Solari specular scene/PT support #20242
Conversation
…into solari6-perfopt-good
JMS55
added
C-Feature
| // https://gpuopen.com/download/Bounded_VNDF_Sampling_for_Smith-GGX_Reflections.pdf (Listing 1) | ||
| fn sample_ggx_vndf(wi_tangent: vec3<f32>, roughness: f32, rng: ptr<function, u32>) -> vec3<f32> { | ||
| let i = wi_tangent; | ||
| let rand = rand_vec2f(rng); | ||
| let i_std = normalize(vec3(i.xy * roughness, i.z)); | ||
| let phi = 2.0 * PI * rand.x; | ||
| let a = roughness; | ||
| let s = 1.0 + length(vec2(i.xy)); | ||
| let a2 = a * a; | ||
| let s2 = s * s; | ||
| let k = (1.0 - a2) * s2 / (s2 + a2 * i.z * i.z); | ||
| let b = select(i_std.z, k * i_std.z, i.z > 0.0); | ||
| let z = fma(1.0 - rand.y, 1.0 + b, -b); | ||
| let sin_theta = sqrt(saturate(1.0 - z * z)); | ||
| let o_std = vec3(sin_theta * cos(phi), sin_theta * sin(phi), z); | ||
| let m_std = i_std + o_std; | ||
| let m = normalize(vec3(m_std.xy * roughness, m_std.z)); | ||
| return 2.0 * dot(i, m) * m - i; | ||
| } | ||
|
|
||
| // https://gpuopen.com/download/Bounded_VNDF_Sampling_for_Smith-GGX_Reflections.pdf (Listing 2) | ||
| fn ggx_vndf_pdf(wi_tangent: vec3<f32>, wo_tangent: vec3<f32>, roughness: f32) -> f32 { | ||
| let i = wi_tangent; | ||
| let o = wo_tangent; | ||
| let m = normalize(i + o); | ||
| let ndf = D_GGX(roughness, saturate(m.z)); | ||
| let ai = roughness * i.xy; | ||
| let len2 = dot(ai, ai); | ||
| let t = sqrt(len2 + i.z * i.z); | ||
| if i.z >= 0.0 { | ||
| let a = roughness; | ||
| let s = 1.0 + length(i.xy); | ||
| let a2 = a * a; | ||
| let s2 = s * s; | ||
| let k = (1.0 - a2) * s2 / (s2 + a2 * i.z * i.z); | ||
| return ndf / (2.0 * (k * i.z + t)); | ||
| } | ||
| return ndf * (t - i.z) / (2.0 * len2); | ||
| } |
There was a problem hiding this comment.
There was a problem hiding this comment.
Since Jasmine's version is less code, I wonder why the big difference and whether i could just use this function in my case as is. I'll test it out and share my findings.
|
I tried this out, started going through the code, and looks very good at a first glance. cargo run --example solari --features="bevy_solari" -- --pathtracerI got curious to see if I could get an infinite mirror in the cornell box, but the reflections beyond a certain number of bounces don't show up. Maybe this is a setting somewhere that I am unaware of? Either way results are neat. |
|
Just came here to add that a 1 to 1 compatibility with the solari material system and materialx is disirable. |
|
There's vague plans to support MaterialX in bevy, but for solari, we're unlikely to support such a wide range of materials. Raytracing is already super expensive, adding lots of varied materials on top makes it even slower :) |
| let B = TBN[1]; | ||
| let N = TBN[2]; | ||
|
|
||
| let wo_tangent = vec3(dot(wo, T), dot(wo, B), dot(wo, N)); |
There was a problem hiding this comment.
If TBN is a mat3<f32>, this is just TBN * wo
There was a problem hiding this comment.
Hmm this is the code we use everywhere else in bevy 😅
| var wi_tangent: vec3<f32>; | ||
| if rand_f(rng) < diffuse_weight { | ||
| wi = sample_cosine_hemisphere(ray_hit.world_normal, rng); | ||
| wi_tangent = vec3(dot(wi, T), dot(wi, B), dot(wi, N)); |
| wi_tangent = vec3(dot(wi, T), dot(wi, B), dot(wi, N)); | ||
| } else { | ||
| wi_tangent = sample_ggx_vndf(wo_tangent, ray_hit.material.roughness, rng); | ||
| wi = wi_tangent.x * T + wi_tangent.y * B + wi_tangent.z * N; |
There was a problem hiding this comment.
And this is a mul with the transpose.
| wi = wi_tangent.x * T + wi_tangent.y * B + wi_tangent.z * N; | ||
| } | ||
|
|
||
| let diffuse_pdf = dot(wi, ray_hit.world_normal) / PI; |
There was a problem hiding this comment.
We have wi_tangent here, can just use the Z component (I assume that's up) instead of the dot.
Note that specular support for the realtime lighting plugin will come at a future date. I think I want to focus on getting GI more stable and hooking up a denoiser first. This is just support for the material properties in RaytracingScenePlugin, and specular lighting in the PathtracingPlugin.