RTX device: path-tracing correctness improvements#293
Merged
Conversation
Split the global `lastBLASChange` into per-domain stamps so a geometry edit no longer invalidates volume BVHs or per-group light index lists, and vice versa. lastSurfaceBLASChange -> bumped by Geometry, Surface lastVolumeBLASChange -> bumped by Volume, SpatialField lastLightSetChange -> bumped on Group commits (light array rebind) lastTLASChange -> unchanged (Instance) Group::markFinalized bumps all three since any of its arrays may have been rebound. Each Group::rebuild* gates on its own stamp; World fans the trigger across the three for the BLAS rebuild path.
The faceVarying.normal/tangent count check dereferenced m_index unconditionally, crashing on triangle-soup geometry (no index buffer).
The computed-tangent Array1D was built without a deleter, so the cudaMalloc'd memory leaked when the array was released. Also remove the unused bitangent parameter on computeTangents calls.
The bitangent formula was (t.x*e1 - s.x*e2)/det — sign-reversed vs. the standard (s.x*e2 - t.x*e1)/det used by MikkTSpace. Normal-mapped lighting on generated tangents was mirrored along the bitangent axis.
Degenerate triangles or NaN/zero input normals produced fixed (1,0,0)/(0,1,0) tangents that conflicted with the actual surface normal. Add safeNormalize, fall back to a Pixar orthonormal basis built from the geometric normal, and factor the per-corner Gram-Schmidt + handedness into a shared orthogonalizeTangent helper.
Per-triangle T at a shared corner depended on each face's own UV gradient, so orthogonalizing against the same vertex normal still produced different tangents per face — the shading frame jumped at every triangle edge. Replace with two passes: pass 1 atomically accumulates angle-weighted T/B/N into per-vertex slots (MikkTSpace's averaging scheme); pass 2 normalizes and orthogonalizes. Output moves from face-varying (3*numTri) to per-vertex 'vertex.tangent'. UV mirror seams must be vertex-split — same constraint as MikkTSpace defaults.
Tangent fetch was discarding .w (vec3 cast on a vec4 array) and reading handedness from vertex 0 alone. Read the vec4 properly and barycentric- interpolate the sign across the corners, then quantize to +-1 for the basis flip.
Numerical fixes: - Rewrite ggxD denom as alpha2*x + (1-x) instead of x*(alpha2-1) + 1. The textbook form cancels catastrophically when alpha2 < eps(1) and collapses to zero at x=1, producing NaN throughput. - Negate NdotL early-outs so a NaN takes the rejection path instead of slipping through (NaN compares false to <=, > etc.). - Fall back from Ns to Ng in Matte and PBR shading state when the normal length squared is non-positive (catches both NaN and zero). Mirror-seam tangents: - Compute per-vertex bitangents B_i = t_i.w * cross(N_i, T_i) and barycentric-blend B and T independently, instead of blending t.w signs and applying once at the hit. Matches glTF Sample Renderer / PBRT / Filament; avoids carving seam edges into the tangent frame.
The lightDotNg gate and the ambient hemisphere normal both used Ng, which carved per-triangle facet shapes into the lit/unlit boundary at grazing angles on smooth-normal meshes. Switch both to Ns so the terminator follows the shading surface; the material's own NdotL guard still rejects light from below.
randomDir was normalizing a cube-uniform vector — that clusters samples toward cube corners, not uniform on the sphere. Replace with the analytic cosTheta = 1 - 2u mapping. sampleHemisphere had Malley's method inverted (z = u, r = sqrt(1-sqrt(u))) instead of (r = sqrt(u), z = sqrt(1-r^2)), biasing AO and diffuse estimates toward grazing. Fold cos(theta)/pi into the ambient LightSample pdf so it matches the corrected density.
For opaque non-metals reflectProb=1 always, so the indirect bounce was purely glossy GGX and multi-bounce Lambertian light was missing entirely — rough dielectrics rendered black on indirect bounces. Add a third cosine-weighted diffuse lobe with importance proxy (1-F)*(1-metallic)*(1-transmission)*luminance(baseColor); existing reflect/transmit weights now divide by their (now <1) lobe-pick probability. Sampled around Ns to match shadeSurface's diffuseBRDF axis.
Two sequential kernels did identical resolveSample work for color and for the albedo/normal guides. Merge into a single kernel; null-pointer checks gate the guide writes when the guides are absent. Halves the prepare-denoise dispatch overhead per frame.
The original `compMax(vec4(abs(P), compMax(abs(dir))*t))` was redundant: for normalized rays `hitpoint = origin + dir*t`, so `|hitpoint|_inf` already bounds the magnitude that drives ULP-scale self-intersection offsets at the surface. Drop the dir/t inputs and use `|P|_inf` alone. The sole caller (`populateSurfaceHit` in populateHit.h) only runs from __closesthit__/__anyhit__ programs, where `optixGetRayTmax()` is the finite hit distance — there was no functional bug, this is a cleanup.
Smooth-normal triangles produce dark, triangle-shaped bands at grazing light: the planar hit point lies below the smooth surface implied by per-vertex normals, so direct-light shadow rays start inside that implied curvature and self-occlude on the tessellation. Add a shadingHitpoint() helper (Hanika RTGII ch. 4: signed distance from each vertex tangent plane, projected back along the vertex normal, barycentric-blended) and call it at shadow/AO ray origins only. Continuation rays keep the unmodified facet hitpoint — transmission needs it: the smoothed point can sit far enough above the facet that the -Ng*epsilon offset still leaves the origin outside the volume, blocking the refracted path from reaching the back wall.
This breaks cutouts. To be revisited.
Both shadeSurface and nextRay evaluated (1-F) at the GGX half-vector, but the diffuse direction is independent of H so the weight didn't match between NEE and the bounce. Evaluate the diffuse Fresnel at NdotV (Frostbite/Disney) so the two estimators agree at any roughness. Also factor the Fresnel+iridescence block into evalFresnelWithIridescence to remove the duplicated block.
The init() function had two post-hoc dot-product guards to swap a NaN/zero shading normal for the geometric one. Doing it inside sampleNormalMap means every caller gets a usable normal back without the cleanup pass downstream.
The field stores the refraction ratio (n1/n2 from the incident side), pre-inverted on front-facing hits so it can feed glm::refract directly. Calling it 'ior' invited reading it as the material's IOR. Move the 'pre-inverted' note onto the struct field so the contract lives with the data.
Clearcoat was only evaluated in NEE, so smooth clearcoats over matte bases never picked up HDRI/sky reflections on the bounce path.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Quality improvements:
Normal mapping fixes:
Stability improvements:
Performance improvements: