GPU for Game Computing - Home - AMDdeveloper.amd.com/wordpress/media/2012/10/Bleiweiss-GameComp… · Multi core CPU, PPU, GPU GPU assisted physics Game play, effects Physics Rendering

GPU for Game Computing

Avi Bleiweiss

GPU transitionCISC to RISC

Dynamic architectureCompute, bandwidth

Data parallelOften algorithm redesign

Game computingSimulation, AI, Audio

GPU, processor array

Compute abstraction

Physics simulation

Mesh video mapping

Results, summary

FYSI, physics simulation case study

GPU, processor array

Compute abstraction

Physics simulation

Mesh video mapping

Results, summary

Deeply programmableVertex, geometry, pixel

Precise and flexible32 bit IEEE float

Unified shader model

Compute pipelineSimplified modality, API

Vertex Shader

Geometry Shader

Pixel Shader

Rasterizer

vertices

vertices

primitives

pixels

pixels

Video Memory

Raw compute power48 pixel engines, eachExecutes 4 float mad per cycle@650 MHz -> 250 GFlops

Shader model 3.0Dynamic branching

Shader model 4.0Integer arithmetic

Memory256 bits wide, double edge@775 MHz -> 50 GBytes/sec0.5 GByte video memory

I/O, PCI Express x164 GBytes/secMulti-GPU shared resources

Relatively small caches

Grid formationALU heavy architecture

Stream data memory system

Many in flight threadsHide memory latency

No static data

No read-modify-write buffers

ScatterIndirect memory write (data[i] = x)Populate data structures Usually performed on CPU

GatherIndirect memory read (x = data[i])Access of data structuresMaps onto texture fetch

Input, multiple arraysDifferent resolutions

Grid computationNo inter-cell dependencies

KernelsApplied to each grid elementHigh arithmetic intensity

Output, multiple arraysMust have same resolution

Porting to GPU non-trivialSignificant speedup payoff

Efficient data structuresIn video memory, sliver

GPU for computing cheapNone of display, texture filtering

Multi-GPU Load partition, shared resources

0 1 2

3 4 5

6

3x3 Grid

7 active elements

2 sliver elements

No double precision floatSingle float good enough

No pixel scatterSelf-modifying location expensive

Dedicated branch unitsBoth sides executed

Grid outputs limited4/8 in DirectX 9/10, respectively

GPU, processor array

Compute abstraction

Physics simulation

Mesh video mapping

Results, summary

Complex, non-intuitivefor general computation

Quest for RISC APIReduced interface set core

Full screen quad

RISC API for renderingRay tracing

Hide graphics APIUnderlying DirectX 9 & 10

Evolving hardware seamless

ScalableMulti GPU support

Automatic multi passingOvercome GPU limitations

Opaque device pointer(s)Single, multi-GPU

Interface objects

Object actionsCreate, assign, remove

ShaderKernel

Target (2D/3D)Scratch

Texture (2D/3D)Resource

GPU ObjectAbstract Object

Data typesScalars, vectors, matrices

Load, storeResource, scratch, respectively

Kernel parametersOne time compile overhead

Compute invocationAcross grid cells

Debug, scratch dump

A single formatFor both resource, scratch

Four float components

Three component vectorAlpha channel for control

Three/four squared matrixThree dimensional array

X Y Z C

00 01 02

10 11 12

20 21 22

vector

matrix

slice #0

slice #1

slice #2

IterativeRecirculate scratches

Parallel setup pathsResources, kernels

Implicit correlationScratches, kernel

Multi pass computeSingle kernel execution

Results CPU access

Assign

Unbind

Bind

Compute

Remove

scratchesresources kernels

results

Create

Remove

Create

Grid parallel traversalTiled pattern, ownership

Point sampled grid cellsVPOS for cell index

Relative resource address Multi resolution resources

Dependent data accessMulti pass reduction

0 1 2

3 4 5

6 70 1

3 4

0 1 2 3

resources

0 1

2 3

scratches

2

5

6 7

0 1

3 4

2

5

6 70 1 2 3

0 1 2 3

3D (4x1x3)

2D (2x2)

2D (3x3)

2D (3x3)

2D (3x3)

