merge

Author: SimonDanisch

docs/code/basic-scene.jl

@@ -72,34 +72,18 @@ end
 
 begin
     integrator = Trace.WhittedIntegrator(cam, Trace.UniformSampler(8), 5)
-    @time integrator(scene, film)
+    @btime integrator(scene, film)
     img = reverse(film.framebuffer, dims=1)
 end
+
 # 6.296157 seconds (17.64 k allocations: 19.796 MiB, 0.13% gc time, 45 lock conflicts)
 # After more GPU optimizations
 # 4.169616 seconds (17.37 k allocations: 19.777 MiB, 0.14% gc time, 20 lock conflicts)
 # After first shading running on GPU
 # 3.835527 seconds (17.36 k allocations: 19.779 MiB, 0.16% gc time, 41 lock conflicts)
+# 4.191 s (4710 allocations: 18.36 MiB)
+# iterative_li: 5.2s -.-
 
-camera_sample = Trace.get_camera_sample(integrator.sampler, Point2f(512))
-ray, ω = Trace.generate_ray_differential(integrator.camera, camera_sample)
-
-@btime Trace.intersect_p(bvh, ray)
-@btime Trace.intersect!(bvh, ray)
-
-###
-# Int32 always
-# 42.000 μs (1 allocation: 624 bytes)
-# Tuple instead of vector for nodes_to_visit
-# 43.400 μs (1 allocation: 624 bytes)
-# AFTER GPU rework
-# intersect!
-# 40.500 μs (1 allocation: 368 bytes)
-# intersect_p
-# 11.500 μs (0 allocations: 0 bytes)
-
-### LinearBVHLeaf as one type
-# 5.247460 seconds (17.55 k allocations: 19.783 MiB, 46 lock conflicts)
 
 # begin
 #     integrator = Trace.SPPMIntegrator(cam, 0.075f0, 5, 1)

src/accel/bvh.jl

@@ -251,7 +251,8 @@ end
             if !ln.is_interior && ln.n_primitives > Int32(0)
                 # Intersect ray with primitives in node.
                 for i in Int32(0):ln.n_primitives - Int32(1)
-                    tmp_primitive = primitives[ln.offset+i]
+                    offset = ln.offset % Int32
+                    tmp_primitive = primitives[offset+i]
                     tmp_hit, ray, tmp_interaction = intersect_p!(
                         tmp_primitive, ray,
                     )
@@ -293,13 +294,15 @@ end
 
     to_visit_offset, current_node_i = Int32(1), Int32(1)
     nodes_to_visit = zeros(MVector{64,Int32})
+    primitives = bvh.primitives
     @inbounds while true
         ln = bvh.nodes[current_node_i]
         if intersect_p(ln.bounds, ray, inv_dir, dir_is_neg)
             if !ln.is_interior && ln.n_primitives > Int32(0)
                 for i in Int32(0):ln.n_primitives-Int32(1)
+                    offset = ln.offset % Int32
                     intersect_p(
-                        bvh.primitives[ln.offset + i], ray,
+                        primitives[offset + i], ray,
                     ) && return true
                 end
                 to_visit_offset == 1 && break

src/gpu-support.jl

@@ -33,13 +33,14 @@ function to_gpu(ArrayType, m::Trace.UberMaterial; preserve=[])
         no_tex_s = typeof(Kr)()
         Kd = Trace.no_texture(m.Kd) ? no_tex_s : to_gpu(ArrayType, m.Kd; preserve=preserve)
     end
-    f_tex = to_gpu(ArrayType, Trace.Texture(ArrayType(zeros(Float32, 1, 1))); preserve=preserve)
+    f_tex = to_gpu(ArrayType, Trace.Texture(zeros(Float32, 1, 1)); preserve=preserve)
     no_tex_f = typeof(f_tex)()
     return Trace.UberMaterial(
         Kd,
         Trace.no_texture(m.Ks) ? no_tex_s : to_gpu(ArrayType, m.Ks; preserve=preserve),
         Trace.no_texture(m.Kr) ? no_tex_s : to_gpu(ArrayType, m.Kr; preserve=preserve),
         Trace.no_texture(m.Kt) ? no_tex_s : to_gpu(ArrayType, m.Kt; preserve=preserve),
+
         Trace.no_texture(m.σ) ? no_tex_f : to_gpu(ArrayType, m.σ; preserve=preserve),
         Trace.no_texture(m.roughness) ? no_tex_f : to_gpu(ArrayType, m.roughness; preserve=preserve),
         Trace.no_texture(m.u_roughness) ? no_tex_f : to_gpu(ArrayType, m.u_roughness; preserve=preserve),