CN110599580A - Multi-seismic mathematical data rapid cross display based on Hilbert space filling curve index - Google Patents

Abstract

The method is characterized in that a concurrent dynamic Hilbert R Tree data index structure is established on a global unified memory of a CPU-GPU (central processing unit-graphics processing unit), and the GPU volume rendering technology is based on a CUDA and OpenGL interoperation technology and a Hilbert coding compact bounding box algorithm. A dynamic concurrent Hilbert R Tree is established on a file, a memory and a cache, so that the working load of a CPU (Central processing Unit) and the occupation of the memory can be effectively reduced, the rendering speed of the mathematical data of multiple earthquakes is improved, the drawing quality is improved, the user interaction time delay is reduced, and the user experience effect is improved.

Description

Multi-seismic mathematical data rapid cross display based on Hilbert space filling curve index

Technical Field

The invention relates to a 3D visualization method of massive seismic mathematical data based on Hilbert space filling curve index.

Background

At present, most oil fields in China enter the middle and later development stages, a large amount of oil gas is extracted, the oil gas exploitation difficulty is increased, and higher requirements are also placed on geophysical prospecting development technology. Through three-dimensional visualization of seismic data, the internal structural characteristics and the overall morphological characteristics of the geologic body can be clearly depicted. The advantages of the three-dimensional data volume are fully exerted, and the seismic wave information (amplitude, frequency, phase and space) is displayed in a three-dimensional visual manner in multiple angles, multiple types and multiple attributes. On the basis, geological interpreters can obtain better understanding and make favorable decisions, so that the efficiency of searching oil and gas reservoirs is greatly improved, and the method has great economic value.

However, as the technology of seismic exploration is continuously improved, the scale and speed of generated data far exceed the memory of a computer, the size of a video memory of a GPU and the limitation of transmission bandwidth. Although the computing power of the CPU and the CPU is significantly increased under the push of moore's law, the CPU is only "impressive" on these mass data due to the size of the manufactured and weak memory, the bus transmission bandwidth between the memory and the memory, and the size of the GPU memory.

It is good that the user only observes the data through the visualization terminal (screen, projector, VR device, AR device) to have a visual meaning, and usually the part of the data is lower than the whole data. And the user can traverse all interested data through the interactive system, which provides possibility for visualization of mass data.

At present, under the condition of not changing computer hardware equipment (increasing a memory bank and replacing a high-level display card), the drawing amount of data can be reduced by compressing mass data; secondly, the massive data are partitioned, and sub blocks are called according to the needs of users in the display process.

In the visualization process, if a data compression method is adopted to realize rapid visualization, lossless compression is generally required to be performed on original data. In the process of data visualization, the phenomenon that the result is not true (the picture is green after repeated compression and decompression by some JPEG picture compression algorithms) due to the processes of interpolation, resampling and the like is prevented.

At present, the massive data is partitioned, the block coding is carried out on octree (quadtree on a plane) in the index establishing process, the data is organized by Z-Order coding, and the coding is simple. However, due to the data organization method of the uniform blocking strategy of the octree, if the termination condition is not properly selected, the number of sub-blocks is increased, the constructed octree is too bulky, and the file operation is rapidly increased. The Z-Order coding has poor local Order retention, and the continuity is poor along with the Order conversion of the curve coding, thereby destroying the spatial continuity of data. If Hilbert is adopted to encode the sub-blocks, the spatial correlation of data can be well guaranteed, the encoding is continuous, and the formed Hilbert-R Tree is convenient to query. And key data needing to be loaded are determined by combining a human-computer interaction model, so that the delay of the data is reduced.

In the current seismic data scientific visualization process, the 2D visualization field mainly uses a B + W seismic profile, the drawing technology is mature, the technologies such as Qt, OpenGL and GPU are adopted by predecessors, and the drawing efficiency and the drawing quality are greatly improved. In the field of 3D visualization, volume rendering techniques are widely used commercially because of their high quality and clear internal details compared to other rendering techniques such as the snowballing method.

The ray casting algorithm is a classic algorithm in a three-dimensional seismic visualization algorithm, and the basic principle is as follows: first, it is determined how rays are cast, each starting from a pixel on the screen and casting along the direction of the line of sight. Next, determining how to traverse the volume data, performing sampling calculation when the projection data intersects with the volume data bounding box, sampling ray segments in the volume data at equal intervals in the ray projection direction, performing interpolation calculation on the optical properties (color value and opacity) of the current sampling point, and then synthesizing the optical properties (from front to back and from back to front) to obtain the color value which is finally displayed by the projection ray.

