Advanced%20D3D10%20Rendering

1
Advanced D3D10 Rendering

• Emil Persson
• May 24, 2007

2
Overview

• Introduction to D3D10
• Rendering techniques in D3D10
• Optimizations

3
Introduction

• Best D3D revision yet! ?
• Clean and powerful API
• Lots of new features
• SM 4.0
• New geometry shader
• Stream Out
• Texture arrays
• Render to volume texture
• MSAA individual sample access
• Constant buffers
• Sampler state decoupled from texture unit
• Dual-source blending
• Etc

4
Clean API

• Vista only
• Everything is mandatory (almost)
• No legacy hardware support
• Clean starting point for future evolution of the
  API
• Limited market short-term
• Some old features deprecated
• Fixed function
• Assembly shaders
• Alpha test
• Triangle fans
• Point sprites
• Clip planes

5
Dealing with deprecated features

• Fixed function
• Write a few über-shaders
• Assembly shaders
• Convert to HLSL
• Alpha test
• Use discard or clip() in pixel shader
• Use alpha-to-coverage
• Triangle fans
• Seldom used anyway, usually just for a quad
• Convert to triangle list or strip
• Point sprites
• Expand point to 2 triangles in GS
• Clip planes
• Use clip distance and/or cull distance

6
SM 4.0

• Geometry shader
• Processes a full primitive (point, line,
  triangle)
• Has access to adjacency information (optional)
• Useful for silhouette detection, shadow volume
  extrusion etc.
• May output multiple primitives
• Output limitation is 1024 floats
• May output nothing (to kill primitive)

7
SM 4.0

• Infinite instruction count
• Very long shaders may have lower throughput
  though
• Integer and bitwise instruction
• Indexable temporaries
• Allows for local arrays
• May be used to emulate a stack
• Useful system generated values
• SV_VertexID
• SV_PrimitiveID
• SV_InstanceID
• SV_Position (Like VPOS, but now .zw are defined
  too)
• SV_IsFrontFace (Like VFACE)
• SV_RenderTargetArrayIndex
• SV_ViewportArrayIndex
• SV_ClipDistance
• SV_CullDistance

8
SM 4.0

• Integer bitwise instructions
• Signed and unsigned
• No idiv though, just udiv
• Same registers as floats
• Can alias without conversion with asint(),
  asuint(), asfloat() etc.
• Integer texture sample values
• Syntax Texture2D ltuint4gt myTex
• Access to individual samples in MSAA surface
• Allows for custom AA resolve
• Syntax Texture2DMS ltfloat4, 4gt myTex

9
Pixel center

• Half pixel offset is gone! ?
• Affects SV_Position as well
• Now matches OpenGL
• DX10
  DX9

10
Pixel center

• Pixels and texels align
• TexCoord SV_Position.xy / float2(width, height)
• Texel center Screenspace

11
The small batch problem

• D3D10 designed to minimize batch overhead
• Pulls work from draw time to creation time
• Validation
• Shader input/output configuration
• Immutable State Objects
• Input layout
• Rasterizer state
• Sampler state
• Depth stencil state
• Blend state

12
The small batch problem

• D3D10 also provides tools to reduce draw calls
• Improved instancing interface
• Geometry shader
• More shader resources
• Constant indexing in PS
• Render target arrays
• Texture arrays

13
Rendering techniques in D3D10
14
Global Illumination
15
Global Illumination

• Probes on a volume grid across the scene
• Each probe captures light environment into a tiny
  cubemap
• Probes are converted to Spherical Harmonics
  coefficients
• Indirect lighting is computed using interpolated
  SH coefficients
• Do the same in probe passes to get multiple light
  bounces

16
Global Illumination

• Awful lot of work
• Each probe is 6 slices. We need loads of probes.
• Sample scene has over 300 probes
• Solution
• Use geometry shader to reduce work
• Distribute work across multiple frames
• Sample updates 40 cubes per frame
• Scatter updates to hide artifacts
• Skip over empty space probes

17
Global Illumination

• The Geometry Shader advantage
• 40 cubes x 6 faces x n draw calls Pain
• DX9 style unrealistic even for simple scenes
• Update multiple slices per pass with GS
• GS output limit is 1024 floats
• Keep number of interpolators down to maximize
  primitive count
• Managed to update 5 probes (30 slices) per pass
• 8 passes is more manageable than 240 ...

