A Complete XBox 360 GPU Particle System: Tech & Pipeline

Sebastian Sylvan & Dr Simon ScarleRare MGS

A Complete XBox 360 GPU Particle System: Tech & Pipeline

Sebastian Sylvan & Dr Simon ScarleRareMGS

A Note on Applicability

What is a Particle System?

Tech overview•Why on the GPU?• Some background• Simulation• Sorting•Rendering

Why on the GPU?

Why on the GPU?

No Synchronisation

Headaches

Why on the GPU?

Performance

Why on the GPU?

Code Specialisation

Background : The Xbox 360

Background: The Xbox 360

Flexible fetching

VS_OUTPUT vs_main( int ix : INDEX ){

int index = (ix+0.5)/a_const.x; float4 pos;

float2 uv;asm{

vfetch pos, index, position0;vfetch uv, index, texcoord0;

};/* .... snip .... */

}

Background: The Xbox 360

Memory Export

Main Memory

Vertex Shade

r

Pixel Shad

er

Frame

BufferDIP

ME

ME

Simulationvoid simulation_shader( int ix : INDEX ){

Particle p = decompress(fetch_particle(ix));

Particle newp = simulate( p, ix, delta_time );

write_particle( ix, compress(newp) );}

Simulation

Newtonian motion

// a = (sum of forces) / massnewp.vel = p.vel + dt*a;newp.pos = p.pos + p.vel*dt;

Simulation

Turbulence :Position-dependent

random forcea += f(tex3D(turb_tex, g(p.pos)));

Scale/bias position into “turbulence

space”

Per-system scale/bias

Simulation

Turbulence:Holy Cache

Thrashing, Batman!

Simulation

Alleviating Cache Thrashing

•Use small texture•Use “thin” format (e.g. 10:11:11)

•Sort the particles

Simulation

Collision detection

(spheres, planes, height fields)

Simulation

Emissionfloat4 rnd;float x = (ix+rnd_offset) % rnd_size;asm{ tfetch1D rnd, x, rnd_tex, UnnormalizedTextureCoords=true};

Particle newp = create_particle(rnd);rnd = ix > half_N ? rnd.xyzw : rnd.wzyx;

Sorting

Sorting

Why sort?• Cache coherency

• Non-commutative rendering

Sorting

Why sort?

Sorting

Need parallel sort for parallel hardware

Parallelize standard algorithms?Sorting networks?

Sorting

Sorting Networks(Bitonic merge sort, odd-even

merge sort)

Sorting

Odd-Even mergemerge(s){ n = size(s) if ( n == 2 ) { return CAS(s(0),s(1)) } else { odds = s(1,3,..,n-1) evens = s(0,2,..,n-2) result = interleave( merge(evens), merge(odds))

for ( i = (1,3,..,n-2) ) { (x,y) = CAS(result(i), result(i+1)) result(i)= x result(i+i)= y } }}

Merge the two halves of the array s

Sorting

Odd-Even merge2 3 8 12 1 6 7 10

2 8 6 103 121 7

2 1 8 7 3 6 12 10

1 7 2 8 3 10 6 12

1 3 2 6 7 10 8 12

Sorting

Odd-Even merge sort

Sorting

Odd-Even merge sortNested algorithm, needs manual

flattening for parallelism:

Instant Insanity!

Sorting

Advice for staying sane

•Write “twin” version in GPPL• Debugging: step through PIX and

your “twin” simultaneously

Sorting

Or, just copy my stuff!

Google: “Particle System Simulation and Rendering on the Xbox 360 GPU“

Sorting

Sorting for cache coherency

Sorting

Sorting performance

Sorting

When incremental sorting fails

Whenever there the position changes too quickly

Prime example: Respawns Solution: Delay simulation/rendering of newly spawned particle

for a few frames for sorting

Rendering

Rendering

Point sprites

Rendering

Geometry AmplificationProduce N vertices for each particle

Index Buffer

Particle Buffer

...

...

