Directable, High-Resolution Simulation of Fire on the GPUSimulation of Fire on the GPUPresentation Date: Sep 22ndChrissie Chi CuiOutline• Introduction• Related work• Highlights•Input Particle System•Input Particle System• View-Specific Fire Refinement• Fire Volume Rendering• Experiment Results• Conclusions & Future WorkIntroduction• Popular techniques in special movie effects• Challenging visual characteristics for simulation– Flickering nature, rapid time-evolution– Large gradients of temperature, turbulent– Vortical behavior in comparatively calm systems•Very fine spatial and temporal resolution • Large Scale combustion Simulation for artist-driven, visual effects production: intractable• Previous methods:– Compositing large amount of filmed, extracted fire:Lack details, not adaptive, E.g. “The Lord of the Rings”– Solving 3D Naiver-Stokes Equations for velocity, density and temperature: Intractable for large spatial temporal resolution, E.g. “The matrix: Reloaded”Related Work• Particle based: – Model individual flames with chain of particles – Lack realistic details and coherence structures• Grid Simulation: – Solving 3D NV Equation –Hard to model the highly turbulent detail of real fire–Hard to model the highly turbulent detail of real fire• Combination of Particle and grid simulation:– A view dependant way– Limited to phenomena that can be decomposed into one or more 2-dimensional slices– Problematic when problems of rendering coherent structures from individual particles domainHighlights• Produce extremely detailed and can be integrated seamlessly into film-resolution images.• A novel combination of coarse particle grid simulation with very fine, view-oriented refinement simulations performed on a GPU.– Split the refinement stage into separable parallel tasks by considering. perceptive limitations likely viewing behavior–Employ multiple independent GPUs for final simulations –Employ multiple independent GPUs for final simulations refinement• A simple, GPU-based volume rendering scheme. • Directability:– Allowing virtually any user-defined particle behavior as an input to the initial coarse simulation. – Minimum physical criteria enforced by the coarse stage– Utilizable on consumer-grade GPUsInput Particle System• A visually plausible fluid-like motion• The particle simulation: fully directable, easy to control and fast• Particle counts: from 20,000 to 100,000 per frame• Particle Attributes:Vectors: position( ), velocity( )–Vectors: position( ), velocity( )– Scalars: radius(Ri), fuel(Ki), mass(Mi), impulse(Ji), Age(Ai)• Implement obstacles or disturbances as “impulse-only” particles – non-zero values for impulse,– zero values for fuel and mass,Coarse Grid Step• Small grid size: no larger than 50×50×50.• Efficiently obtaining an approximate of non-divergent particle velocity• Inserted into the update step between the calculation of the new velocity and the calculation of the new position•Augment PIC/FLIP by using wavelet decomposition on the •Augment PIC/FLIP by using wavelet decomposition on the velocity grid to produce multiple levels of detail.– Amplify vorticity at multiple scales – Accelerating convergence of the iterative incompressibility solution.• Enforce incompressibility at each detail level from the lowest level• Solving pressure Poisson Equation using Jacobi iterationsCoarse Grid Step AlgorithmView Specific Slice Refinement Overview• Goal: – Add high resolution details– Add fire-like physics• Observation Grounds for view specific refinement:– Fire is very fast-moving and turbulent, even in calm situations. – Fine details are short-lived and tend to change quickly. –The turbulent, fine details do not have a large visual impact on the overall motion of the flame flow. –The turbulent, fine details do not have a large visual impact on the overall motion of the flame flow. – From a particular viewpoint, fine variations of detail and movement perpendicular to the projection plane are not individually visible so they do not affect the large-scale flow very much.• To-do-list: – Create a set of independent image planes parallel to a camera viewing plane; – Project attributes, e.g. forces and fuel, from the Coarse Grid Step particles onto these planes; – Solve the two dimensional, incompressible Navier-Stokes equations for each image plane on GPUs independently.View Specific Slice Refinement DemoFigure. Refinement simulation slices are aligned to the projection plane. They are spaced coarsely and evenly along the projection axis. Slices have very high resolution in the dimension parallel to the projectionView Specific Slice Refinement stepsGPU Configuration• The 2D refinement simulations are computed entirely on the GPU• Built in OpenGL2.1 using GLSLSlice Refinement Details I• Compare to the work done by Harris:– New ordering of the major sub-steps in the update step– The addition of temperature, cooling, and thermal buoyancy• Inputs from Particles: Particle Fuel, Particle Mass, Particle Weighted Velocity• Simulation State: Density, Temperature, Texture, Fuel, Velocity, Artificial Pressure• Output per Slice:–Image I: simDensity(Ch.I), simTemperature(Ch.II), textureDetail(Ch. III) and –Image I: simDensity(Ch.I), simTemperature(Ch.II), textureDetail(Ch. III) and fluidFuel (Ch. IV)– Image II: simVelocityU (Ch. I) and simVelocityV (Ch. II)• Particle Projection onto refinement slices– Create three planes of data: Fuel( adding temperature), Mass(adding density), WeightedVelocity(an impulse, ultimately adding velocity).– Used additively with the existing simulation– Each particle’s contribution to the slice is weighted based on its perpendicular distance from the slice, using a Gaussian kernel.The uniform distance between slicesPerpendicular distance between particle i and this sliceSlice Refinement Details II• Cooling, dissipation, heating and density step– the first of the “slab” operations: several two-dimensional data planes are operated upon to create an output data plane of exactly the same rectangular size– Highly parallelizable on GPUs– Cooling:– Dissipate:– Heating:– Density:• Boundary Conditions– Collision only resolved in the input simulation–“impulse only particles” still affect simulation velocity in the refinement stage–“impulse only