Introduction to
Geometry Shaders
Patrick Cozzi
Analytical Graphics, Inc.
Overview
Geometry Shaders in the Pipeline
 Primitive Types
 Applications
 Performance

Birds Eye View
Create or destroy primitives on the GPU
 Requires DirectX 10 or GL_ARB_geometry_shader4

Geometry
Shader
Geometry Shaders in the
Pipeline
Vertices in
world coordinates
Vertex
Shader
Perspective Divide
Fragment
and
Shader
Viewport Transformation
clip coordinates
Fragment
Shader
window coordinates
Geometry Shaders in the
Pipeline
window coordinates
clip coordinates
Vertex
Shader
clip coordinates
Primitive
Geometry
Assembly
Shader
PD
Geometry
and
Shader
VT
PD
PD
Fragment
Fragment
and
and
Clipping
Shader
Shader
VT
VT
Fragment
Shader
Primitive Types
Geometry
Shader
Output primitives can be disconnected
Primitive Types

Input Primitives




GL_POINTS
GL_LINES
GL_TRIANGLES
Adjacency

Output Primitives



GL_POINTS
GL_LINE_STRIP
GL_TRIANGLE_STRIP
Primitive Types

Input primitive type doesn’t have to
equal output primitive type
blogs.agi.com/insight3d/index.php/2008/10/23/geometry-shader-for-debugging-normals/
Applications

Implement glPolygonMode


Triangles
Points or Line Strips
Emulate GL_ARB_point_sprite

Points
Triangle Strips
Applications
Displacement Mapping
 Single pass generate a cube map
 Extrusions

Shadow volumes
 Fins along silhouettes for fur rendering

Applications: Fur in Lost Planet

Render surface, write buffers for
Fur Color
 Angle
 Length


GS turns each pixel into a translucent
polyline

Automatic LOD
Applications: Fur in Lost Planet
color
angle
length
Images from meshula.net/wordpress/?p=124
Performance

Duplicates per-vertex operations for
vertices shared by primitives
5 vertices processed
9 vertices processed
Vertex
Shader
Geometry
Shader
Performance

Must guarantee order in == order out
Geometry
Shader
Performance

Order guarantee affects parallelism
Geometry
Shader
Geometry
Shader
Geometry
Shader
Reorder Buffer
Clipping
Performance

Buffer size needs to support a number
of threads running in parallel
Performance

Maximum number of vertices a GS will
output:
GEOMETRY_VERTICES_OUT_ARB
Determines the speed of GS execution
 Make this and vertex sizes as small as
possible

Performance

GeForce 8, 9, and GTX2xx

Output size = vertex size *
GEOMETRY_VERTICES_OUT_ARB
Maximum output size: 1,024 scalars
 Performance is inversely proportional
to output size
 Not a continuous function:

• 1-20 scalars: Peak Performance
• 27-40 scalars: 50% Performance
• On GeForce 8800 GTX
Performance

Benefits

Reduces vertex buffer memory usage
• Compute in GS, e.g. normals
• Create more geometry
• No need to duplicate (e.g. compared to
equivalent VS implementation)
Less memory == less bus traffic
 Reduces vertex attribute setup cost

Summary
Modify incoming primitive or make a
limited number of copies
 Not for

Large scale amplification
 Instancing

Resources
Introduction to Direct3D 10
SIGGRAPH 2007 Course Notes
www.microsoft.com/downloads/details.aspx?FamilyId=96CD28D5-4C15-475E-A2DC-1D37F67FA6CD&displaylang=en
Resources
GL_ARB_geometry_shader4
www.opengl.org/registry/specs/ARB/geometry_shader4.txt
Resources
Section 3.5
www.realtimerendering.com
Resources
Section 4.6
developer.nvidia.com/object/gpu_programming_guide.html