Therefore, the ray casting algorithm has a large calculation amount and a low drawing speed, so that the real-time interactivity is poor, and the application of the ray casting algorithm in more fields is limited. Currently, to solve this problem, there are two methods: firstly, reasonably block-coding the minimum compact bounding box of the bounding voxels, eliminating intersecting boxes which do not intersect with rays as much as possible, and reducing the task amount of drawing; and secondly, compressing the voxels by adopting an asymmetric compression algorithm (the performance requirement on the compression algorithm is not high and the performance requirement on the decompression algorithm is high), and decompressing the GPU in real time during drawing. In the 3D visualization process, the two solutions are incompatible with the previous mass data visualization idea, and how to skillfully design a block coding scheme and a data compression method becomes a core problem of improving the mass data visualization drawing efficiency, drawing quality and real-time interaction.

In addition, the time of processing seismic data by the general rendering engine based on Qt, OpenGL and the like is long, the rasterization processing efficiency of the graph is low, the requirement of rapid visualization application of mass data is difficult to meet, the real-time interaction function of a user cannot be achieved, and particularly, a volume rendering algorithm with huge calculated amount is adopted. The GPU is adopted for drawing, so that the drawing efficiency is improved, and the drawing quality is improved: the modern graphics API (DirectX D3D12 OpenGL 4.5 Vulkan) also adopts more GPUs and less CPUs, and the drawing quality and the drawing efficiency are greatly improved compared with the prior graphics API (DirectX D3D 11).

Disclosure of Invention

The invention aims to provide a multi-seismic mathematical data rapid cross display method which is based on Hilbert indexes, constructs a dynamic concurrent Hilbert R tree in a CPU-GPU global unified memory, supports a plurality of rendering threads and data compression threads to operate the R tree simultaneously, and supports a plurality of view modes of single data. The method comprises the steps of establishing a Hilbert R tree data structure suitable for volume rendering in a GPU video memory based on Hilbert indexes, establishing a lossless bzip2 compression algorithm based on the GPU, reducing the transmission size of data from a file to the memory to the video memory based on a CUDAUVA global unified memory technology and a CUDA-OpenGL interoperation technology, and avoiding data from being copied between the CPU and the GPU.

In order to achieve the above technical objects, the present invention provides the following technical solutions (as shown in fig. 1). The core technology of the method is to establish a dynamic concurrent Hilbert R Tree on a global unified memory of a CPU and a GPU, so that the space query efficiency is improved, the buffer hit rate is improved, and the user operation delay is reduced. Various technical schemes are comprehensively applied: the method comprises the steps of memory mapping reading and writing of large files, concurrent parallel technology of CUDA and OpenMP (CUDA and MPI parallel technology), global unified memory technology in CUDA (unified device architecture) (UVA technology), interoperation technology of CUDA and OpenGL, and establishment of a high-speed rendering channel from the files to a memory, a video memory and a frame cache.

The Hilbert R Tree in a local cache in the rendering thread is a child node of a global CPU-GPU memory Hilbert R Tree; in the global CPU-GPU memory, the Hilbert R Tree is in a sub-node of the Hilbert R Tree which is subjected to Hilbert coding: the spatial structure relationship of the three levels of caches is guaranteed by the structure of the Hilbert R Tree (shown in FIG. 2). All human interaction operations of the user can be mapped to a Hilbert R Tree operation (as shown in FIG. 6).

The method sequentially comprises the following steps:

(1) file IO interacts with global memory: reading mass data in the file, and establishing a dynamic concurrent Hilbert tree on a global unified memory of the CPU-GPU.

(2) File IO interacts with global memory: reading mass data (compressed data) in a file, compressing (decompressing) sub-blocks of the Hilbert tree by using a GPU lossless compression algorithm (decompression algorithm), and storing the compressed (decompressed) sub-blocks into a file written according to the Hilbert node index sequence;

(3) interaction between the global memory and the local cache: the dynamic concurrent Hilbert tree established in the global unified memory of the CPU-GPU is interacted with the Hilbert tree sub-block data in the volume rendering method in the rendering thread;

(4) human-computer interaction and local cache interaction: the viewpoint transformation of the interactive window dynamically updates the Hilbert tree data structure in the rendering thread.

Drawings

Fig. 1 is a flow chart of the solution.

