bevy/crates/bevy_pbr/src/render/pbr_lighting.wgsl
Robert Swain b44b606d29 bevy_pbr: Avoid copying structs and using registers in shaders (#7069)
# Objective

- The #7064 PR had poor performance on an M1 Max in MacOS due to significant overuse of registers resulting in 'register spilling' where data that would normally be stored in registers on the GPU is instead stored in VRAM. The latency to read from/write to VRAM instead of registers incurs a significant performance penalty.
- Use of registers is a limiting factor in shader performance. Assignment of a struct from memory to a local variable can incur copies. Passing a variable that has struct type as an argument to a function can also incur copies. As such, these two cases can incur increased register usage and decreased performance.

## Solution

- Remove/avoid a number of assignments of light struct type data to local variables.
- Remove/avoid a number of passing light struct type variables/data as value arguments to shader functions.
2023-01-02 22:07:33 +00:00

250 lines
11 KiB
WebGPU Shading Language
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#define_import_path bevy_pbr::lighting
// From the Filament design doc
// https://google.github.io/filament/Filament.html#table_symbols
// Symbol Definition
// v View unit vector
// l Incident light unit vector
// n Surface normal unit vector
// h Half unit vector between l and v
// f BRDF
// f_d Diffuse component of a BRDF
// f_r Specular component of a BRDF
// α Roughness, remapped from using input perceptualRoughness
// σ Diffuse reflectance
// Ω Spherical domain
// f0 Reflectance at normal incidence
// f90 Reflectance at grazing angle
// χ+(a) Heaviside function (1 if a>0 and 0 otherwise)
// nior Index of refraction (IOR) of an interface
// ⟨n⋅l⟩ Dot product clamped to [0..1]
// ⟨a⟩ Saturated value (clamped to [0..1])
// The Bidirectional Reflectance Distribution Function (BRDF) describes the surface response of a standard material
// and consists of two components, the diffuse component (f_d) and the specular component (f_r):
// f(v,l) = f_d(v,l) + f_r(v,l)
//
// The form of the microfacet model is the same for diffuse and specular
// f_r(v,l) = f_d(v,l) = 1 / { |n⋅v||n⋅l| } ∫_Ω D(m,α) G(v,l,m) f_m(v,l,m) (v⋅m) (l⋅m) dm
//
// In which:
// D, also called the Normal Distribution Function (NDF) models the distribution of the microfacets
// G models the visibility (or occlusion or shadow-masking) of the microfacets
// f_m is the microfacet BRDF and differs between specular and diffuse components
//
// The above integration needs to be approximated.
// distanceAttenuation is simply the square falloff of light intensity
// combined with a smooth attenuation at the edge of the light radius
//
// light radius is a non-physical construct for efficiency purposes,
// because otherwise every light affects every fragment in the scene
fn getDistanceAttenuation(distanceSquare: f32, inverseRangeSquared: f32) -> f32 {
let factor = distanceSquare * inverseRangeSquared;
let smoothFactor = saturate(1.0 - factor * factor);
let attenuation = smoothFactor * smoothFactor;
return attenuation * 1.0 / max(distanceSquare, 0.0001);
}
// Normal distribution function (specular D)
// Based on https://google.github.io/filament/Filament.html#citation-walter07
// D_GGX(h,α) = α^2 / { π ((n⋅h)^2 (α21) + 1)^2 }
// Simple implementation, has precision problems when using fp16 instead of fp32
// see https://google.github.io/filament/Filament.html#listing_speculardfp16
fn D_GGX(roughness: f32, NoH: f32, h: vec3<f32>) -> f32 {
let oneMinusNoHSquared = 1.0 - NoH * NoH;
let a = NoH * roughness;
let k = roughness / (oneMinusNoHSquared + a * a);
let d = k * k * (1.0 / PI);
return d;
}
// Visibility function (Specular G)
// V(v,l,a) = G(v,l,α) / { 4 (n⋅v) (n⋅l) }
// such that f_r becomes
