浏览代码

Save 1 VGPR by removing 'unNdotV' from PreLightData

/Branch_Batching2
Evgenii Golubev 8 年前
当前提交
80077f9a
共有 2 个文件被更改,包括 49 次插入44 次删除
  1. 69
      Assets/ScriptableRenderPipeline/HDRenderPipeline/Material/Lit/Lit.hlsl
  2. 24
      Assets/ScriptableRenderPipeline/ShaderLibrary/CommonLighting.hlsl

69
Assets/ScriptableRenderPipeline/HDRenderPipeline/Material/Lit/Lit.hlsl


#define LTC_LUT_SCALE ((LTC_LUT_SIZE - 1) * rcp(LTC_LUT_SIZE))
#define LTC_LUT_OFFSET (0.5 * rcp(LTC_LUT_SIZE))
#define MIN_N_DOT_V 0.0001 // The minimum value of 'NdotV'
// SSS parameters
#define SSS_N_PROFILES 8
#define SSS_LOW_THICKNESS 0.005 // 0.5 cm

struct PreLightData
{
// General
float NdotV; // Between 0.0001 and 1
float unNdotV; // Between -1 and 1
float NdotV; // Geometric version (not clamped)
// GGX iso
float ggxLambdaV;

float anisoGGXLambdaV;
// IBL
float3 iblDirWS; // Dominant specular direction, used for IBL in EvaluateBSDF_Env()
float3 iblDirWS; // Dominant specular direction, used for IBL in EvaluateBSDF_Env()
float3 specularFGD; // Store preconvole BRDF for both specular and diffuse
float3 specularFGD; // Store preconvoled BRDF for both specular and diffuse
float diffuseFGD;
// area light

{
PreLightData preLightData;
// General
float3 iblNormalWS = bsdfData.normalWS;
float NdotV = dot(bsdfData.normalWS, V);
float3 iblNormalWS = GetViewShiftedNormal(bsdfData.normalWS, V, NdotV, MIN_N_DOT_V);
preLightData.unNdotV = dot(bsdfData.normalWS, V);
// GetShiftedNdotV return a positive NdotV
// In case a material use negative normal for double sided lighting like Speedtree they need to do a new calculation
preLightData.NdotV = GetShiftedNdotV(iblNormalWS, V, preLightData.unNdotV);
preLightData.NdotV = NdotV; // Store the unaltered (geometric) version
NdotV = max(NdotV, MIN_N_DOT_V); // Use the modified (clamped) version
float3 iblR = reflect(-V, iblNormalWS);
preLightData.ggxLambdaV = GetSmithJointGGXLambdaV(preLightData.NdotV, bsdfData.roughness);
preLightData.ggxLambdaV = GetSmithJointGGXLambdaV(NdotV, bsdfData.roughness);
// GGX aniso
if (bsdfData.materialId == MATERIALID_LIT_ANISO)

preLightData.anisoGGXLambdaV = GetSmithJointGGXAnisoLambdaV(preLightData.TdotV, preLightData.BdotV, preLightData.NdotV, bsdfData.roughnessT, bsdfData.roughnessB);
preLightData.anisoGGXLambdaV = GetSmithJointGGXAnisoLambdaV(preLightData.TdotV, preLightData.BdotV, NdotV, bsdfData.roughnessT, bsdfData.roughnessB);
iblNormalWS = GetAnisotropicModifiedNormal(bsdfData.bitangentWS, iblNormalWS, V, bsdfData.anisotropy);
float3 anisoIblNormalWS = GetAnisotropicModifiedNormal(bsdfData.bitangentWS, iblNormalWS, V, bsdfData.anisotropy);
iblR = reflect(-V, anisoIblNormalWS);
// @Sébastien: I preserved the original behavior by creating 'anisoIblNormalWS', but,
// from the performance standpoint, it would be best to store the value in 'iblNormalWS'
// and then use it in GetSpecularDominantDir(). Please reconsider. :-) -Evgenii
GetPreIntegratedFGD(preLightData.NdotV, bsdfData.perceptualRoughness, bsdfData.fresnel0, preLightData.specularFGD, preLightData.diffuseFGD);
GetPreIntegratedFGD(NdotV, bsdfData.perceptualRoughness, bsdfData.fresnel0, preLightData.specularFGD, preLightData.diffuseFGD);
// We need to take into account the modified normal for faking anisotropic here.
float3 iblR = reflect(-V, iblNormalWS);
preLightData.iblDirWS = GetSpecularDominantDir(bsdfData.normalWS, iblR, bsdfData.roughness, preLightData.NdotV);
preLightData.iblDirWS = GetSpecularDominantDir(iblNormalWS, iblR, bsdfData.roughness, NdotV);
float theta = FastACos(preLightData.NdotV);
float theta = FastACos(NdotV);
float2 uv = LTC_LUT_OFFSET + LTC_LUT_SCALE * float2(bsdfData.perceptualRoughness, theta * INV_HALF_PI);
// Get the inverse LTC matrix for GGX

