Isosurface extraction and interpretation on very large datasets

Isosurface extraction and interpretation on very large datasets in geophysics Guilhem Dupuy LIUPPA Bruno Jobard LIUPPA, INRIA-Magique 3 D Sébastien Guillon Noomane Keskes Total Guilhem Dupuy 6/19/2021 Dimitri Komatitsch MIGP, INRIA-Magique 3 D

Outline • Work context • Tandem algorithm • Our approach : a component-based parallel isosurface extraction • Application on realistic case • Conclusion Guilhem Dupuy 6/19/2021 2

Input dataset : seismic attribute • Example of seismic attribute : § Highlight chaotic and high amplitude area • Dataset size often > 50 Go • Isosurfaces on noisy attributes § millions of triangles § many disconnected components Guilhem Dupuy 6/19/2021 3

Outline • Work context • Tandem algorithm • Our approach : a component-based parallel isosurface extraction • Application on realistic case • Conclusion Guilhem Dupuy 6/19/2021 4

Tandem algorithm [Attali & al, 2005] • Extension of the marching cubes • Extraction and simplification in tandem : § Alternate an extraction of a layer with a simplification § New metric formulation to simplify and optimize triangles’ quality § Notion of time lag to prevent simplification artifacts Guilhem Dupuy 6/19/2021 5

How to prevent artifacts time lag • Without Edge contractions are scheduled in order to prevent artifacts due to bias in shape during reconstruction • Notion of time lag : With time lag X front • Initially height(u)=uz and rad(u)=1 • The contraction ab → c is prevent as long as reach(c)>front Guilhem Dupuy 6/19/2021 6

A new approach suited for complex geological environments [Pivot & al, 2007] Input relevant attribute (volume) + Geological interpretation (surfaces) constrained surface extraction Tandem algorithm Interpretationconsistencywithfaultsand sedimentaryarchitecture, … Size of the advancing front Simplification could be insufficient to fit in memory Too time consuming Filtering Guilhem Dupuy 6/19/2021 7

Outline • Work context • Tandem algorithm • Our approach : a component-based parallel isosurface extraction • Application on realistic case • Conclusion Guilhem Dupuy 6/19/2021 8

Our approach : a component-based parallel isosurface extraction • Processing : § Recursive binary splitting of the dataset § Process each separately § Preserve refinements for cuts § Stitch parts back together • Constraints : § Minimize memory consumption = merge surface with maximum interface § Various number of processors during computation Guilhem Dupuy 6/19/2021 9

Our approach : a component-based parallel isosurface extraction binary splitting triangulated surface sub bloc … triangulated surface … 6/19/2021 sub bloc … Guilhem Dupuy … Dataset merge sub bloc triangulated surface 10

Criterion for merging parts of the dataset • We define type of cut results: § left opposite to right § up opposite to down § front opposite to bottom • A part could be merge with its brother and children excepted them tagged with an opposite type 6/19/2021 right O left • If a part is forbidden for a merge then its children are also forbidden Guilhem Dupuy brother left right left OO 11 X right X X

Extension of the time lag during bloc merging Without time lag With time lag Guilhem Dupuy 6/19/2021 12

Dumping completed components • Observation : § In geophysics isosurfaces generate many disconnected components § A component which was extracted and simplified use memory • Solution : § Dump completed components in a streamed file format to release memory Guilhem Dupuy 6/19/2021 13

Dumping completed components • Streaming format [Isenburg & Lindstrom, 2005] is used to append component to current output • Accompanying metadata index Standard Format Guilhem Dupuy 6/19/2021 14 Streaming Format

Parallel evaluation : efficiency and speedup • seq. Time : runtime with 1 processor • p : number of processor • para. Time(p) : runtime with p processors • Efficiency is given by : • Speedup is given by : Guilhem Dupuy 6/19/2021 15

Outline • Work context • Tandem algorithm • Our approach : a component-based parallel isosurface extraction • Application on realistic case • Conclusion Guilhem Dupuy 6/19/2021 16

Now we have a new problem What can we do with that ? Guilhem Dupuy 6/19/2021 17

Exploration of resulting isosurfaces • Study of connected components § Computation of metadata : volume, main direction, … • New storage of triangular meshes § Two files : description and raw data § Components sorting based only on description file Component 1 Streamed component 1 volume, oriented bounding box, … Component n Streamed component n volume, oriented bounding box, … Guilhem Dupuy 6/19/2021 18

The use of metadata : selection of a subset of components • Many filters possibility : § Statistical-based filters § Geometric-based filters • Example : § Volume-based filter § 18 111 to 328 disconnected components Guilhem Dupuy 6/19/2021 19