Compute bound process

Asymmetric processing

Computation distributionSimple grid subdivision

Synchronization overheadPCI Express contention

Shared resources

GPU, processor array

Compute abstraction

Physics simulation

Mesh video mapping

Results, summary

Game engineSimulation, visual rendering

Physics scope Rigid, deformable bodies

Physics platformsMulti core CPU, PPU, GPU

GPU assisted physicsGame play, effects

Physics

Rendering

scene description

display

Physics scene description formatFYSL, GPU oriented

Consistent I/O abstractionSimulation input and results

Stream APIAsynchronous, discrete simulation

GPU, CPU implementationPerformance analysis

Iterative physics pipelineSystem Setup, Solver, Collision

Physics to rendering interfacePosition, transform update

Compute Abstraction LayerOn top of DirectX API

Simulation controlType, time step

Actors, hierarchicalBounding volumes, meshesLinear, angular motion

Joints, motion constraintsSpring, distance

FeedbackPath decision

Actor definition Joints definition

System Setup

Solver

Collision

results (FYSL)

initial scene (FYSL)

current state resolution

per stepnext state resolution

detection and response

Grid based simulationEuler method (x(t + dt) = x(t) + t * dx/dt;)

Geometry, property resourcesTexture array

Portable shading library

Resume from last stepResult caching

Adaptive time step

GPU, processor array

Compute abstraction

Physics simulation

Mesh video mapping

Results, summary

Object space representationPrecise, CPU equivalent

Geometry in video memoryMultiple arbitrary meshes

GPU texture addressability4K by 4K for 2D

Mesh representation1D vertex and index buffers

Vertex and index buffersPositions and faces, respectively

Counter clockwise triangles

vertex buffer

position # 0

position # 1

position # 2

position # 3

position # 4

face # 0 (0, 1, 2)

index buffer

0 1

2 3

4

5

6

position # 5

position # 6

triangular mesh

face # 1 (2, 1, 3)

face # 2 (1, 4, 3)

face # 3 (3, 4, 5)

face # 4 (5, 4, 6)

Mesh as a resource2D video memory array

Three floats element{x,y,z} and {i0,i1,i2}

Arbitrary dimensionsPadded as necessary

Dependent textureFor each triangle access

0 1 2

3 4 5

6

vertex buffer (3x3)

0 1 2

3 4

index buffer (3x2)

Many mesh scene

16 textures constraint

A single super meshCoalesces multiple meshesVertex and index uber buffer

Index buffer offset

Boundary sub mesh id

b # 0 b # 1 b # n

b # 0 b # 1

b # 1

b # n

b # n

1D composite

2D composite

Low creation overhead

Sub mesh collisionTri-tri intersection, contact

Avoids self intersectionMulti pass compute

Intersect, response, integrate

Dependent texture accessFace to vertex fetch

GPU, processor array

Compute abstraction

Physics simulation

Mesh video mapping

Results, summary

GPU superior to CPUCollision detection, response

Understand compute behaviorGrid distributionFlow control, multi pass

Scalability across GPUsMore pipes, ALUs

Single CPU No dual core, no SSE2

System performancePhysics process, no rendering

Shape based benchmarksLinear motion

Absolute, normalized results

8655802 24 GeForce 7800

800650224GeForce 7900

775650316Radeon X1900

750625116Radeon X1800

70060034Radeon X1600

53334001(4)1Pentium 4

Memory (MHz)Core (MHz)ALUsPipesProcessor

7.5775367mesh update

30.138241300mesh-mesh

102

119

Ops

17.40587volume-volume

19

Texture

72

ALU

3.79impact

RatioShader Type

562768(0.0002)

197508(0.0001)

74793(0.0004)

101404(0.0004)

CPU

703(0.177)

187(0.171)

188(0.164)

203(0.231)

7800

482(0.259)

47(2.659)

110(1.136)

125(1.000)

tetrahedron128x128

247(0.129)

30(1.066)

16(2.000)

32(1.000)

mesh64x64(62471 tris)

201(0.154)

15(2.067)

15(2.067)

31(1.000)

sphere256x256

220(0.213)

15(3.133)

31(1.516)

47(1.000)

aabb256x256

7900X1900X1800X1600Benchmark

Compute draw calls (figures in msec)

CPU wins for small gridsExpected, setup overhead

GPU faster elsewhereUp to an order of magnitude

GPU ScalabilityFlow control might be stalling

GeForce 7800/7900 slower Across benchmarks

Unified shader architecture

Higher concurrency level64 pixel enginesFlexible mix of scalar/vector

DirectX 10Constant buffersTexture array, indexingNon-power-of-2 3D textures3D render target

GPU game computing Still a challenge

Parallel programmingMacro and micro level

Software productivityTools, tools, tools

Emerging CPU/GPU platformsAdaptive load balance

Thank You!

Questions?