DirectX Graphics 9 Overview

1
DirectX Graphics 9 Overview

• PM
• DirectX Graphics Team
• Microsoft Corporation

2
Overview

• Higher Order Surface update
• Displacement Mapping
• Vertex Shader 2.0
• Pixel Shader 2.0
• Infrastructure
• Future plans

3
API Update

• DXG9 is a new .dll
• DX8 interfaces handed to previous dll
• API tokens renamed from 8 to 9
• All object handles are converted to interfaces so
  can be Released()
• TSS split up to handle samplers vs texture
  coord/parameter iterators
• Will clearly delineate Renderstates that still
  apply when using shaders

4
Scissor Plane

• One per-device rect. that can be set to clip
  pixels
• Supported in all HW
• Interesting for productivity apps
• API cap and enable flag
• Devcaps-gtScissor
• D3DRS_SCISSORENABLE

5
Quads

• May return as higher order primitive
• Shared vertex patch scheme, and/or
• Catmull-Clark subdivision surfaces
• Implementations can use 2 tris for planar case
• Would use gl-style color provoking
• Last vert of quad sets flat-shade color
• Not a high priority for DXG9

6
Fog

• Still supported outside shaders
• In FF pipeline controlled by renderstate
• Old apps still rely on this behavior
• Many newer apps will use shader code to implement
  it instead
• May move into shader in DX10

7
Deprecated Features

• W-Buffering
• Less important with high-precision pixel ops
• May be slower than just using z depth
• Color key
• Likely to be cut from textures
• May support some keying in video path

8
HO Primitive Overview

• More support for float tessellation levels
• Allows smoother variation of LOD
• Rational surface support
• Adaptive tessellation
• Displacement mapping
• 2 methods

9
Rational Surfaces

• Improves quality of surface model
• Reduces primitives reqd. in content
• Enabled in most current hardware
• API support
• D3DRS_RATIONALSURFACE
• Devcaps-gtRationalSurface
• Affects N-patches and RT patches

10
Adaptive Tessellation

• Screen-space edge length-based criterion
• Avoids cracks across shared edges
• Supported for both RT-patches and N-patches
• API specifies
• Desired edge length in screen space
• Min/max limits on resulting segment count

11
Displacement Mapping

• Implemented as perturbation to tessellated
  vertices
• Not happening per pixel yet
• Displacement is special register in vertex shader
• Defined by vertex shader declaration
• If present in FF pipeline will be
• Scaled by normal and add to position

12
D-Mapping Methods

• 2 types of capability supported
• 1. Precomputed Displacement Mapping
• Better for objects
• 2. Sampled Displacement Mapping
• Better for terain

13
1. Precomputed D-Mapping

• Assumes all primitives tessellated to same level
• Displacements are pre-computed at points to be
  produced by tessellator
• Ordering may be HW-specific
• Displacements provided as additional
  asynchronous vertex stream
• Displacement available in vertex shader as
  read-only register

14
Precomputed D-Map API

• API routine provided to generate displacement
  values
• Syntax proposals
• Set result vector as vertex shader input stream
  component
• Or possibly as vshader texture

15
Precomputed D-Map Issues

• Advantages
• Precomputation enables fast runtime
• Useful for characters and compact objects
• Limitations
• Wont work with adaptive tessellation
• gt May not be as good for terrain
• Cant animate displacement map
• Even scrolling it requires regeneration
• Procedural mods to d-map likewise

16
2. Sampled D-Mapping

• Filter/sample dmap at render time
• Input surface specified as special texture for
  vertex access
• Evaluator samples using tri-linear
• Makes scalar result available in special register
  for use by vertex shader
• No support for dependent reads
• ie no shader modification of u/v values before
  sampling

17
Sampled D-Map Issues

• Advantages
• Works with adaptive tessellation
• Therefore faster on terrain
• Provides automatic LOD for objects
• Limitations
• Must use care with displacement maps across seams
  to avoid cracking
• Implementations must guarantee order of operations

18
D-Map Usage Scenario A

• Encode normal maps assuming displacements present
• Compute lighting basis from base mesh normals
  only
• Do all lighting per-pixel
• Dmap provides occlusion effect only
• Can be easily skipped in far field
• Requires dmap and nmap data to be synched -cant
  tile independently

19
D-Map Usage Scenario B

• Compute pixel lighting basis per post-tessellated
  polygon
• Use with per-pixel or per-vertex lighting
• Requires displacement map to contain normal
  displacements also

20
D-Map Authoring

• Tool development underway in D3DX
• Take high polycount models
• Convert to low-polycount displacement map
• Many other issues being handled

21
DMAP Data Formats

• 32-bit float pixel format is ideal
• Other formats supported
• 16-bit float
• 16 and 8-bit signed integer
• 2 Channel counts supported
• 1-channel fmt. for displacement only
• 4-channel fmt. for normal displacement

22
Vertex Shader Update

• Minimal changes from DirectX8
• Double instruction and const count
• Both set at 256
• No new math ops
• Add support for flow control

