fix bugs and reclaim performance

Author: SimonDanisch

Project.toml

@@ -23,7 +23,6 @@ GeometryBasics = "0.4"
 ImageCore = "0.10"
 ImageIO = "0.6"
 ProgressMeter = "1.10"
-Random = "1.11.0"
 RandomNumbers = "1.6.0"
 StaticArrays = "1.9.7"
 StructArrays = "0.6.18"

docs/code/caustic_glass.jl

@@ -45,7 +45,7 @@ for t in floor_triangles
     push!(primitives, Trace.GeometricPrimitive(t, plastic))
 end
 
-bvh = Trace.BVHAccel(primitives, 1);
+bvh = Trace.BVHAccel(map(identity, primitives), 1);
 
 from, to = Point3f(0, 2, 0), Point3f(-5, 0, 5)
 cone_angle, cone_δ_angle = 30f0, 10f0
@@ -73,14 +73,14 @@ lights = [
 
 scene = Trace.Scene(lights, bvh)
 
-resolution = Point2f(1024)
 n_samples = 8
 ray_depth = 5
 
 look_point = Point3f(-3, 0, -91)
 screen = Trace.Bounds2(Point2f(-1f0), Point2f(1f0))
 filter = Trace.LanczosSincFilter(Point2f(1f0), 3f0)
 
+resolution = Point2f(1024)
 ir = Int64.(resolution)
 film = Trace.Film(
     resolution, Trace.Bounds2(Point2f(0), Point2f(1)),
@@ -92,5 +92,6 @@ camera = Trace.PerspectiveCamera(
     screen, 0f0, 1f0, 0f0, 1f6, 90f0, film,
 )
 
-integrator = Trace.SPPMIntegrator(camera, 0.075f0, ray_depth, 10)
-integrator(scene)
+integrator = Trace.SPPMIntegrator(camera, 0.075f0, ray_depth, 1)
+@profview_allocs integrator(scene)
+@time integrator(scene)

src/accel/bvh.jl

@@ -225,17 +225,18 @@ function intersect!(bvh::BVHAccel{P}, ray::AbstractRay)::Tuple{Bool,P,SurfaceInt
         nodes_to_visit[i] = Int32(0)
     end
     primitive::P = first(bvh.primitives)
+    primitives = bvh.primitives::Vector{P}
     @inbounds while true
         ln = bvh.nodes[current_node_i]
         if intersect_p(ln.bounds, ray, inv_dir, dir_is_neg)
             if ln isa LinearBVHLeaf && ln.n_primitives > 0
                 # Intersect ray with primitives in node.
                 for i in 0:ln.n_primitives-1
-                    tmp_primitive = bvh.primitives[ln.primitives_offset+i]
+                    tmp_primitive::P = primitives[ln.primitives_offset+i]
                     tmp_hit, ray, tmp_interaction = intersect_p!(
                         tmp_primitive, ray,
                     )
-                    if tmp_hit && !(tmp_interaction isa Vec3)
+                    if tmp_hit
                         hit = tmp_hit
                         interaction = tmp_interaction
                         primitive = tmp_primitive

src/bounds.jl

@@ -181,7 +181,7 @@ end
 """
 dir_is_negative: 1 -- false, 2 -- true
 """
-function intersect_p(
+@inline function intersect_p(
         b::Bounds3, ray::AbstractRay,
         inv_dir::Vec3f, dir_is_negative::Point3{UInt8},
     )::Bool

src/camera/perspective.jl

@@ -108,6 +108,5 @@ end
     )
     # TODO add medium
     ray = apply(camera.core.core.camera_to_world, ray)
-    Ray(ray, d=normalize(ray.d))
-    ray, 1f0
+    return Ray(ray, d=normalize(ray.d), time=time), 1.0f0
 end

src/film.jl

@@ -108,6 +108,7 @@ function FilmTile(
         bounds::Bounds2, filter_radius::Point2f,
         filter_table_width::Int32,
     )
+    # Include some padding for over rounding (since we re-use the tiles)
     tile_res = (Int32.(inclusive_sides(bounds))) .+ 2
     contrib_sum = fill(RGBSpectrum(), tile_res[2], tile_res[1])
     filter_weight_sum = fill(0.0f0, tile_res[2], tile_res[1])

src/integrators/sampler.jl

@@ -6,7 +6,7 @@ struct WhittedIntegrator{C<: Camera, S <: AbstractSampler} <: SamplerIntegrator
     max_depth::Int64
 end
 
-@noinline function sample_kernel_inner(i, scene, t_sampler, film, film_tile, camera, pixel, spp_sqr)
+function sample_kernel_inner(i::A, scene::B, t_sampler::C, film::D, film_tile::E, camera::F, pixel::G, spp_sqr::H) where {A, B, C, D, E, F, G, H}
     while has_next_sample(t_sampler)
         camera_sample = get_camera_sample(t_sampler, pixel)
         ray, ω = generate_ray_differential(camera, camera_sample)
@@ -57,7 +57,7 @@ function (i::SamplerIntegrator)(scene::Scene)
     _tb_max = min.(_tb_min .+ (tile_size - 1), sample_bounds.p_max)
     _tile_bounds = Bounds2(_tb_min, _tb_max)
     filmtiles = [FilmTile(film, _tile_bounds, filter_radius) for _ in 1:Threads.maxthreadid()]
-    Threads.@threads for k in 0:total_tiles
+    Threads.@threads :greedy for k in 0:total_tiles
         x, y = k % width, k ÷ width
         tile = Point2f(x, y)
         tb_min = sample_bounds.p_min .+ tile .* tile_size