FIG. 2 is a file, global unified memory, rendering thread cache constituting a Hilbert R Tree.

Fig. 3 is a schematic diagram of the concurrent read-write compression flow in fig. 1.

FIG. 4 is a flow chart of the bzip2 compression algorithm of FIG. 1 at 2.

FIG. 5 is a table of Hilbert curve binary codes.

FIG. 6 user interaction mapping to Hilbert R Tree currently shows node transformation.

FIG. 7 CUDA & OpenGL interoperation APIs.

FIG. 8 is a schematic view of a user interaction model hit cache.

Detailed Description

The invention is further described below with reference to the accompanying drawings.

The method comprises the following steps: reading mass data in the file, performing regular blocking, and performing compression coding on each block. From the original uncompressed mass data, regular blocking is performed according to the following principle: one is that the block size must be an integer multiple of the memory page size. This is because the starting offset size of the read/write file is integral multiple of the size of the paged memory by using the memory mapping method. Reading the file by adopting a memory mapping mode can read the subblocks by adopting an MPI multi-process parallel or an OpenMP multi-thread concurrent mode; secondly, the size of the block needs to consider the compression efficiency of the compression algorithm, the size limit of the memory and the size limit of the video memory. The maximum subblocks are divided as far as possible under the size limit of the memory and the video memory, so that the compression efficiency of the compression algorithm is high, the data transmission quantity is remarkably reduced, meanwhile, the man-machine interaction operation can be well covered, and frequent data transmission is avoided. Finally, for the convenience of algorithm implementation, fixed-size partitioning is adopted.

Compressing each fixed sub-block using a GPU-based bzip2 compression algorithm (as shown in fig. 4) to form a compressed sub-block; and simultaneously, Hilbert coding is carried out on each sub-block, and the sub-blocks are written into a compressed file. The process can be performed silently as a background thread (process) for performing the function of mass data compression.

Meanwhile, a concurrent dynamic Hilbert R Tree is globally constructed on the CPU-GPU according to the content to be displayed in the current window of the user.

Step two: and establishing a dynamic Hilbert R Tree in a global unified memory of the CPU-GPU according to the content displayed by the window. The method for establishing the dynamic Hilbert R Tree in the global unified memory of the CPU-GPU has the following characteristics that the height of the Tree is determined, and each sub-block is uniformly partitioned.

And calculating Hilbert code values of the loaded subblocks required by the window display contents according to the window display contents.

Reading the compressed data file into a memory for decompression according to the Hilbert code value of the subblock to be loaded, and encoding the spatial coordinates of each subblock into the Hilbert code value according to the mapping mode of the graph 5. And establishing a Hilbert R Tree for the uniformly partitioned sub-blocks from bottom to top according to the sequence of Hilbert code values. (if read from the original data file, it is necessary to inverse map the Hilbert code to a linear code value according to the Hilbert curve binary code table of FIG. 5, and then operate).

The CPU-GPU global unified memory (UVA technology in CUDA) is provided by the following CUDA APIs: the functions provided by the functional families of cudaMallocManged, cudaFree and the like avoid back and forth copying of CPU and GPU data.

Step three: and the dynamic concurrent Hilbert tree established in the global unified memory of the CPU-GPU is interacted with the Hilbert tree sub-block data in the volume rendering method in the rendering thread. The local cache in the rendering thread is established by an OpenGL function cluster, leaf nodes of a proper level are selected from Hilbert Tree in a memory to serve as root nodes of the Hilbert R Tree in the local cache of the rendering thread, and the Hilbert R Tree which can be complete is established. The human-machine operations of the user will be mapped to interoperate with the Hilbert R Tree in the rendering thread and the Hilbert R Tree in memory (as shown in FIG. 6).

Take the CUDA and OpenGL interoperation as an example (as shown in FIG. 7): OpenGL elements for rendering such as Pixel Buffer (PBO), Texture buffer Object (Texture Object), Screen frame buffer (FrameBuffer), vertex buffer Object (VB0), OpenGL programmable rendering pipeline buffer Object: firstly, the buffer areas can map CUDA pointers at low cost through a CUDA-OpenGL interoperation API, then the CUDA pointers are handed to a kernel function for execution, and after the execution is finished, the mapping is released, and the CUDA pointers are handed to OpenGL for drawing. There are also corresponding interoperation mechanisms for other graphics APIs (DirectX D3D11, DirectX D3D12, Vukan).