18
Post tone-mapping resolve

• D3D10 allows for custom AA resolves
• Can drastically improve HDR AA quality
• Standard resolve occurs before tone-mapping
• Ideally resolve should be done after tone-mapping
• Standard resolve Custom
  resolve

19
Post-tonemapping resolve

• Texture2DMSltfloat4, SAMPLESgt tHDR
• float4 main(float4 pos SV_Position)
  SV_Target
• int3 coord
• coord.xy (int2) pos.xy
• coord.z 0
• // Tone-map individual samples and sum it
  up
• float4 sum 0
• unroll
• for (int i 0 i lt SAMPLES i)
• float4 c tHDR.Load(coord, i)
• sum.rgb 1.0 exp2(-exposure
  c.rgb)
• // Average
• sum (1.0 / SAMPLES)

20
Optimizations
21
Geometry shader

• GS optimizations
• Input/output usually the bottleneck
• Reduce outputs with frustum and/or backface
  culling
• Keep input small by packing data
• TexCoord could be 2x16 bits in an uint
• Or use for instance asuint(normal.w)
• Merge to full float4 vectors
• Dont do 2x float2
• Keep output small
• Could be faster to trade for some work in PS
• Pass just position, dont interpolate both
  lightVec and viewVec
• Or even back-project SV_Position.xyz to world
  space in PS
• Small output means more work fits within 1024
  floats limit

22
GS frustum and backface culling

• // Transform to clip space
• float4 pos3
• pos0 mul(mvp, In0.pos)
• pos1 mul(mvp, In1.pos)
• pos2 mul(mvp, In2.pos)
• // Use frustum culling to improve performance
• float4 t0 saturate(pos0.xyxy float4(-1,
  -1, 1, 1) - pos0.w)
• float4 t1 saturate(pos1.xyxy float4(-1,
  -1, 1, 1) - pos1.w)
• float4 t2 saturate(pos2.xyxy float4(-1,
  -1, 1, 1) - pos2.w)
• float4 t t0 t1 t2
• branch
• if (!any(t))
• // Use backface culling to improve
  performance
• float2 d0 pos1.xy pos0.w -
  pos0.xy pos1.w
• float2 d1 pos2.xy pos0.w -
  pos0.xy pos2.w

23
Miscellaneous optimizations

• Pre-baked constant buffers
• Dont update per-material constants in DX9 style
• PS dont need to return float4 anymore
• Use float3 if you only care about RGB
• May reduce instruction count
• Use GS to reduce draw calls
• Single pass render-to-cubemap
• Update multiple render targets per pass

24
The new shader compiler

• SM4 shader compiler preserves semantics better
• This means more responsibility for you guys
• Be careful about your assumptions
• Periodically check the resulting assembly
• D3D10DisassembleShader()
• Use GPUShaderAnalyzer for performance critical
  shaders

25
The new shader compiler
Example

• HLSL code
• float4 main(float4 t TEXCOORD0) SV_Target
• if (t.x gt t.y)
• return t.xyzw
• else
• return t.wzyx

DX9 assembly add r0.x, -v0.x, v0.y cmp
oC0, r0.x, v0.wzyx, v0
DX10 assembly lt r0.x, v0.y, v0.x if_nz
r0.x // lt--- Did you really want a branch here?
mov o0.xyzw, v0.xyzw ret else
mov o0.xyzw, v0.wzyx ret endif
26
The new shader compiler

• Use branch, flatten, unroll loop to
  control output code
• This is not for everyone
• Poor use could reduce performance
• Make sure you know what youre doing
• Only use if youre familiar with assembly code
• Verify that you get the code you expect
• Always benchmark both options

New DX10 assembly (using flatten) lt
r0.x, v0.y, v0.x movc o0.xyzw, r0.xxxx,
v0.xyzw, v0.wzyx ret
27
Questions? emil.persson_at_amd.com