(click to see it in action)
In this article I will discuss about something which is best known as numerical integration. I have always been fascinated by the cloth simulations, and wondered how it can be possible to reproduce this visually appealing effect using Actionscript, knowing that the AVM (Adobe Virtual Machine) is lacking well behind other languages like Java, not to mention C++ in terms of pure computation performance. Well to my surprise, after implementation I managed to obtain a pretty solid frame rate.
The concept
The core algorithm of the simulation is based on two concept known in physics as the Newton law's of motion and restoration of force. This method is reasonably simple and robust and is a good general candidate for numerical solution of differential equations when combined with an intelligent adaptive step-size routine. After surfing the net for some tips and inspiration and reading Keith Peter's book (AdvancED ActionScript 3.0 Animation), I got a grasp about the whole concept and physics acting behind it.
The last temptation to implement my version came when I saw this amazing experiment implemented in Processing.
The backbone of every clothing simulation is the so called numerical integration. There are many different techniques and implementation for physics simulation but, most of them use one of the three widespread integration method: Euler integration, Verlet integration and Runge-Kutta. The most significant difference between these methods consist in the accuracy of the movement.
The Euler integration is the simplest one and uses the following algorithm: first calculate the actual position of the object, concatenate by the velocity and then apply acceleration. This is a very straightforward approximation with a very good performance, only it's not accurate (unless we use an extremely small timestep size - which we have no control over).
The Verlet integration is very similar with Euler integration only this time we'll get rid of the velocity component. We only need to calculate the change in position applying the acceleration value (A = F/m) which is a scalar component (the position and velocity actually are vectors). This is however velocity. So knowing the previous position and acceleration we could calculate the current position at the next frame. Keith Peters explain the differences between implementations in a very fashionable and enjoyable manner, so you should understand it easily. However my implementation differs quite enough from that one used in Keith's book.
(click to see it in action)
The implementation
In this experiment i'm gonna use Runge Kutta or simply RK4 implementation, but we can implement almost with the same accuracy the Verlet integration too. Very simply put, the algorithm first calculate the position and velocity at a certain interval (this calculation is done 4 times) and then make an approximation of them, and the average should be the object position and velocity at a certain position in time. The strength of this method is it's accuracy, considering the minimal performance loss (i got almost constant 60 fps on my machine).
Translated to AS3 it looks like this:
/**
* Get initial position, K1 velocity
* apply forces and velocity, the result is K2
*/
for (i = 0; i < numPart; i++)
{
p = particles[i];
if (!p.fixed)
{
originalPos = originalPosV[i];
k1Vel = k1VelV[i];
s = (0.5 * t);
k1Vel.x *= s;
k1Vel.y *= s;
k1Vel.z *= s;
p.position.x = originalPos.x + k1Vel.x;
p.position.y = originalPos.y + k1Vel.y;
p.position.z = originalPos.z + k1Vel.z;
originalVel = originalVelV[i];
k1Force = k1ForceV[i];
s = (0.5 * t / p.mass);
k1Force.x *= s;
k1Force.y *= s;
k1Force.z *= s;
p.velocity.x = originalVel.x + k1Force.x;
p.velocity.y = originalVel.y + k1Force.y;
p.velocity.z = originalVel.z + k1Force.z;
}
}
particleSystem.applyForces();
for (i = 0; i < numPart; i++)
{
p = particles[i];
if (!p.fixed)
{
k2ForceV[i].x = p.force.x;
k2ForceV[i].y = p.force.y;
k2ForceV[i].z = p.force.z;
k2VelV[i].x = p.velocity.x;
