AnyGL: A Large Scale Hybrid Distributed Graphics System

1
AnyGL A Large Scale Hybrid Distributed Graphics
System

• Ph.D Candidate Yang Jian(??)
• Supervisor Jiaoying Shi(???)
• State Key Lab of CADCG, Zhejiang Univ.
• 2002

2
Contents

• Goal of Computer Graphics Architecture
• Parallel Graphics Architecture
• AnyGL in Details
• Conclusion
• Further Research

3
Goal of Graphics Architecture

• Real-time
• Triangle rendering capacity
• Rasterization rate
• Texture size
• Display Resolution
• Realistic
• Local illumination, Gouraud, Phong
• Global illumination, ray-trace, ray-casting,
  Radiosity

4
Goal of Graphics Architecture

• Programmable,
• more effects
• faster,
• realistic
• non-realistic

5
Contents

• Goal of Computer Graphics Architecture
• Parallel Graphics Architecture
• AnyGL in Details
• Conclusion
• Further Research

6
Parallel Graphics Architecture

• Development of Graphcis Architecure
• Classification of Parallel Graphics Architecure
• Hardware System
• Software System
• Scalability of Parallel Parallel Graphics
  Architecure

7
Development of Graphics Architecture

• 1rd Generantion, 1975 1985,
• Geometry transfromation, clip, render objects
  based on wire frame
• 2rd Generation, mid 1985 1990
• Video memory, depth buffer, back face cull,
• gouraud rendering
• 3rd Generation, 1990 2000s
• Texture Mapping
• anti-alias, mutlis-sample

8
Development of Graphics Architecture

• 4th Generation,1997
• Local illumination
• Hardware lighting, complex texture mapping(bump ,
  compressed, 2000
• SIMD, 2001
• Programmable processor, 2001
• More programmable processors, 2002
• 5th Generations, 2005?
• Global illumination

9
Parallel Graphics Architecture

• Development of Graphcis Architecure
• Sorting Classification of Parallel Graphics
  Architecure
• Hardware System
• Software System
• Scalability of Parallel Parallel Graphics
  Architecure
• Misc.

10
Sorting Classification

• S. Molnar et al. 1994
• sort-first
• sort-middle
• sort-last
• M. Eldridge 2000, Pomegranate
• Transform Lighting
• Rasterization
• Texture
• Fragment
• Display

11
Typical Graphics Pipeline
12
sort-first
13
sort-middle
14
sort-last
15
Parallel Graphics Architecture

• Development of Graphcis Architecure
• Sorting Classification of Parallel Graphics
  Architecure
• Hardware System
• Software System
• Scalability of Parallel Parallel Graphics
  Architecure

16
Hardware System of PGA

• SGI Reality Engine
• sort-middle
• share bus, broadcast
• every 2 scanlines one processor

17
SGI Reality Engine
18
Hardware System of PGA

• PixelFlow
• Sort-last
• ring-net, compose image
• SIMD is highly expolited
• Pomegranate
• 5 stages
• 64 nodes

19
Pomegranate
20
Hardware System of PGA

• Evans Sutherland Freedom 3000
• sort-last
• Kubota Denali
• sort-last
• Sepia, Compaq Reserch 2000
• sort-last
• ServerNet-II

21
Parallel Graphics Architecture

• Development of Graphcis Architecure
• Sorting Classification of Parallel Graphics
  Architecure
• Hardware System
• Software System
• Scalability of Parallel Parallel Graphics
  Architecure
• Misc.

22
Sortware System

• GLR
• Client/Server
• GLX, X Windows
• Princeton Multi-projector System
• sort-first
• one node distriutes commands
• 8 nodes receive and render commands
• Parallel Mesa
• sort-last
• binary swapping, parallel pipeline compositing

23
Sortware System

• WireGL
• sort-first
• state tracking and state switch, dirty-bits,
  Lazy-update
• 3 kinds of OpenGL commands
• 16 clients 16 servers
• Lightning-2 assembles images

24
WireGL(Chromium)
25
Parallel Graphics Architecture

• Development of Graphcis Architecure
• Sorting Classification of Parallel Graphics
  Architecure
• Hardware System
• Software System
• Scalability of Parallel Parallel Graphics
  Architecure
• Misc.

