Exploiting the GPU for High Performance Geospatial Situational Awareness Involving Massive and Dynamic Data Sets

Transcription

1 Exploiting the GPU for High Performance Geospatial Situational Awareness Involving Massive and Dynamic Data Sets Intro text for this chapter Bart Adams Software Engineering Manager

2 Software Components for Geospatial Situational Awareness Target markets Security, Defense, Emergency, Aviation, Maritime 2 Mission critical deployments Focus on precision, performance, and interoperability

3 Foundations Our Vision GPU Acceleration Main Challenges No Pre-Processing Big Data Customizable and Extensible API Exploiting the GPU Orthorectified Video Feeds Dynamic Plot Filtering and Styling Line-of-Sight Computations Intro text for this chapter

4 Foundations Our Vision GPU Acceleration Main Challenges No Pre-Processing Big Data Customizable and Extensible API Exploiting the GPU Orthorectified Video Feeds Dynamic Plot Filtering and Styling Line-of-Sight Computations Intro text for this chapter

5 1999: Luciad s Founding Vision Organizations like NATO and EUROCONTROL were struggling with data Existing tools could not deal with different sources Often long pre-processing needed Our vision Use the data as is - in real-time No pre-processing Still stands after 15 years of innovation 5

6 Key Product: LuciadLightspeed Performance obtained through intelligent use of Graphics Processing Unit (GPU) Visualization Analysis 6

7 Key Product: LuciadLightspeed Target also low-end hardware OpenGL 2.0 Some advanced features require ARB_framebuffer_object GLSL 1.1 OpenCL 1.0 With software fallbacks Important for virtualized and/or certified hardware, legacy systems 7

8 Foundations Our Vision GPU Acceleration Main Challenges No Pre-Processing Big Data Customizable and Extensible API Exploiting the GPU Orthorectified Video Feeds Dynamic Plot Filtering and Styling Line-of-Sight Computations 8

9 Challenge 1: No Pre-Processing Support over 40 formats and 100 standards Shape, GML, DTED, S-57, AIXM5, NITF, GRIB, GeoTIFF, Most not optimized for GPU rendering Input often very high-level description <gml:polygon srsname="epsg:4326"> <gml:linearring> <gml:coordinates> lon0, lat0, lon1, lat1,... </gml:coordinates> </gml:linearring> </gml:polygon> 9

10 Challenge 1: No Pre-Processing Need to transform between coordinate reference systems (CRS) Need to adaptively discretize shape boundaries E.g., introduce vertices for ellipsoidal shortest-path interpolation (geodesics) <gml:polygon srsname="epsg:4326"> <gml:linearring> <gml:coordinates> lon0, lat0, lon1, lat1,... </gml:coordinates> </gml:linearring> </gml:polygon> Need to triangulate interior on-the-fly Custom high-performance tesselation algorithm Clip against projection boundaries Non-planar tesselation in 3D 10

11 Challenge 2: Big Data Can expect a large number of big data sets Hundreds of data sources not uncommon Often covering the same area Many pixels touched multiple times Also in 2D due to layered rendering 11

12 Challenge 2: Big Data Keep object count manageable Frustum culling Object-size culling Distance-based culling and LOD Scale-based culling and LOD Adaptive scale-based discretization Pixel-density based culling and LOD Label deconfliction 12

13 Challenge 3: Customizable and Extensible API Our API allows to dynamically generate any geometry and style for any input object public class MyStyler extends Styler public void style( Collection aobjects, StyleCollector astylecollector) { Example: Input = Shape and properties of France Output = Animated population pie chart for (Object object: aobjects) { astylecollector.object( object ).geometry( getgeometry ( object ) ).styles( getfillstyle( object ), getlinestyle( object )).submit(); 13

14 Challenge 3: Customizable and Extensible API Our API allows to dynamically generate any geometry and style for any input object Customers want performance without any fuss Avoid state changes, group commonly styled shapes in batches Dynamic texture atlases for icons, labels, and textures Order-independent rendering algorithms Example: Input = Shape and properties of France Output = Animated population pie chart 14

15 Challenge 3: Customizable and Extensible API Customers need capability to perform own OpenGL rendering public class MyPainter extends Painter public void paintobjects( Collection aobjects ) { // Do any OpenGL painting here gl.gldrawarrays(...); and automatically want to benefit from Single code path for 2D and 3D rendering Automatic draping on terrain if needed Correct handling of transparent shapes 15

16 Challenge 3: Customizable and Extensible API Work with 3 global paint passes 1. Paint draped shapes 2. Paint non-draped opaque shapes 3. Paint non-draped transparent shapes 16

17 Challenge 3: Customizable and Extensible API Work with 3 global paint passes 1. Paint draped shapes 2. Paint non-draped opaque shapes 3. Paint non-draped transparent shapes 17

