Gladiator Deferred Shading Demo

1
Gladiator Deferred Shading Demo
2
Raw Stats

• 150k tris/frame at 30Hz
• 25 skinned characters (64 bones each), 5k-15k
  tris each, tangent space bump mapped
• Single scene pass deferred shading, four material
  channels albedo, detail albedo, gloss, normal
• Up to 24 real-time omnidirectional lights
• Unified per-pixel lighting model two-hemisphere
  area, directional, and omnidirectional (point)
  lights
• Fully dynamic shadow volumes, CPU silhouette
  extraction from low poly (400-1000 tri) meshes
• Near-linear light accumulation (gamma 1.3), 2x
  overbright lighting, hardware gamma ramp converts
  from gamma 1.3 to sRGB (http//www.srgb.com)

3
Deferred Shading Observations

• A significant fraction of the total scene time
  can be spent just rendering attributes.
• Reducing the number of scene passes frees up lots
  of CPU/GPU time for other stuff shadows, lights,
  etc.
• Why not the holy grail ONE scene pass?
• For a single pass scheme to work on Xbox, all
  attributes must be all packed into a single
  32-bit ARGB color. Were dont have the luxury of
  MRTs (Multiple Render Targets).

4
Deferred Shading Observations

• What are the minimum attributes needed for decent
  per-pixel lighting with bump mapping?
• Depth, Albedo, Normal, Gloss
• For a workable solution the attributes must be
  packed after texture filtering. Without this
  requirement the problem is much easier point
  sample prepacked textures.
• At first glance this doesnt seem possible How
  to pack all attributes needed for decent
  per-pixel lighting into one 32-bit ARGB color
  using the Xboxs texaddrops (Texture Address
  Operations) and combiners?

5
Identifying Limitations and Whats Really
Important

• In order to pack all attributes so tightly its
  necessary to identify the perceptual importance
  of each attribute.
• The pack/unpack process described here was
  heavily influenced by the Xboxs GPU
  architecture, NTSC (Never the Same Color) display
  output, and the genre of our demo.

6
Looking at the Problem Backwards

• Unpacking
• Assume the packing problem has been solved How
  to quickly unpack the attribute data during the
  screenspace lighting passes into a form suitable
  for lighting in the combiners?
• Xbox supports 2D dependent texture lookups using
  AR or GB as the source texture coordinates. So it
  seems natural to pack albedo/gloss into one pair
  of components (AR) and the normal into the other
  set (GB).
• Unpacking in the light shaders will be easy, and
  not require any combiners.
• Requires two precomputed lookup tables (256x256
  2D texture maps).

7
Know Your Attributes

• Albedo Packing
• Good albedo textures shouldnt have much if any
  burned-in lighting
• With per-pixel lighting their influence on the
  final look doesnt justify a full 24-bit color in
  many cases 16-bits or even less can look OK
• Detail Texturing
• Shrek Xbox made heavy use of 2-3 layer detail
  albedo texturing (modulation and compositing), so
  it seems prudent to require some type of detail
  map support.
• The packing process must occur after the texture
  lookups, ruling out texaddrop (lookup table)
  assisted packing.

8
Know Your Attributes

• Gloss Packing
• Some type of per-pixel or per-material gloss
  (specular intensity) control is necessary.
• The grayscale gloss maps used in Shrek where
  usually plain, so scrunching gloss down to a few
  bits (2-4) shouldnt be a show stopper.
• Normal Packing
• Output is constrained to a unit length vector.
  Its possible to discard the magnitude of one
  component and recover it later
• z zSign sqrt(1 x x y y)
• Sign of discarded component can be stored as a
  single bit.
• Z Recovery op can be done in the 2D dependent
  lookup used for normal unpacking.

9
Attribute Packing Pixel Shader

• Samples albedo and tangent space normal maps
• Cubemap renormalizes iterated geometric normal
• Cant renorm. the iterated tangent and binormals.
  They arent as perceptually important anyway.
• Rotates normal from tangent to view space in the
  combiners
• Approximately renormalizes rotated normal
• Equation n n .5 (1 - dot(n, n))
• Packs normal into Output.GB
• Packs albedo and gloss into Output.AR
• Note Shader detailed here only supports 1
  albedo map. Detail texturing is an exercise for
  the reader.

10
Attribute Packing Pixel Shader Final Bit Layout
-1 of 2

• Output AR BRGL 4462
• Output GB ZYX 178
• Albedo can be encoded as sRGB or YCbCr
• 4 bits for Albedo.B/Cb and Albedo.R/Cr
• 6 bits for Albedo.G/Y
• 2 bits for Gloss (L)
• Normals are packed in view space
• 1 bit for Normal.Z (sign)
• 7 bits for Normal.Y
• 8 bits for Normal.X

11
Attribute Packing Pixel Shader Final Bit Layout
2 of 2

• Output Color AAAA AAAA RRRR RRRR GGGG
  GGGG BBBB BBBB
• Packed Bits BBBB RRRR GGGG GGLL ZYYY
  YYYY XXXX XXXX

Key to Packed Bits B Blue/Cb (4 bits) R
Red/Cr (4 bits) G Green/Y (6 bits) L Gloss (2
bits) X Normal.X (8 bits) Y Normal.Y(7
bits) Z Normal.Z sign (1 bit)
12
Load Time Texture Map MungingSimplifies
Packing Pixel Shader

• The color components of the albedo/normal
  texture palettes are rearranged, or duplicated to
  the alpha channel.
• Albedo Texture Components
• albedoTexture.A Blue or Cb
• albedoTexture.R Green or Y
• albedoTexture.G Gloss (specular intensity)
• albedoTexture.B Red or Cr
• Normal Texture Components
• Alpha channel of tangent space normal maps
  contains Y
• normalTexture.A Y
• normalTexture.R X
• normalTexture.G Y (ignored by packing shader)
• normalTexture.B Z

13
Packing Pixel Shader Inputs Are Swizzled
Simplifies Packing Pixel Shader

• T1, V0, and V1 are the rows of a 3x3 tangent
  space to view space rotation matrix
• The components of each row are inverted in the
  vertex shader to simplify packing
• T1 View Space Normal.zyx
• V0 View Space Tangent.zyx
• V1 View Space Binormal.zyx

14
Shader Flowchart

• John Brooks (CEO, Blue Shift, Inc.) made me
  delete it!