// f_r(v,l) = D(h,α) V(v,l,α) F(v,h,f0)
// where
// V(v,l,α) = 0.5 / { n⋅l sqrt((n⋅v)^2 (1α2) + α2) + n⋅v sqrt((n⋅l)^2 (1α2) + α2) }
// Note the two sqrt's, that may be slow on mobile, see https://google.github.io/filament/Filament.html#listing_approximatedspecularv
fn V_SmithGGXCorrelated(roughness: f32, NoV: f32, NoL: f32) -> f32 {
let a2 = roughness * roughness;
let lambdaV = NoL * sqrt((NoV - a2 * NoV) * NoV + a2);
let lambdaL = NoV * sqrt((NoL - a2 * NoL) * NoL + a2);
let v = 0.5 / (lambdaV + lambdaL);
return v;
}
// Fresnel function
// see https://google.github.io/filament/Filament.html#citation-schlick94
// F_Schlick(v,h,f_0,f_90) = f_0 + (f_90 f_0) (1 v⋅h)^5
fn F_Schlick_vec(f0: vec3<f32>, f90: f32, VoH: f32) -> vec3<f32> {
// not using mix to keep the vec3 and float versions identical
return f0 + (f90 - f0) * pow(1.0 - VoH, 5.0);
}
fn F_Schlick(f0: f32, f90: f32, VoH: f32) -> f32 {
// not using mix to keep the vec3 and float versions identical
return f0 + (f90 - f0) * pow(1.0 - VoH, 5.0);
}
fn fresnel(f0: vec3<f32>, LoH: f32) -> vec3<f32> {
// f_90 suitable for ambient occlusion
// see https://google.github.io/filament/Filament.html#lighting/occlusion
let f90 = saturate(dot(f0, vec3<f32>(50.0 * 0.33)));
return F_Schlick_vec(f0, f90, LoH);
}
// Specular BRDF
// https://google.github.io/filament/Filament.html#materialsystem/specularbrdf
// Cook-Torrance approximation of the microfacet model integration using Fresnel law F to model f_m
// f_r(v,l) = { D(h,α) G(v,l,α) F(v,h,f0) } / { 4 (n⋅v) (n⋅l) }
fn specular(f0: vec3<f32>, roughness: f32, h: vec3<f32>, NoV: f32, NoL: f32,
NoH: f32, LoH: f32, specularIntensity: f32) -> vec3<f32> {
let D = D_GGX(roughness, NoH, h);
let V = V_SmithGGXCorrelated(roughness, NoV, NoL);
let F = fresnel(f0, LoH);
return (specularIntensity * D * V) * F;
}
// Diffuse BRDF
// https://google.github.io/filament/Filament.html#materialsystem/diffusebrdf
// fd(v,l) = σ/π * 1 / { |n⋅v||n⋅l| } ∫Ω D(m,α) G(v,l,m) (v⋅m) (l⋅m) dm
//
// simplest approximation
// float Fd_Lambert() {
// return 1.0 / PI;
// }
//
// vec3 Fd = diffuseColor * Fd_Lambert();
//
// Disney approximation
// See https://google.github.io/filament/Filament.html#citation-burley12
// minimal quality difference
fn Fd_Burley(roughness: f32, NoV: f32, NoL: f32, LoH: f32) -> f32 {
let f90 = 0.5 + 2.0 * roughness * LoH * LoH;
let lightScatter = F_Schlick(1.0, f90, NoL);
let viewScatter = F_Schlick(1.0, f90, NoV);
return lightScatter * viewScatter * (1.0 / PI);
}
// From https://www.unrealengine.com/en-US/blog/physically-based-shading-on-mobile
fn EnvBRDFApprox(f0: vec3<f32>, perceptual_roughness: f32, NoV: f32) -> vec3<f32> {
let c0 = vec4<f32>(-1.0, -0.0275, -0.572, 0.022);
let c1 = vec4<f32>(1.0, 0.0425, 1.04, -0.04);
let r = perceptual_roughness * c0 + c1;
let a004 = min(r.x * r.x, exp2(-9.28 * NoV)) * r.x + r.y;
let AB = vec2<f32>(-1.04, 1.04) * a004 + r.zw;
return f0 * AB.x + AB.y;
}
fn perceptualRoughnessToRoughness(perceptualRoughness: f32) -> f32 {
// clamp perceptual roughness to prevent precision problems
// According to Filament design 0.089 is recommended for mobile
// Filament uses 0.045 for non-mobile
let clampedPerceptualRoughness = clamp(perceptualRoughness, 0.089, 1.0);
return clampedPerceptualRoughness * clampedPerceptualRoughness;
}
fn point_light(
world_position: vec3<f32>, light_id: u32, roughness: f32, NdotV: f32, N: vec3<f32>, V: vec3<f32>,
R: vec3<f32>, F0: vec3<f32>, diffuseColor: vec3<f32>
) -> vec3<f32> {
let light = &point_lights.data[light_id];
let light_to_frag = (*light).position_radius.xyz - world_position.xyz;
let distance_square = dot(light_to_frag, light_to_frag);
let rangeAttenuation =
getDistanceAttenuation(distance_square, (*light).color_inverse_square_range.w);
// Specular.
// Representative Point Area Lights.