23
Vs.2.0 Model

• 4-D address register a0 w-o 1
• 16 vertex inputs v r-o 1
• 16 temporary registers r r/w 3
• 256 constants c r-o 1
• 256 instructions
• using same constant twice in one instruction is
  ok

24
Vertex Shader Flow Control

• DX9 shaders will support flow control, based on
  constants only
• Designed to solve current issues
• enable/disable envt. mapping, etc.
• varying of lights problem
• Otherwise shaders are regression vs. fixed
  function DX7-style capability

25
Instruction Counts vs. Slots

• Flow control means slots ! counts
• Instruction store is 256, but more instructions
  can be executed than are stored
• Actual instruction count limit is higher

26
Vertex Shader Jump

• Syntax
• jump condition, label
• Where
• condition evaluates to a constant reg, and
• label is an instruction downstream of this one

27
Vertex Shader Loop

• Syntax
• loop condition, label
• Where
• condition evaluates to a constant reg, and
• label is an instruction downstream of this one
• Indexing of constant array bank via
  auto-increment counter

28
Vertex Shader Subroutine

• Syntax
• jsr condition, label
• Where
• condition evaluates to a constant reg, and
• label is an instruction
• Ret statement marks end of routine
• Consumes an instruction slot

29
Vertex Flow Rules

• Jumps can appear inside loops
• Subroutines can be called inside loops
• Jumps and loops can be in subroutines
• Jumps cannot enter or leave loops
• Subroutines cannot be nested

30
Vertex Shader Declaration

• Syntax will be updated to decouple declaration
  from execution
• Declaration will allow redefinition of VB inputs
  without requiring recompilation of shader body
• E.g. Separate handle for declaration
• Will also support more orthogonal feature set

31
Vertex Input Formats

• Orthogonal Types
• 1-, 2-, 3-, and 4-D float
• 2 and 4-D short/word
• 11-11-10 for normals
• BYTE4 0xWWZZYYXX
• D3DCOLOR 0xWWXXYYZZ
• Pad unspecified elements with
• x,y,z,w 0,0,0,1

32
Orthogonal Data Mappings

• Support all the preceding with
• Signed and unsigned
• 2s complement or not
• Normalized and not
• Normalized signed values are -1.0 to 1.0

33
Pixel Shading Overview

• 2 Levels of Shader Programming
• High Level Language
• Assembly Language
• Both are implemented in D3DX
• Just as in DX8.1

34
Assembly Language ps.2.0

• Provides more Direct control
• Exposes details of hw
• Register counts
• Instruction counts
• Will support .cpp macro capability
• Can be embedded in higher-level language as in
  effect files

35
Syntax Note

• Syntax today is not final
• It is intended to represent the logical model
• Please provide feedback on syntax

36
ps.2.0 Model

• 4-D Vector Registers
• 2 color iterators v r-o 1
• 8 texcoord iterators t r/w 2
• 16 textures/samplers s r-o 1
• 16 temporary registers r r/w 3
• 32 constant registers c r-o 2
• Arbitrarily intermixable Instructions
• 32 address ops
• 64 math ops

37
ps.2.0 Sample Shader

• ps.2.0
• dcl v0.rgba // diffuse color
• dcl t0.rg // comes from oT0
• dcl s0.a, 1D // scalar texture
• dcl s1.rgb, 2D
• dcl s2.rgb, 3D, bx2// convert to signed
• dcl s3.rgb, cube
• texld r0, s1, t0 // sample 1-D texture
• mul r1, r0, v0 // modulate diffuse
• out r1 // emit

38
Register Declarations

• Iterators and samplers must be declared before
  use
• Registers v, t, and s
• Temp regs r possible
• Provides opportunity for comment
• Identifies resources allocated for use in shader
• Using less helps performance

39
Declaration Instruction

• dcl regname.components
• Identifies components that this shader will need
  to use
• Must specify components used
• No defaults in declarations
• Identifies 1-D, 2-D, 3-D, vs cube map
• All other state is in TSS

40
Initializer Instruction (def)

• def c, val, val, val, val
• Just an initializer
• Not a declaration
• Can be overridden by subsequent calls to
  SetShaderConstant()

41
Address Instructions

• texld r, s, t/r
• Loads r with value sampled from stage s at
  coordinate t or r
• Using r as sampling address indicates dependent
  read, t is non-dependent

42
Address Instructions2

• texldp r, s, t/r
• Loads r with value sampled from stage s at
  coordinate t or r dividing by w just before
  sampling
• Using r as sampling address indicates dependent
  read, t is non-dependent
• Used for projected texturing

43
Address Instructions3

• texkill t/r
• Works with either t or r register inputs
• Kills pixel if any input components lt0
• As in DX8
• Needs flag for gt vs
• Introduces aliasing artifacts that wont be
  helped by most FSAA schemes

44
Output Instructions

• out r
• Indicates register to contain final color
• Register r cannot be re-used
• zout r
• Indicates register to contain final z-buffer
  value
• Register r