The content displayed by the window is a subblock of a certain level of the Hilbert R Tree, the subblock can be used as a root node in a rendering thread local cache, the subblock is partitioned and encoded according to a fixed partitioning principle, and the Hilbert R Tree is established on the local cache. And performing volume rendering on the subblocks by using a ray projection algorithm, wherein the subblocks which do not pass through the rays do not participate in calculation, and finally, the corresponding picture is rendered on the visual terminal by using the volume rendering algorithm.

Step four: the viewpoint transformation of the interactive window dynamically updates the Hilbert tree data structure in the rendering thread.

The 3D rotation view, which is the rotation view operation of the user, appears to a rendering thread based on volume rendering, only the viewpoint is rotated, and the direction of the sight line needs to be changed to perform calculation again. At this point the Hilbert R Tree of the rendering thread does not have any data update operations occurring. Its user operation may hit the rendered thread's local cache, i.e., operation a of the representation of fig. 8.

The zooming-in operation of the user can be analyzed into two situations, one is that the view window transformation is needed, the view window transformation is captured by an OpenGL camera model, and the zooming-in operation can be realized by reducing the distance of a viewpoint imaging plane. When the magnification operation causes the distortion of the visualization operation image, a certain sub-block of the original Hilbert-R Tree needs to be moved and rendered from the beginning. As in the zoom operation of fig. 6. In either case, no additional data is required. The zoom-in operation represents the a operation in fig. 8.

The zoom-out operation of the user can be resolved into two cases, one is view window transformation, which is similar to the zoom-in operation and will be captured by the OpenGL camera model, and the zoom-out operation, which is represented as operation a in fig. 8, can be implemented by increasing the distance from the viewpoint to the imaging plane. When the visual image is distorted due to the zoom-out operation, the child block displayed on the window needs to be replaced by the parent node child block, and the new rendering operation is performed, namely B operation shown in fig. 8. At this time, the display node of the current window is updated to the parent node of the original node, as in the zoom-out operation in fig. 6.

The part of the user's translation attempt operation is captured by the OpenGL view window operation, and exceeding the OpenGL translation attempt operation capability will cause the current window to become its sibling, as in the translation operation in fig. 6. This causes the root node in the rendering thread to replace its sibling node, and also retrieves the data from memory, as in operation B of fig. 8.

Random operation of a user may break through a cache mechanism of the Hilbert R Tree, which may cause that a compressed file is read from a newly read compressed file, and the decompressed file is sent to a global memory to establish the Hilbert R Tree, as shown in operation C in fig. 8, the Hilbert R Tree is established in a local cache according to window display content, and a donor rendering method is used for rendering a 3D image.

Claims (3)

1. The method comprises the steps of fast cross-displaying multi-seismic mathematical data based on Hilbert space filling curve indexes, specifically, based on a bzip2 compression algorithm of a GPU, and based on OpenMP (MPI) and CUDA memory mapping read-write large file technologies, establishing Hilbert R Tree in a file; based on a dynamic concurrent Hilbert R Tree algorithm, a Hilbert R Tree data index structure is arranged on a global unified memory of a CPU-GPU; establishing a Hilbert R Tree file in a rendering thread cache based on a CUDA and OpenGL interoperation technology, a GPU ray projection algorithm based on a GPU and a GPU volume rendering technology of a Hilbert coding compact bounding box algorithm; based on the OpenGL visualization technology, establishing the mapping of user interactive operation and Hilbert RTree operation in a cache, a memory and a file: the method has the advantages that the workload of the CPU and the occupation of the memory can be effectively reduced, the utilization rate of the GPU is improved, the copy transmission of data in files and memories and the multiple times of the CPU and the GPU is reduced, the rendering speed of the mathematical data of multiple earthquakes is improved, the drawing quality is improved, the user interaction time delay is reduced, and the user experience effect is improved.

2. The method is characterized in that the method comprises the following steps of (1) rapidly and alternately displaying multi-seismic mathematical data based on Hilbert space filling curve index: based on a Hilbert coding technology, establishing a Hilbert R Tree from a file, a memory and a rendering thread cache; based on the OpenGL visualization technology, user interaction operation is mapped into operation of a corresponding Hilbert R Tree, and a high-speed rendering channel from a file to a visualization terminal is established.

3. A visualization method as recited in claim 1, wherein: in the drawing process, the GPU participates in each visual process in the whole process, the utilization rate of the GPU is high, the load of the CPU is reduced, the drawing speed is high, and the drawing efficiency is high.