// see http://blog.selfshadow.com/publications/s2013-shading-course/karis/s2013_pbs_epic_notes_v2.pdf p14-16
let a = roughness;
let centerToRay = dot(light_to_frag, R) * R - light_to_frag;
let closestPoint = light_to_frag + centerToRay * saturate((*light).position_radius.w * inverseSqrt(dot(centerToRay, centerToRay)));
let LspecLengthInverse = inverseSqrt(dot(closestPoint, closestPoint));
let normalizationFactor = a / saturate(a + ((*light).position_radius.w * 0.5 * LspecLengthInverse));
let specularIntensity = normalizationFactor * normalizationFactor;
var L: vec3<f32> = closestPoint * LspecLengthInverse; // normalize() equivalent?
var H: vec3<f32> = normalize(L + V);
var NoL: f32 = saturate(dot(N, L));
var NoH: f32 = saturate(dot(N, H));
var LoH: f32 = saturate(dot(L, H));
let specular_light = specular(F0, roughness, H, NdotV, NoL, NoH, LoH, specularIntensity);
// Diffuse.
// Comes after specular since its NoL is used in the lighting equation.
L = normalize(light_to_frag);
H = normalize(L + V);
NoL = saturate(dot(N, L));
NoH = saturate(dot(N, H));
LoH = saturate(dot(L, H));
let diffuse = diffuseColor * Fd_Burley(roughness, NdotV, NoL, LoH);
// See https://google.github.io/filament/Filament.html#mjx-eqn-pointLightLuminanceEquation
// Lout = f(v,l) Φ / { 4 π d^2 }⟨n⋅l⟩
// where
// f(v,l) = (f_d(v,l) + f_r(v,l)) * light_color
// Φ is luminous power in lumens
// our rangeAttentuation = 1 / d^2 multiplied with an attenuation factor for smoothing at the edge of the non-physical maximum light radius
// For a point light, luminous intensity, I, in lumens per steradian is given by:
// I = Φ / 4 π
// The derivation of this can be seen here: https://google.github.io/filament/Filament.html#mjx-eqn-pointLightLuminousPower
// NOTE: (*light).color.rgb is premultiplied with (*light).intensity / 4 π (which would be the luminous intensity) on the CPU
// TODO compensate for energy loss https://google.github.io/filament/Filament.html#materialsystem/improvingthebrdfs/energylossinspecularreflectance
return ((diffuse + specular_light) * (*light).color_inverse_square_range.rgb) * (rangeAttenuation * NoL);
}
fn spot_light(
world_position: vec3<f32>, light_id: u32, roughness: f32, NdotV: f32, N: vec3<f32>, V: vec3<f32>,
R: vec3<f32>, F0: vec3<f32>, diffuseColor: vec3<f32>
) -> vec3<f32> {
// reuse the point light calculations
let point_light = point_light(world_position, light_id, roughness, NdotV, N, V, R, F0, diffuseColor);
let light = &point_lights.data[light_id];
// reconstruct spot dir from x/z and y-direction flag
var spot_dir = vec3<f32>((*light).light_custom_data.x, 0.0, (*light).light_custom_data.y);
spot_dir.y = sqrt(max(0.0, 1.0 - spot_dir.x * spot_dir.x - spot_dir.z * spot_dir.z));
if (((*light).flags & POINT_LIGHT_FLAGS_SPOT_LIGHT_Y_NEGATIVE) != 0u) {
spot_dir.y = -spot_dir.y;
}
let light_to_frag = (*light).position_radius.xyz - world_position.xyz;
// calculate attenuation based on filament formula https://google.github.io/filament/Filament.html#listing_glslpunctuallight
// spot_scale and spot_offset have been precomputed
// note we normalize here to get "l" from the filament listing. spot_dir is already normalized
let cd = dot(-spot_dir, normalize(light_to_frag));
let attenuation = saturate(cd * (*light).light_custom_data.z + (*light).light_custom_data.w);
let spot_attenuation = attenuation * attenuation;
return point_light * spot_attenuation;
}
fn directional_light(light_id: u32, roughness: f32, NdotV: f32, normal: vec3<f32>, view: vec3<f32>, R: vec3<f32>, F0: vec3<f32>, diffuseColor: vec3<f32>) -> vec3<f32> {
let light = &lights.directional_lights[light_id];
let incident_light = (*light).direction_to_light.xyz;
let half_vector = normalize(incident_light + view);
let NoL = saturate(dot(normal, incident_light));
let NoH = saturate(dot(normal, half_vector));
let LoH = saturate(dot(incident_light, half_vector));
let diffuse = diffuseColor * Fd_Burley(roughness, NdotV, NoL, LoH);
let specularIntensity = 1.0;
let specular_light = specular(F0, roughness, half_vector, NdotV, NoL, NoH, LoH, specularIntensity);
return (specular_light + diffuse) * (*light).color.rgb * NoL;
}