26
Scalability of Parallel Parallel Graphics
Architecure

• 5 parameters, Eldridge et al.2000
• triangle input rate
• triangle rendering rate
• texture size
• pixel rate
• display resolution

27
Contents

• Goal of Computer Graphics Architecture
• Parallel Graphics Architecture
• AnyGL in Details
• Conclusion
• Further Research

28
AnyGL

• Goal
• Architecture
• Classification of OpenGL Commands
• State Tracking
• Compression
• Scalability
• Parallel Graphics API

29
Goal of AnyGL

• Hibrid Distributed Graphics System
• sort-first
• sort-last
• Large scale
• sort-first, new state tracking
• sort-last, image compression
• Good Scalability

30
AnyGL

• Goal
• Architecture
• Command classification
• Analysis of Dependency Relation
• State Tracking
• Compression
• Scalability

31
Architecture of AnyGL

• 4 kinds of Nodes
• Geometry Distributing Node ( G-node )
• Goemetry Rendering None ( R-node )
• Image Composite Node ( C-node )
• Display Node ( D-node )
• Sort-first
• G-node, R-node
• Sort-Last
• C-node, D-node

32
Architecture of AnyGL
33
(No Transcript)
34
Geometry Rendering Node

• Capture opengl commands of application
• Packet OpenGL commands
• Computer State changes
• Send OpenGL command packets

35
Geometry Rendering Node

• Receive Packet
• Check if context switch is needed
• Context switch
• Unpack the command packet
• Send corlor depth buffer to C-node

36
Image Composite Node

• Receive color and depth data from R-nodes
• If they are compressed, decompress them.
• Compose the received data
• Send composed color and depth buffer to D-nodes

37
Display Node

• Receive color and depth data from C-node
• Assemble image for final display
• Display the final image by CRT, Projector, LCD...

38
Logical Division of Command Buffer

• Command buffer is divided into 2 parts
• Command Code buffer
• Parameter Buffer
• Command Code buffer
• ptr--
• Parameter Buffer
• ptr

39
(No Transcript)
40
Command Packet Structure
41
Subdivision of glBegin-glEnd

• Condition
• lots of commands appear between a pair of
  glBegin-glEnd
• The command buffer or parameter is full, while
  glBegin does not meet the corresponding glEnd
• Requirement
• Make sure that glBegin and glEnd are matched in
  each packet
• Maitain completeness of primitive

42
Subdivision of glBegin-glEnd

• ??2.1 glBegin????
• ?glBegin?glEnd?????????glBeginEnd,????1
• ?glVertex??????vtxCount,?????0
• ?nPrimtiveType???????
• glBegin?????????????
• glBegin??(????)??????????
• glBeginEnd 0
• nPrimitiveType glBegin????

43
Subdivision of glBegin-glEnd

• ??2.2 glVertex????
• ?glVertex???????????
• ????????????
• if(glBeginEnd1)vtxCount,??????????
• ??2.3 glEnd??????
• ?glEnd???????????
• glBeginEnd1
• vtxVount0

44
Subdivision of glBegin-glEnd

• ??2.4????????OpenGL???????
• if (glBeginEnd) then ??,?????????
• if(vtxCount 0 )??
• ????????????vtxBackdate
• ???????GL_TRIANGLES_FAN??GL_POLYGON,???????????,bR
  ecoverFirst1
• ?glEnd???????????
• ?????????
• ??glBegin??
• if(bRecoverFirst 1)
• ??????????glBegin????
• ??????????glVertex?????????
• ??vBackdate?glVertex????
• ??????????OpenGL????????????????????????

45
AnyGL

• Goal
• Architecture
• Command classification
• Analysis of Dependency Relation
• State Tracking
• Compression
• Scalability

46
Command Classification

• OpenGL 1.3
• 17 kinds
• Four kinds in AnyGL
• primitive commands
• state commands
• remote remapping commands
• speical commands

47
Remote Mapping Commands

• G-nodes send textures to R-nodes
• R-nodes receive textures from G-nodes
• For R-nodes
• textures with same texture ID may reresent
  different textures, since they come from
  different textures
• textures with different texture IDs maybe same
  texture.

