Deferred shading and G-buffer encoding

Enhanced deferred pixel programs write several render targets that later lighting programs consume. The layout below describes inspected PS_DeferredTextured* programs; individual effects can write constants, discard pixels or use a different encoding. Observed · program evaluation and disassembly

Target	Channels	Role	Confidence
0	RGB	Diffuse colour, with material/texture and wetness influences	Observed
0	A	Draw-state byte from engine constants	Observed
1	RGB	World normal, usually 0.5*N + 0.5	Observed
1	A	Additional normal precision; some effects pack 3/3/2 bits of normal components	Observed
2	R	Square root of specular intensity	Observed
2	G	Square root of specular falloff divided by 512	Observed
2	B	Fresnel parameter	Observed
3	R/G	Natural/artificial ambient occlusion	Observed
3	B	Emission divided by 16	Observed
3	A	Ambient-path flag; values above one half select an additional multiplier	Observed
4	R/G	Previous-frame screen displacement in pixels	Observed

Target 0 holds the square root of the diffuse colour: the sampled point- and sun-light programs square it before use. Observed · light-program execution

Normal encoding varies: the twiddle effects quantize RGB as floor(128*N + 128)/256 and carry residual bits in alpha. Other inspected effects write saturate(128*N.z - 127) in alpha. A fourth normal component is not a material opacity channel. Observed · program evaluation

The inspected occlusion path stores sqrt((a + k*vertexColour)/2) through engine constants; the ambient pass applies the square of twice the stored square. Treating the encoded occlusion as a direct linear multiplier changes the response. Observed · program evaluation

Specular material path

For the inspected normal_spec_decal deferred program, with specular texture sample s, the output is as follows. Observed · DXIL disassembly

Output	Expression	Confidence
RT2.x	sqrt(SpecularColor * dot((s.r², s.g², s.b), specMapIntMask))	Observed
RT2.y	sqrt(Specular * s.a / 512)	Observed
RT2.z	Fresnel	Observed

The directional-light program reconstructs intensity I = RT2.x² and falloff n = RT2.y² * 512. Its exponent is 3*n below 500 and 1500 + 558*(n-500) above that point. The specular term includes (exponent+2)/8, the Blinn-Phong power of N·H, Schlick Fresnel with F0 = 1-Fresnel and F90 = 1, intensity, N·L and light colour. Observed · DXIL disassembly

The same inspected light path scales diffuse by 1 - I*F(N·V). Dry default has no material specular parameters and writes zero intensity and falloff; a parameter-free material is not evidence for a nonzero default specular term. Observed · DXIL disassembly

Road markings with normal_spec_decal draw in the deferred_draw pass over the G-buffer. Their pixel program writes the alpha diffuse.w * COLOR0.w * globalScalars.x (the diffuse sample’s alpha, the vertex colour’s alpha and the visibility factor) into the alpha of targets 0, 1 and 2, which blend SrcAlpha / InvSrcAlpha; with globalScalars.x at 0 a crosswalk’s markings leave nothing, at 1 they appear. Observed · DXIL disassembly, and drawing the program with both values

Sampling and scope

Inspected lighting passes sample the G-buffer using TEXCOORD0.xy / TEXCOORD0.w, a projected interpolant, rather than an integer load at SV_Position. Observed · program inspection

The formulas on this page describe sampled compiled permutations. Shared effect-family naming does not prove every permutation has identical output stores; per-program output inspection is needed for that claim. Unknown