to CUDA-LOD-0 cannot meet the display requirements, and enters the next layer of CUDA-LOD for depth-first traversal. CUDA-LOD-02, 03 meet the screen error threshold, load CUDA-LOD-02, 03, CUDA-LOD-00, 01 does not meet the screen error threshold, and then enter the next layer of CUDA-LOD, since it has reached the bottom layer, no need To judge the screen error, directly load 000, 001, 002, 003, 010, 011, 012, 013, and display after all reading is completed.

3. Dynamic and static mixed display:

The formula for surface runoff erosion can be simplified as

$\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; h</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}{\mathrm{<span\; class="notranslate">\; Q</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}$

Among them, Q represents runoff per unit width, S represents slope, and β, m, n, k are model parameters.

The user selects an area in the scene that needs to dynamically display the soil erosion process, as shown in Figure 2. According to the selected screen coordinates, the geometric coordinates are calculated; this area partially overlaps 000, 001, 002, 003 and 020, 021 in the tile DEM. Read the corresponding tile DEM and tile rainfall data, and use CUDA to calculate the elevation change of each point in each tile DEM according to the tile DEM, tile thematic data, and dynamic area range. According to the simplified surface runoff erosion formula, it is first necessary to use DEM neighborhood analysis to calculate the slope S, then use S and rainfall data to calculate the runoff per unit width Q, and finally substitute Q, S, β, m, n, k into the formula , to calculate the elevation change. After the above steps are completed, start the rendering of the scene graph. It is necessary to display the nodes 000, 001, 002, 003, 010, 011, 012, 013, 02, 03, and the 010, 011, 012, 013, 03 nodes and their The bottom child node does not correspond to elevation change resources, so the calculation function is not called. 000, 001, 002, and 003 correspond to elevation change resources 000, 001, 002, and 003 respectively, and node 02 corresponds to elevation change resources 020 and 021, and the calculation function needs to be called. The polygon grids of 000, 001, 002, and 003 are mapped to the elevation change resources one by one, and the mapping relationship between the 02 node and the elevation change resource 020 is (x, y)-(2x, 2y) (x∈(0, 32) , y ∈ (0, 32)), the mapping relationship between node 02 and elevation change resource 021 is (x, y)-(2x-64, 2y) (x ∈ (32, 64), y ∈ (0, 32) ). According to the mapping relationship, update the polygon vertices of all dynamic nodes, and finally draw the graph.

The user continues to roam in the scene, and due to the change of viewpoint, the display nodes become 00, 01, 02, 03. Among them, 00 and 01 need to be reloaded from the external memory. The nodes that need to be dynamically displayed (that is, the nodes corresponding to the elevation change resources) are 00 and 02. Among them, 02 has been dynamically displayed in the previous frame, and the new elevation can be calculated by superimposing the "DEM elevation change" data on the LOD; 00 is the newly loaded LOD, and the "DEM total elevation change" data needs to be superimposed to calculate the new elevation.

Claims (9)

1. DEM dynamic and visual accelerating system based on CUDA, comprising: data preprocessing module, dynamic and visual module and concurrent collaborative scheduler module is characterized in that:

Described data preprocessing module cuts into tile for the data that will show and calculate usefulness, and the dynamic load during for demonstration is prepared data; Pre-service is finished under distributed environment, compares stand-alone environment and can significantly obtain acceleration;

Described dynamic and visual module comprises visual accelerator module and dynamic calculation accelerator module;

Wherein, the purpose of visual accelerator module is to reduce the polygon vertex quantity that shows with screen, uses LOD to finish this target, uses scene graph to organize LOD simultaneously, makes hierarchical structure more clear, and accelerates to play up flow process;

The purpose of dynamic calculation accelerator module be calculate the variation of DEM between two frames fast and with change application in LOD, use CUDA replaced C PU to calculate to finish this target, simultaneously CUDA calculating is combined with the scene graph tissue, make procedure for displaying correct, efficient;

Described concurrent collaborative scheduler module makes that external memory-internal storage data exchange, internal memory-video memory exchanges data, CPU calculate, CUDA calculates the four kinds of operations that can carry out simultaneously, not influencing parallel processing under the prerequisite that shows the result, improves display efficiency.

2. the DEM dynamic and visual accelerating system based on CUDA as claimed in claim 1 is characterized in that, described data preprocessing module, and its implementation is as follows: the raw data of cutting is divided into dem data and thematic data; The result of cutting is LOD, tile DEM, tile thematic data; Have hierarchical structure between the LOD, the corresponding different display resolution of different levels LOD, the LOD of each level is divided into a plurality of small documents, and the represented graticule mesh number of each small documents is identical; The tile data are not had a hierarchical structure, only big file division are become the identical file of size, the geographic range of each small documents correspondence, and the scope corresponding with file among the bottom LOD is identical; In the distributed data pre-service, as server, be a plurality of client allocating tasks with a computing machine.

3. the DEM dynamic and visual accelerating system based on CUDA as claimed in claim 1, it is characterized in that, described visual accelerator module, its implementation is as follows: scene graph is organized LOD with the form stratification of quaternary tree, namely, the LOD of a low resolution is to there being the high-resolution slightly LOD of 4 same geographic ranges, with the child node of these 4 LOD as low resolution LOD; And by that analogy, these 4 LOD also will have the child node of oneself; By this method for organizing, make playing up of scene graph can reduce some unnecessary calculating in the flow process, obtain to accelerate.

4. the DEM dynamic and visual accelerating system based on CUDA as claimed in claim 1 is characterized in that, described dynamic calculation accelerator module, and its implementation is as follows: based on tile DEM and tile thematic data, the elevation that uses CUDA to calculate DEM changes; Be computing module of node increase of scene graph, in the renewal traversal of scene graph, this computing module uses CUDA that elevation is changed the LOD that is added to; Last rendering result will demonstrate dynamic DEM.

5. the DEM dynamic and visual accelerating system based on CUDA as claimed in claim 1, it is characterized in that, described concurrent collaborative scheduler module, its implementation is as follows: the optimization of inside and outside deposit data exchange is mainly by reading realization in advance, internal memory-video memory exchanges data does not have the obvious results optimization method, the optimization that CPU calculates scheduling realizes that by the dynamic LOD of calculating of elder generation the optimization that CUDA calculates scheduling is distributed in the different step realization of playing up by the different step with CUDA calculating.

6. the DEM dynamic and visual accelerated method based on CUDA comprises the steps:

1) generate to show LOD, tile DEM, the tile thematic data that needs usefulness by the data pre-service, as after the data presented basis;

2) dynamically the demonstration of DEM is divided into four steps:

2.1) DEM elevation change calculations: using CUDA to begin to calculate needs the dynamically DEM elevation variation in the zone of demonstration;

2.2) renewal of scene graph traversal: the traversal scene graph, choosing needs the LOD that shows; Use CUDA that the DEM elevation is changed on the LOD that needs dynamically to show that is added to;

2.3) selection of scene graph traversal: the selection traversal is the steps necessary that scene graph is played up, and for purposes of the invention, choose earlier and could determine scene graph, just can carry out the stack that the DEM elevation changes to LOD; So the selection of LOD is 2.2) in finish;

2.4) scene graph play up traversal: the LOD that needs are played up uses GPU to finish final rendering.

7. the DEM dynamic and visual accelerated method based on CUDA as claimed in claim 6 is characterized in that described data pre-service comprises:

The raw data of cutting is divided into dem data and thematic data; The result of cutting is LOD, tile DEM, tile thematic data; Have hierarchical structure between the LOD, the corresponding different display resolution of different levels LOD, the LOD of each level is divided into a plurality of small documents, and the represented graticule mesh number of each small documents is identical; The tile data are not had a hierarchical structure, only big file division are become the identical file of size, the geographic range of each small documents correspondence, and the scope corresponding with file among the bottom LOD is identical; In the distributed data pre-service, as server, be a plurality of client allocating tasks with a computing machine.

8. the DEM dynamic and visual accelerated method based on CUDA as claimed in claim 6 is characterized in that, described DEM elevation change calculations comprises:

The DEM elevation changes and carries out in order following three steps: 1) elevation change calculations; 2) the total change calculations of elevation; 3) new DEM elevation calculates; This three step needn't finish before upgrading traversal, can travel through executed in parallel with renewal, but be the correctness that guarantees to calculate, before in upgrading traversal, carrying out the variation of LOD stack elevation, finish " 1) DEM elevation change calculations " with the corresponding tile of LOD, will finish sometimes " 2) the total change calculations of elevation ".

9. the DEM dynamic and visual accelerated method based on CUDA as claimed in claim 6 is characterized in that, the renewal traversal of scene graph comprised for two steps:

The first step: the hierarchical structure of utilizing scene graph, select the LOD that needs demonstration efficiently, implementation method is: scene graph is organized LOD with the form stratification of quaternary tree, the LOD of a low resolution, to the high-resolution slightly LOD of 4 same geographic ranges should be arranged, with the child node of these 4 LOD as low resolution LOD; And by that analogy, these 4 LOD also will have the child node of oneself; If the LOD of low resolution is not in indication range, then corresponding high-resolution LOD needn't repeat to judge also scarcely in indication range;

Second step: use CUDA that the DEM elevation is changed on the LOD that needs dynamically to show that is added to, implementation method is: each node of scene graph is embodied as the LOD node that can call the CUDA function, hereinafter to be referred as the CUDA-LOD node; The CUDA-LOD node comprises computing function and quoting computational resource for the LOD node has increased a computing module in the computing module; The CUDA-LOD node allows computing module that the resource of this node and all child nodes thereof is conducted interviews, and will carry out the computing function in the computing module when node updates travels through, and computing function calls the CUDA kernel function and finishes parallel computation; Because the CUDA-LOD resolution difference of different levels, net point wherein is not corresponding one by one with the elevation delta data of tile DEM, but has mapping relations, need realize these mapping relations in computing function.

Images