Outline • Work context • Tandem algorithm • Our approach : a component-based parallel isosurface extraction • Application on realistic case • Conclusion Guilhem Dupuy 6/19/2021 20

Conclusion • A binary splitting strategy designed to improve parallel isosurface extraction scalability § Extension of the time lag improve mesh quality § Merging strategy to reduce synchronisation delay § Adapted to all parallel strategy • A component based dumping method § Another way to improve scalability § Intrinsic result file organisation allow an efficient exploration Guilhem Dupuy 6/19/2021 21

Any Questions ? Guilhem Dupuy 6/19/2021 22

Back up Guilhem Dupuy 6/19/2021 23

Geological constraints • Introduction of geological knowledge : § New limits in the dataset § Control connectivity of objects Guilhem Dupuy 6/19/2021 24

How to prevent artifacts • We need to schedule edge contractions to prevent artifacts due to bias in shape during reconstruction • Notion of time lag : • Initially height(u)=uz and rad(u)=1 • The contraction ab → c is prevent as long as reach(c)>front Guilhem Dupuy 6/19/2021 25

Effect of the time lag during extraction Without time lag With time lag Computations are done with same error criteria and isotropy parameter Guilhem Dupuy 6/19/2021 26

Used quadric error matrix for edge collapse Q= q 00 q 10 q 20 q 30 q 01 q 11 q 21 q 31 q 02 q 12 q 22 q 32 q 03 q 13 q 23 q 33 figure courtesy of Michael Garland [Garland & Heckbert 1997] Hc represents displacement introduce by contractions Gc represents mesh anisotropy introduce by contractions is named isotropy parameter Guilhem Dupuy 6/19/2021 27

The used quadric error Consider edge contraction ab → c as a function that maps vertices a and b to c Simplification is a composition of edge contractions Represent displacement introduce by the contractions Uc is the set of the triangles incident to vertices in the initial triangulation that map to c Py is the plane spanned by triangle d(x, Py) is the Euclidean distance between x and its closest point on P y Guilhem Dupuy 6/19/2021 28

The used quadric error Consider the edge contraction ab → c, Sab are the triangles containing a, b or both Represent mesh anisotropy introduce by the contraction Guilhem Dupuy 6/19/2021 29

The used quadric error The candidate for contraction c minimize : where W is a term for normalization : E 0 is the error threshold and [0. . 1] represents a compromise between shape measure and anisotropy measure Guilhem Dupuy 6/19/2021 30

Inertia matrix • Considering the triangle t defined by points a, b and c and • Its centroid is notated : • Inertia matrix is defined by: Guilhem Dupuy 6/19/2021 31

Mesh quality • Considering and the greatest eigenvalues of the inertia matrix of a triangle • Isotropy of a triangle is defined by • Anisotropy of a mesh K is defined by Guilhem Dupuy 6/19/2021 32

Dataset splitting Guilhem Dupuy 6/19/2021 33

Mesh Cutting • • Cut dataset into small pieces Process each separately Preserve refinements for cuts Stitch dataset parts back together figure courtesy of Hugues Hoppe Guilhem Dupuy 6/19/2021 34

Dumping unused components Guilhem Dupuy 6/19/2021 35

Parallel evaluation : runtime Guilhem Dupuy 6/19/2021 36

Streaming format for triangular meshes Classical format V V V F F 0. 3 0. 4 0. 9 1. 0 1. 4 1 1 3 3 1. 1 0. 4 0. 5 0. 1 0. 8 2 3 2 5 0. 2 0. 5 0. 7 1. 1 1. 2 3 4 5 4 Pre order format V V V F V F F 0. 3 0. 4 0. 9 1 1. 0 -1 1. 4 -3 -3 1. 1 0. 4 0. 5 2 0. 1 3 0. 8 -2 -5 1 0. 2 0. 5 0. 7 3 1. 1 4 1. 2 5 -4 1 2 3 3 4 2 4 5 • In classical format we need to read all the vertices to instantiate facets • Pre order format could be see as a stream where vertices are described just before a facet need them. • A negative value coincides with the finalization of a vertex. Guilhem Dupuy 6/19/2021 37

Proposed pipeline [Isenburg & Lindstrom, 2005] regular volume grid V V V F ⋮ 256 1. 32 0. 12 0. 23 1. 43 0. 23 0. 92 0. 91 0. 15 0. 62 1 -2 3 0. 72 0. 34 0. 35 4 -1 3 ⋮ ⋮ ⋮ 256 smextract | smclean | smsimp | smcompress P 1 P 2 P 3 grid. raw Guilhem Dupuy 6/19/2021 P 4 mesh. smc 38