48
AnyGL

• Goal
• Architecture
• Command classification
• Analysis of Dependency Relation
• State Tracking
• Compression
• Scalability

49
Analysis of Dependency Relation

• OpenGL is client/server based state machine, its
  commands have serious order
• See OpenGL 1.1 state machine

50
Analysis of Dependency Relation

• Graphics Pipeline
• Transformation
• Lighting
• Rasterization
• Fragment
• According to OpenGL 1.3
• 17 categories

51
AnyGL

• Goal
• Architecture
• Command classification
• Analysis of Dependency Relation
• State Tracking
• Compression
• Scalability

52
State Tacking of AnyGL

• History of Sate Tracking
• Sate Tracking Based On Logic Timestamp
• Test Result

53
History of State Tracking

• GLS, no
• GLR, pixel format
• GLX, X Windows, all 2d, pixelformat vertex array
• Multi-projector by Princeton, no
• Parallel Mesa, no
• Argus, notice the problem

54
History of State Tracking

• M. Cox, parallel renderman
• WireGL, Dirty-bits, Lazy-update
• Chromium, user-defined

55
State Tacking of AnyGL

• History of Sate Tracking
• Sate Tracking Based On Logic Timestamp
• Test Result

56
State Tacking

• GLS
• GLR, Pixel Format
• GLX X Windows(R11) , Vertex Array
• M.Cox, RenderMan, Approximate
• WireGL
• Dirty-bits and Lazy-update

57
State Tacking Based on Logic Timestamp

• G-node
• Maintain n1 virtual context,
• n is the number of R-nodes connected to this
  G-node
• The extra virtual context is called application
  virtual context, same as the context of current
  process
• R-node
• Maintain m virtual context
• m is the number of G-nodes connected to this
  R-node

58
State Tacking Based on Logic Timestamp

• Application Virtual Context
• Always same as the context of the current process
• Active Virtual Context (G-node)
• If and only if it will be the command packages
  destination
• Active Virtual Context(R-node)
• If and only if it represents the source node of
  the current command packet

59
State Tacking Based on Logic Timestamp

• 5 Functions
• State tracking of G-node
• Context Difference
• Context Synchronization
• Context Switch
• State tracking of R-node

60
(No Transcript)
61
State Tacking on G-node
glLightf_state(light, pname, param) check
validate of parameters check change of the
light state if (changed) lightslight.pnamex
param lightslight.pnamex.timestamp lights
light.timestamp
62
Context Difference
glLight_Diff(src, dest) if timestamp of src,
dest are equal return if (lighting.timestamp of
src,dest not equal dest.lighting.enable) pack
glEnable(GL_LGHTING) else pack
glDisable(GL_LGHTING) if( dst.lighting) for(I
0 I Ilt num_lights, I) if(lightI.timestamp of
src,dest not equal) pack changed parameters of
dest into geometry packet call difference
of material
63
Context Synchronization
glFragment_synch(src, dest) if(
fragement.timestamp of src, dest are equal)
return if(src.fog.timestamp ! dest.
fog.timestamp) src.fog dest.fog src.fog.tim
estamp dest.fog.timestamp call
sychronization of depth
64
Context Switch
glTransform_switch(src, dest)
if(transform.timestamp of src ,dest are
equal) return if(modelview.timestamp of src,
dest not equal) //call hardware
api glLoadMartix(dest_model_view) src.modelview
dest.modelview src. modelview.timestamp
dest. modelview.timestamp call
projection switch of src , dest
65
State Tracking on R-node

• Similar to state tracking on R-node
• It call hardware API directly

66
State Tacking of AnyGL

• History of Sate Tracking
• Sate Tracking Based On Logic Timestamp
• Test Result

67
State Tracking Performance

• It is almost as fast as WireGL
• Context Difference is fast
• Context Switch is fast
• Context synchronization
• lt 1 microsecond

68
(No Transcript)
69
(No Transcript)
70
AnyGL

• Goal
• Architecture
• Command classification
• Analysis of Dependency Relation
• State Tracking
• Compression
• Scalability

71
Compression

• Image Compression
• Geometry Compression
• Normal
• Position
• Color

72
Image Compression

• VCM(Video Compression Manager)
• Default Color Image Compression
• HUFFYUV
• Default Grayscale Image Compression
• JPEG-LS

73
Geometry Compression

• Deering95
• Normal
• Position
• Color

74
Geometry Compression

• Normal
• Deer95 software implementation
• 18-bits
• Programmer can choose compression precision

75
Geometry Compression

• Position
• Linear Prediction,
• GL_POINT, GL_LINES, GL_LINESTRIP,
  GL_TRIANGLES,GL_QUADS
• Circle prediction,
• GL_POLYGON, GL_LINE_LOOP
• Parallelogram prediction,
• GL_TRIANGLE_STRIP, GL_QUAD_STRIP
• Fan Prediction
• GL_TRIANGLE_FAN