18 Challenge 3: Customizable and Extensible API Work with 3 global paint passes 1. Paint draped shapes 2. Paint non-draped opaque shapes 3. Paint non-draped transparent shapes 18

19 Foundations Our Vision GPU Acceleration Main Challenges No Pre-Processing Big Data Customizable and Extensible API Exploiting the GPU Orthorectified Video Feeds Dynamic Plot Filtering and Styling Line-of-Sight Computations 19

20 Exploiting the GPU 1: Orthorectified Video Feeds Improve Situational Awareness by draping UAV/Satellite video feeds on terrain (in 3D and 2D) 20

21 Exploiting the GPU 1: Orthorectified Video Feeds This process is called orthorectification Trace ray back from terrain to UAV camera Sample pixel color from UAV video frame UAV camera frustum Wait - isn t this projective texturing? Not quite Treated as separate layer to support hundreds of video feeds and arbitrary layer ordering UAV CRS not necessarily the same as the terrain CRS Additional vertex attribute corresponding to 3D terrain coordinates in UAV CRS Also works in 2D, in any projection terrain 21

22 Exploiting the GPU 1: Orthorectified Video Feeds GPU benefits Orthorectification almost as efficient as simple texture lookup Can handle multiple UAV feeds with almost no CPU overhead Improved Situational Awareness by combining video feeds with other layers GPU technical hindsight Avoid floating point accuracy issues by using local coordinate origins 22

23 Exploiting the GPU 2: Dynamic Plot Filtering and Styling Satellite AIS - Automatic ship identification and tracking system Installed on thousands of ships worldwide > 500,000,000 position updates per day 23

24 Exploiting the GPU 2: Dynamic Plot Filtering and Styling Filter and style using expressions based on plot properties Visually detect patterns, anomalies, Example: Only paint plots in time range Expression visibilityfunction = between( timestampattr, timeminparam, timemaxparam ); 24

25 Exploiting the GPU 2: Dynamic Plot Filtering and Styling GPU implementation Expression visibilityfunction = between( timestampattr, timeminparam, timemaxparam ); #version 110 attribute float timestamp; uniform float timemin; uniform float timemax; void main() { if (timestamp>=timemin && timestamp<=timemax) { // Compute dynamic styling // Transform vertex to view coordinates else { // Discard vertex 25

26 Exploiting the GPU 2: Dynamic Plot Filtering and Styling Parameters become vertex shader uniforms Parameter timeminparam = parameter( "timemin", getmintime() ); Expression visibilityfunction = between( timestampattr, timeminparam, timemaxparam ); #version 110 attribute float timestamp; uniform float timemin; uniform float timemax; void main() { if (timestamp>=timemin && timestamp<=timemax) { // Compute dynamic styling // Transform vertex to view coordinates else { // Discard vertex 26

27 Exploiting the GPU 2: Dynamic Plot Filtering and Styling Attribute expressions become vertex attributes Expression timestampattr = attribute( "timestamp", (AttributeValueProvider) (aobject) -> { return ((AISPlot)aObject).getTimeStamp(); ); Expression visibilityfunction = between( timestampattr, timeminparam, timemaxparam ); #version 110 attribute float timestamp; uniform float timemin; uniform float timemax; void main() { if (timestamp>=timemin && timestamp<=timemax) { // Compute dynamic styling // Transform vertex to view coordinates else { // Discard vertex 27

28 Exploiting the GPU 2: Dynamic Plot Filtering and Styling Resulting expression compiled to GLSL vertex shader code Expression visibilityfunction = between( timestampattr, timeminparam, timemaxparam ); #version 110 attribute float timestamp; uniform float timemin; uniform float timemax; void main() { if (timestamp>=timemin && timestamp<=timemax) { // Compute dynamic styling // Transform vertex to view coordinates else { // Discard vertex 28

29 29 Exploiting the GPU 2: Dynamic Plot Filtering and Styling

30 Exploiting the GPU 2: Dynamic Plot Filtering and Styling GPU benefits Highly parallelizable dynamic filtering and styling on millions of plots Complex and dynamic expressions and styling Very smooth interactive user experience GPU technical hindsight Dynamic GLSL shader generation Mixed experiences with asynchronous shader compilation implemented using context sharing 30

31 Exploiting the GPU 3,4,5, : Demo! Line-of-Sight Analysis Terrain Analysis Image Processing Complex Stroking Cross-terrain Routing 31

32 Conclusion Situational Awareness applications require connecting to, visualizing, and analyzing any data source on-the-fly Such applications preclude pre-processing and optimizing the data as is done for example in geodatabases, computer games, and classical mapping engines Performance of GPU allows handling real-time, big, and dynamic data sets enabling a new generation of Situational Awareness applications 32

33 Thank You 33