Fluid Simulation using Navier-Stokes Equations

Simulation of Fluids using the Navier-Stokes Equations Kartik Ramakrishnan

Presentation Scope Introduction Theory Adaptation to CPU and GPU Results Optimizations and Future Work

Introduction Physically based simulation of fluids produces very good visual effects Computationally expensive process In the past,usually done offline and not in real time Real time fluid simulation was done using particle systems Current hardware makes physically based fluid simulation possible in real time

Navier-Stokes Theory The fluid being rendered is assumed to be incompressible (fluid volume remains constant over time) The fluid being rendered is assumed to be homogeneous (fluid density remains constant in space)

Navier-Stokes Theory RHS of the equation, proceeding left to right, Advection/velocity term Pressure term Diffusion term Force term

Navier-Stokes Theory Advection - Describes the transport mechanism of the fluid - Also describes the transport mechanism of any material the fluid may be carrying - For example , ink in water, the advection term describes the movement of both the water and the ink

Navier-Stokes Theory Diffusion - Describes the resistance of the fluid to flow - Viscosity of the fluid is a factor in this term - More viscous the fluid, the more resistant it is to flow - Impedes the velocity of the fluid

Navier-Stokes Theory Pressure - Describes the behavior of the fluid due to pressure induced among fluid particles - Think fluid particles squished against each other due to the application of force - Pressure leads to increase in the velocity

Navier-Stokes Theory Force - Describes the behavior the fluid due to any externally applied force or any bodily force (gravity) - Application of force causes an increase in the fluid velocity

Adaptation to CPU and GPU The NS equation is not in a finite form Need to reduce each of the above RHS terms to a discrete representation that can be implemented on the CPU and the GPU

Adaptation to CPU and GPU Finite representation of the NS Equation

Adaptation to CPU and GPU Discrete NS algorithm - Assume a 2D bounding grid for the fluid - Assume one such grid for each fluid property of interest (velocity, pressure, force, dye etc.) - Assume that velocity and pressure are 0 at time t = 0 - Assume that velocity is negative and that pressure is 0 at the grid boundaries

Adaptation to CPU and GPU Algorithm steps - Apply force: w1 = u + Force* timestep - Advect: w2 = w1(x timestep*w1) - Diffuse term: del2 w3 = w2 - Pressure term: del2 p = w3 - Subtract pressure gradient: w4 = w3 del p

Adaptation to CPU and GPU void Fluids::Simulate() { //Advect Advect((); SwapVelocityGrid(); SwapColorGrid(); //Add force AddForce(); SwapVelocityGrid(); //Diffuse Diffuse(); SwapVelocityGrid(); //Compute Pressure ComputePressure(); SwapPressureGrid(); //Project Project(); SwapVelocityGrid(); SetBounds(); } timestep++;

Results AMD Athalon 3200+ 2.19 GHz, 2GB RAM AMD Radeon 4870 1GB Brook+ 64x64 fluid grid, 50 frames and 70 jacobian iterations CPU (averaged over 50 frames) = 0.83 sec GPU (averaged over 50 frames) = 0.035 sec

Optimizations and Future Work Use float 4 for the pressure computation instead of float Pinned memory, new to Brook+1.4 Shared memory, supposed to be implemented, but think its not GPU version still glitchy needs a few tweaks Other fluid grid sizes

Distant future Better visuals Port to DirectX 10 : dedicated texture hardware support available, bilinear interpolation available for free

Questions