76
Circle Prediction
Parallelogram Prediction
77
Triangle Fan Prediction
78
Geometry Compression

• Linear prediction

• Circle prediction

79
Geometry Compression

• Parallelogram prediction

Triangle fan prediction
80
AnyGL

• Goal
• Architecture
• Command classification
• Analysis of Dependency Relation
• State Tracking
• Compression
• Scalability
• Parallel Graphics API

81
Scalability

• Scalability of Parallel System
• Component
• System Size
• Problem Size

82
Scalability

• 5 Parameters in Pomegranate,
• Triangle input rate
• Triangle Rendering Rate
• Pixel Rate
• Texture Size
• Display Resolution

83
Scalability of Nodes

• G-node
• Pack capability, 20M triangles /s
• R-node
• Render capability
• C-node
• Composite capability
• D-node
• Display Resolution

84
Scalability of Nodes

• No Node Number Limit
• Main bottleneck
• Network Bandwidth
• Compression is Adopted to solve the bottle- neck
• Geometry Compression
• Image Compression

85
AnyGL

• Goal
• Architecture
• Command classification
• Analysis of Dependency Relation
• State Tracking
• Compression
• Scalability
• Parallel Graphics API

86
Parallel Graphics API

• Traditional Parallel Programming
• Parallel Graphics API in Argus
• Global Share Extension of AnyGL
• Parallel Programming Frame of AnyGL

87
Traditional Parallel Programming

• Single Process, Vector Machine, HPF and Fortran
  90
• DSM
• Message Passing, PVM, MPI,
• Multi-Process, global variables, SGI Power C,
  p-tread

88
Parallel Graphics API in Argus

• Barrier
• Distributed Mutex
• Typical Application
• Volume Rendering

89
Global Share Extension

• Assume
• m G-nodes have same texture tex0
• Tex0 occupies K MB memory
• Each R-node allocate M K MB memory to store
  Tex0!

90
Global Share Extension

• GL_TEXTURE_nD, n 1, 2, 3
• Local Texture
• R-node allocates memory for each local texture
• GL_GLOBAL_TEXTURE_nD, n 1,2,3
• Global Texture
• R-node allocates shared memory for global texture

91
Global Share Extension

• GL_GLOBAL_TEXTURE_nD, n 1,2,3

92
Parallel Frame of AnyGL

• int main(int argc, char argv)
• glboals.bWorking 1
• globals.M_PolygonMode GL_FILL
• globals.S_PolygonMode GL_FILL
• MPI_Init(argc, argv) /MPI???/
• MPI_Comm_rank( MPI_COMM_WORLD, rank_id)
• MPI_Comm_size( MPI_COMM_WORLD, nprocs )
• if(rank_id 0) /?????/
• MasterGUIInit(argc, argv) /??GUI??/
• Create_Context() /??GLcontext/
• MasterMainLoop(Display)
• else /?????/
• Create_Context()
• for() Displau()
• MPI_Finalize( ) / ??????/
• return 0

93
Contents

• Goal of Computer Graphics Architecture
• Parallel Graphics Architecture
• AnyGL in Details
• Conclusion
• Further Research

94
Conclusion

• 1.AnyGL is a larger-scale hybrid distributed
  graphics system, which has integrated sort-first
  and sort-last parallel graphics architecture. It
  provides unlimited scalability for distributed
  graphics system.
• 2.Its virtual parallel graphics pipeline is
  composed of 4 kinds of logic nodes, G-Node,
  R-Node, C-Node, D-Node

95

• 3. A new kind of dependency relation, graphics
  context dependency relation of OpenGL pipeline,
  is analyzed for distributed graphics architecture
  in details.
• 4. A new state-tracking algorithm is implemented
  in AnyGL, which is called Logic Timestamp
  Graphics State Tracking.
• 5. Image compression and geometry compression are
  deeply exploited in AnyGL.

96

• 6. A new parallel programming interface, global
  share of texture object and display list, are
  proposed for future parallel graphics
  architecture research.
• AnyGL provides a new solution for larger scale
  distributed graphics system in theory.
  Performance test proves the high scalability of
  AnyGL. It does not limit the nodes number of
  logic node. Graphics state tracking algorithm
  based on Logic timestamp introduces almost no
  extra overheads. It research is the basis of LAN
  larger scale distributed graphics system.

97
Further Research

• D3D
• COM ? DCOM
• Parallel of Stream Based Graphics Architecture?
• Global Illumination for Parallel Architecture
• Programmable Shading Language Support for
  Distributed Graphics System?

98
Thanks for Your Attention!