45
Math Instructions

• Parallel ops
• add, sub, mul, mad, frc, cmp
• Vector ops
• dp3, dp4
• Scalar ops
• rcp, rsq, exp2, log2 , pow
• Logic ops
• min, max, sge, slt

46
Possible Macro Instructions

• From DX8 pixel shaders
• lrp, cnd
• From DX8 vertex shaders
• m4x4, m3x3, dst, lit
• Common requests
• norm, cross, abs, sincos

47
Input Argument Modifiers

• Negate any input
• Arbitrary swizzles/replicates on inputs
• r0.xyzw r2.rgba r0.a
• No complement or bias/_bx2 modifiers

48
Note on Unsigned Textures

• _bx2 functionality still available for loading
  unsigned formats as signed data in range -1..1
• DX6/7/8 content needs this
• Specified on sampler stage declaration or in
  TextureStageState

49
Output Modifiers

• Arbitrary mask on outputs
• r0.y
• _sat is still required as in frc
• Clamps to range 0..1
• No shifts supported
• _d2, _d4, _x2, etc.

50
Co-Issue

• None.

51
Dependent Reads

• Can be serialized, but only to a max depth of 4
• dcl t0.xy
• dcl s0.rg, 2D
• ld r0, s0, t0
• ld r1, s1, r0
• ld r2, s1, r1
• ld r3, s1, r2
• Is legal

52
TextureStageState

• Most components unchanged
• Number of stages updated to 16
• Some components affect the 8 texture parameters
• D3DTSS_WRAPUV
• All others control the 16 samplers
• D3DTSS_FILTER, D3DTSS

53
DXG9 Surface Formats

• Format Filter Blend
• AGBR8 y y
• ABGR10 y y
• ABGR16 y n
• ABGR16f n n
• ABGR32f n n

54
Channel Counts

• Will expose support for the following
• 1-, 2- and 4-channel 32-bit formats
• 1-, 2- and 4-channel 16-bit formats
• No odd-byte formats

55
Multiple Color Buffers

• Looking at writing to multiple color buffers from
  one shader pass
• Concept of compound render targets
• 2 different surfaces in same RT
• Read by setting each surface at separate stage
• No filtering or blending supported in early
  implementations
• Reason for out ps.2.0 instruction

56
DXG Infrastructure Update

• Color converting Present()
• Perf improvements likely
• Including stretch case
• Copy with Vertical Synch flag
• D3DSWAPEFFECT_GDI_MODE
• Enables better GDI integration
• Flag bits at Reset() vs Create()
• Enables more flexible reset

57
Volume Textures

• More compression techniques
• DXTc supported since DX8
• DXVc added in DX9

58
Gamma Correction

• Many apps need correct g math
• Rendering quality
• Math (adds) for light or shadow intensities
  breaks down if g ! 1.0
• Compositing quality
• a-blending transparent objects
• e.g video effects, decals
• a-blending for antialiasing
• greatly improves quality IF g is handled

59
Gamma Correction

• Hardware now has enough precision to handle gamma
  correctly
• Correcting gamma from 8-bit textures was
  pointless on 8-bit hardware
• DXG9 adds explicit API support for gamma
  correction on reads and writes
• 2 color space standards supported
• sRGB gamma 2.2 all current content
• scRGB gamma 1.0 (linear) 16-bit precision

60
Gamma Input Syntax

• New TextureStageState in DXG9
• D3DTSS_GAMMA
• sRGB
• default
• Indicates that gamma correction from 2.2 space to
  linear will occur on read
• Performs exponentiation of each channel by 2.2
  before pixel shader

61
Gamma Output Syntax

• Need to correct back to g 2.2 for framebuffer
  display
• Involves taking high-precision data to 0.45th
  power before write to fb blender
• Enabled by Renderstate
• D3DRS_RTsRGB, TRUE
• Indicates that RenderTarget is in sRGB format
  that is gamma corrected
• relative to pixel shader math

62
Potential Scenarios

• All current texture art authored at g2.2
• 8-bit g 2.2 render target
• 10-bit g 2.2 render target
• 10-bit linear render target
• Correct in DACs?
• 16-bit integer render target

63
Gamma Usage Patterns

• Some applications need to use multipass blending
• Gamma correction is in pixel shader, not frame
  buffer blender
• Apps will need to do lots of
• SetTexture( RenderTarget )

64
DirectX10

• Continue to increase generality
• Conditionals (predicate)
• Subroutines (finite depth stack)
• Still arg calling conventions
• Pixel-Vertex shader integration

65
Pixel-Vertex Integration

• Equivalent precisions
• No double requirement
• Same instruction set
• Implementations may share math units?
• Opaque routines (provided by impl)
• Tessellate( vertex value, vertex value)
• Evaluate( pixel value, vertex value)
• Rationalized input formats
• Vertex streams vs texture pixel fmts

66
DX10 Gamma

• Full gamma support in ps for DX9
• DXG10 hw should support gamma in FB blender
• Performs xg on reads and x1/g on write
• With blend math at gt14-bit precision
• Enables blending of data with arbitrary input
  gamma for multi-pass methods