out float3 diffuseLighting,
out float3 specularLighting)
{
// Optimized math. Ref: PBR Diffuse Lighting for GGX + Smith Microsurfaces (slide 114).
float NdotV = preLightData.unNdotV; // This value must not be clamped
float NdotV = preLightData.NdotV; // Get the unaltered (geometric) version
// GCN Optimization: reference PBR Diffuse Lighting for GGX + Smith Microsurfaces
float invLenLV = rsqrt(abs(2.0 * LdotV + 2.0)); // invLenLV = rcp(length(L + V))
float invLenLV = rsqrt(abs(2 * LdotV + 2)); // invLenLV = rcp(length(L + V))
NdotV = max(NdotV, MIN_N_DOT_V); // Use the modified (clamped) version
float3 F = F_Schlick(bsdfData.fresnel0, LdotH);

bsdfData.roughnessB = ClampRoughnessForAnalyticalLights(bsdfData.roughnessB);
#ifdef LIT_USE_BSDF_PRE_LAMBDAV
Vis = V_SmithJointGGXAnisoLambdaV(preLightData.TdotV, preLightData.BdotV, preLightData.NdotV, TdotL, BdotL, NdotL,
Vis = V_SmithJointGGXAnisoLambdaV(preLightData.TdotV, preLightData.BdotV, NdotV, TdotL, BdotL, NdotL,
Vis = V_SmithJointGGXAniso(preLightData.TdotV, preLightData.BdotV, preLightData.NdotV, TdotL, BdotL, NdotL,
Vis = V_SmithJointGGXAniso(preLightData.TdotV, preLightData.BdotV, NdotV, TdotL, BdotL, NdotL,
bsdfData.roughnessT, bsdfData.roughnessB);
#endif

bsdfData.roughness = ClampRoughnessForAnalyticalLights(bsdfData.roughness);
#ifdef LIT_USE_BSDF_PRE_LAMBDAV
Vis = V_SmithJointGGX(NdotL, preLightData.NdotV, bsdfData.roughness, preLightData.ggxLambdaV);
Vis = V_SmithJointGGX(NdotL, NdotV, bsdfData.roughness, preLightData.ggxLambdaV);
Vis = V_SmithJointGGX(NdotL, preLightData.NdotV, bsdfData.roughness);
Vis = V_SmithJointGGX(NdotL, NdotV, bsdfData.roughness);
#endif
D = D_GGX(NdotH, bsdfData.roughness);
}

float diffuseTerm = Lambert();
#elif LIT_DIFFUSE_GGX_BRDF
float3 diffuseTerm = DiffuseGGX(bsdfData.diffuseColor, preLightData.NdotV, NdotL, NdotH, LdotV, bsdfData.perceptualRoughness);
float3 diffuseTerm = DiffuseGGX(bsdfData.diffuseColor, NdotV, NdotL, NdotH, LdotV, bsdfData.perceptualRoughness);
float diffuseTerm = DisneyDiffuse(preLightData.NdotV, NdotL, LdotH, bsdfData.perceptualRoughness);
float diffuseTerm = DisneyDiffuse(NdotV, NdotL, LdotH, bsdfData.perceptualRoughness);
#endif
diffuseLighting = bsdfData.diffuseColor * diffuseTerm;