Results

Results

Results

Results

A Complete XBox 360 GPU Particle System: Tech & Pipeline

Sebastian Sylvan & Dr Simon ScarleRare MGS

A Complete XBox 360 GPU Particle System: Tech & Pipeline

Sebastian Sylvan & Dr Simon ScarleRareMGS

An Complete XBox 360 GPU Particle System: Tech & Pipeline

Sebastian Sylvan & Dr Simon ScarleRare MGS

almost

^

Old System

1. Artist wants a new particle system2. Artist talks to Programmer3. Programmer implements code/shaders to

produce the effect4. Artist looks at particles and wants it exactly

the same only completely different5. Goto 2 : Rinse & Repeat

New System

• Rare usually takes an Art Driven approach• Creation of particle effects via art tool• Sebastian's Tech demos implemented in our

pipeline with this in mind– highly customisable– efficiency

Overview

Artist Input

Runtime Asset

Interim Build Format

Artist Input

• Particle.mb produced by Maya– Main art package at Rare– Use Maya particle options as base

• Extend – custom attributes on particle nodes• Random Scaling/orientation• Soft Particles• Timing

Interim Build Format

• Parse particle data from artist input– Rparticle

• Currently undergoing a de-Maya-ification process– i.e. add another rparticle producer– rparticle objects produced in WorldBuilder/Viewer

Rparticle

• Collection of arrays of custom data storage classes – RparticleObject – RparticleEmitter– RparticleField– RparticleRamp

• And a couple of R1 data classes– RmodelShader– RmodelNurbsCurve

Runtime Asset

Simulation Shader

Parameters

Render Shader

(Sorting Shader)

Simulation Shader

• Lego-ized/modular construction• “Death to the Über-Particle by meta-coding”• RparticleObject is parsed into two lists of

ISimulationFragment– Initialisation– Simulation

• Fragments output needed HLSL to build up the shader

Simulation Shader

• Helper mini-fragments• Initialisation Frags– Create a new particle– Depends on emitter

• Directional• Omni-directional• Volume• Curve/Texture

• Simulation Frags– Euler integration for

Newtonian motion– Fields

• Gravity/Uniform• Drag• radial• vortex

– Collisions• Sphere• plane

InitFragmentOmniinitcode << GetRandomNumGenerator() <<

‘’ void InitParticle( int index, inout Part particle ){‘’‘’ float4 rnd = Random( index );’’ ’’ particle.pos = float3( 0, 0, 0);’’

‘’ float3 buildVel = 2 * ( rnd.x – 0.5 );““ float fac = sqrt ( 1 - buildVel.z* buildVel.z );”<< SinCos( ‘’theta’’, ‘’2* PI* rnd.y’’ )<<“ buildVel.x = fac* cosTheta;"“ buildVel.y = fac* sinTheta;”“ particle. Vel = buildVel;“;

initcode << OffSetMinMaxDistance( m_emitter->GetMinDist(), m_emitter->GetMaxDist() )<<

“ particle.vel *= ( 2* rnd.w - 1) *”<< m_emitter->GetSpeedRandom() <<‘’ +’’ << m_emitter->GetSpeed() <<‘’;’’<< Epilogue( 0, pdesc );

Simulation Shader

• Very easy to add new types of emitter and dynamics– implement HLSL code in a ISimulationFragment

• Not a huge Über-shader – Only the Code needed

• Slight problem adding a new Feature to an “Active” Asset– Shader rebuild not just a parameter tweak

Render Shader

• Three customizations of our standard shader builder– no Über-shader again– Easy use of current special effects

• Three standard base types– Pointsprites– Line / Streaks– Quad / billboards / sprites

Render Shader

• geometry amplification for non-point sprite primitives– Use velocity to get 2nd vertex for line particles– Produce further vertices for camera facing quads

• Once again only required parts of shader are present

• Line fade / colour change• ramp textures to rotate/scale sprites

Parameters

• Broadest sense– Parameters/Data for effects• Softness• Spread• Emitter transform

– Position, direction

– Textures• Ramps

– scaleX, scaleY, rotation, RGBA

• Curve / texture emitters

DEMONSTRATION

DEMONSTRATION

• Emitters– Directional– Omni-directional– Volume

• Box• Sphere• Cylinder• Cone• torus

– Curve

• Visual Effects– Normal map– Cube Map– Projection Textures

• Fields– Gravity– Radial– Vortex– Drag

• Ramps– Scaling– Rotation– RGBA