k2VelV[i].y = p.velocity.y;
k2VelV[i].z = p.velocity.z;
}
p.force.x = 0;
p.force.y = 0;
p.force.z = 0;
}
/**
* Get initial position, K2 velocity
* apply forces and velocity, the result is K3
*/
for (i = 0; i < numPart; i++)
{
p = particles[i];
if (!p.fixed)
{
originalPos = originalPosV[i];
k2Vel = k2VelV[i];
s = (0.5 * t);
k2Vel.x *= s;
k2Vel.y *= s;
k2Vel.z *= s;
p.position.x = originalPos.x + k2Vel.x;
p.position.y = originalPos.y + k2Vel.y;
p.position.z = originalPos.z + k2Vel.z;
originalVel = originalVelV[i];
k2Force = k2ForceV[i];
s = (0.5 * t / p.mass);
k2Force.x *= s;
k2Force.y *= s;
k2Force.z *= s;
p.velocity.x = originalVel.x + k2Force.x;
p.velocity.y = originalVel.y + k2Force.y;
p.velocity.z = originalVel.z + k2Force.z;
}
}
particleSystem.applyForces();
for (i = 0; i < numPart; i++)
{
p = particles[i];
if (!p.fixed)
{
k3ForceV[i].x = p.force.x;
k3ForceV[i].y = p.force.y;
k3ForceV[i].z = p.force.z;
k3VelV[i].x = p.velocity.x;
k3VelV[i].y = p.velocity.y;
k3VelV[i].z = p.velocity.z;
}
p.force.x = 0;
p.force.y = 0;
p.force.z = 0;
}
/**
* Get initial position, K3 velocity
* apply forces and velocity, the result is K4
*/
for (i = 0; i < numPart; i++)
{
p = particles[i];
if (!p.fixed)
{
originalPos = originalPosV[i];
k3Vel = k3VelV[i];
s = (0.5 * t);
k3Vel.x *= s;
k3Vel.y *= s;
k3Vel.z *= s;
p.position.x = originalPos.x + k2Vel.x;
p.position.y = originalPos.y + k2Vel.y;
p.position.z = originalPos.z + k2Vel.z;
originalVel = originalVelV[i];
k3Force = k3ForceV[i];
s = (0.5 * t / p.mass);
k3Force.x *= s;
k3Force.y *= s;
k3Force.z *= s;
p.velocity.x = originalVel.x + k3Force.x;
p.velocity.y = originalVel.y + k3Force.y;
p.velocity.z = originalVel.z + k3Force.z;
}
}
particleSystem.applyForces();
for (i = 0; i < numPart; i++)
{
p = particles[i];
if (!p.fixed)
{
k4ForceV[i].x = p.force.x;
k4ForceV[i].y = p.force.y;
k4ForceV[i].z = p.force.z;
k4VelV[i].x = p.velocity.x;
k4VelV[i].y = p.velocity.y;
k4VelV[i].z = p.velocity.z;
}
p.force.x = 0;
p.force.y = 0;
p.force.z = 0;
}
/**
* Update initial position and velocity based on intermediate values
*/
for (i = 0; i < numPart; i++)
{
p = particles[i];
if (!p.fixed)
{
//position
originalPos = originalPosV[i];
k1Vel = k1VelV[i];
k2Vel = k2VelV[i];
k3Vel = k3VelV[i];
k4Vel = k4VelV[i];
k2VelInt.x = k2Vel.x;
k2VelInt.y = k2Vel.y;
k2VelInt.z = k2Vel.z;
s = (2);
k2VelInt.x *= s;
k2VelInt.y *= s;
k2VelInt.z *= s;
k3VelInt.x = k3Vel.x;
k3VelInt.y = k3Vel.y;
k3VelInt.z = k3Vel.z;
s = (2);
k3VelInt.x *= s;
k3VelInt.y *= s;
k3VelInt.z *= s;
intVel.x = k1Vel.x + k2Vel.x + k3Vel.x + k4Vel.x;
intVel.y = k1Vel.y + k2Vel.y + k3Vel.y + k4Vel.y;
intVel.z = k1Vel.z + k2Vel.z + k3Vel.z + k4Vel.z;
s = (t / 6);
intVel.x *= s;
intVel.y *= s;
intVel.z *= s;
p.position.x = originalPos.x + intVel.x;
p.position.y = originalPos.y + intVel.y;
p.position.z = originalPos.z + intVel.z;
// velocity
originalVel = originalVelV[i];
k1Force = k1ForceV[i];
k2Force = k2ForceV[i];
k3Force = k3ForceV[i];
k4Force = k4ForceV[i];
k2ForceInt.x = k2Force.x;
k2ForceInt.y = k2Force.y;
k2ForceInt.z = k2Force.z;
s = (2);
k2ForceInt.x *= s;
k2ForceInt.y *= s;
k2ForceInt.z *= s;
k3ForceInt.x = k3Force.x;
k3ForceInt.y = k3Force.y;
k3ForceInt.z = k3Force.z;
s = (2);
k3ForceInt.x *= s;
k3ForceInt.y *= s;
k3ForceInt.z *= s;
intForce.x = k1Force.x + k2Force.x + k3Force.x + k4Force.x;
intForce.y = k1Force.y + k2Force.y + k3Force.y + k4Force.y;
intForce.z = k1Force.z + k2Force.z + k3Force.z + k4Force.z;
s = (t / 6 * p.mass);
intForce.x *= s;
intForce.y *= s;
intForce.z *= s;
p.velocity.x = originalVel.x + intForce.x;
p.velocity.y = originalVel.y + intForce.y;
p.velocity.z = originalVel.z + intForce.z;
}
}(For a comparison between different numerical integration methods read this article: http://codeflow.org/entries/2010/aug/28/integration-by-example-euler-vs-verlet-vs-runge-kutta/.)
Here is the final result (press space to show/hide the stats).
Press right arrow to show wire frame and texture simultaneously, up arrow to show only bitmap texture and left arrow to show only the wire-frame.