uint sampleCount = 4096)
{
float3x3 localToWorld = float3x3(bsdfData.tangentWS, bsdfData.bitangentWS, bsdfData.normalWS);
float NdotV = max(preLightData.NdotV, MIN_N_DOT_V);
float3 acc = float3(0.0, 0.0, 0.0);
// Add some jittering on Hammersley2d

float LdotH = dot(L, H);
// Note: we call DisneyDiffuse that require to multiply by Albedo / PI. Divide by PI is already taken into account
// in weightOverPdf of ImportanceSampleLambert call.
float disneyDiffuse = DisneyDiffuse(preLightData.NdotV, NdotL, LdotH, bsdfData.perceptualRoughness);
float disneyDiffuse = DisneyDiffuse(NdotV, NdotL, LdotH, bsdfData.perceptualRoughness);
// diffuse Albedo is apply here as describe in ImportanceSampleLambert function
float4 val = SampleEnv(lightLoopContext, lightData.envIndex, L, 0);

uint sampleCount = 4096)
{
float3x3 localToWorld = float3x3(bsdfData.tangentWS, bsdfData.bitangentWS, bsdfData.normalWS);
float NdotV = max(preLightData.NdotV, MIN_N_DOT_V);
float3 acc = float3(0.0, 0.0, 0.0);
// Add some jittering on Hammersley2d

// GGX BRDF
if (bsdfData.materialId == MATERIALID_LIT_ANISO)
{
ImportanceSampleAnisoGGX(u, V, localToWorld, bsdfData.roughnessT, bsdfData.roughnessB, preLightData.NdotV, L, VdotH, NdotL, weightOverPdf);
ImportanceSampleAnisoGGX(u, V, localToWorld, bsdfData.roughnessT, bsdfData.roughnessB, NdotV, L, VdotH, NdotL, weightOverPdf);
ImportanceSampleGGX(u, V, localToWorld, bsdfData.roughness, preLightData.NdotV, L, VdotH, NdotL, weightOverPdf);
ImportanceSampleGGX(u, V, localToWorld, bsdfData.roughness, NdotV, L, VdotH, NdotL, weightOverPdf);
}

24
Assets/ScriptableRenderPipeline/ShaderLibrary/CommonLighting.hlsl


// Helper functions
//-----------------------------------------------------------------------------
// NdotV can be negative for visible pixels due to the perspective projection, the normal mapping and decals.
// This can produce visible artifacts with direct specular lighting (white point, black point) and indirect specular (artifact with cubemap fetch)
// A way to reduce artifact is to limit NdotV value to not be negative and calculate reflection vector for cubemap with a shifted normal (i.e what depends on the view)
// This is what provide this function
// Note: NdotV return by this function is always positive, no need for saturate
float GetShiftedNdotV(inout float3 N, float3 V, float NdotV)
// 'NdotV' can become negative for visible pixels due to the perspective projection, normal mapping and decals.
// This can produce visible artifacts under specular lighting, both direct (overly dark/bright pixels) and indirect (incorrect cubemap direction).
// One way of avoiding these artifacts is to limit the value of 'NdotV' to a small positive number,
// and calculate the reflection vector for the cubemap fetch using a normal shifted into view.
float3 GetViewShiftedNormal(float3 N, float3 V, float NdotV, float minNdotV)
const float limit = 0.0001; // Epsilon value that avoid divide by 0 (several BSDF divide by NdotV)
if (NdotV < limit)
if (NdotV < minNdotV)
// We do not renormalize the normal because { abs(length(N) - 1.0) < limit } + It is use for cubemap
N += (-NdotV + limit) * V;
NdotV = limit;
// We do not renormalize the normal to save a few clock cycles.
// The magnitude difference is typically negligible, and the normal is only used to compute
// the reflection vector for the IBL cube map fetch (which does not depend on the magnitude).
N += (-NdotV + minNdotV) * V;
return NdotV;
return N;
}
// Generates an orthonormal basis from a unit vector.

正在加载...
取消
保存