JuliaGeometry · SimonDanisch · Dec 22, 2025 · Dec 23, 2025 · Dec 24, 2025 · Jan 1, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -25,6 +25,8 @@ jobs:
           version: 1
           arch: x64
       - uses: julia-actions/cache@v2
-      - name: Install pkgs dependencies
-        run: julia --project=@. -e 'using Pkg; Pkg.test("Raycore", coverage=true)'
       - uses: julia-actions/julia-runtest@v1
+        with:
+          # --check-bounds=auto: GPU kernels (OpenCL/pocl) crash with --check-bounds=yes
+          # because bounds-check code generates SPIR-V that pocl cannot handle
+          check_bounds: auto
diff --git a/Project.toml b/Project.toml
@@ -4,6 +4,10 @@ version = "0.1.1"
 authors = ["Anton Smirnov <tonysmn97@gmail.com>", "Simon Danisch <sdanisch@protonmail.com"]
 
 [deps]
+AcceleratedKernels = "6a4ca0a5-0e36-4168-a932-d9be78d558f1"
+Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+Atomix = "a9b6321e-bd34-4604-b9c9-b65b8de01458"
+GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
 GeometryBasics = "5c1252a2-5f33-56bf-86c9-59e7332b4326"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
@@ -12,28 +16,39 @@ StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 
 [weakdeps]
+Lava = "3a680b1f-cb25-4bee-9cf7-bc880b76dc8c"
 Makie = "ee78f7c6-11fb-53f2-987a-cfe4a2b5a57a"
 
 [extensions]
+RaycoreLavaExt = "Lava"
 RaycoreMakieExt = "Makie"
 
 [compat]
-julia = "1.10"
+AcceleratedKernels = "0.4.3"
+Adapt = "4.4.0"
+Aqua = "0.8"
+Atomix = "1.1.2"
+GPUArraysCore = "0.2.0"
 GeometryBasics = "0.5"
+JET = "0.11"
 KernelAbstractions = "0.9, 0.10"
 LinearAlgebra = "1"
+Makie = "0.24"
+OpenCL = "0.10.8"
 Random = "1"
 StaticArrays = "1.9.7"
 Statistics = "1"
-Makie = "0.24"
-Aqua = "0.8"
-JET = "0.11"
 Test = "1"
+julia = "1.10"
+pocl_jll = "7.1.0"
 
 [extras]
+Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"
 JET = "c3a54625-cd67-489e-a8e7-0a5a0ff4e31b"
+OpenCL = "08131aa3-fb12-5dee-8b74-c09406e224a2"
+StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
-Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"
+pocl_jll = "627d6b7a-bbe6-5189-83e7-98cc0a5aeadd"
 
 [targets]
-test = ["Test", "JET", "Aqua"]
+test = ["Test", "JET", "Aqua", "pocl_jll", "OpenCL", "StaticArrays"]
diff --git a/README.md b/README.md
@@ -4,12 +4,14 @@
 [![](https://img.shields.io/badge/docs-stable-blue.svg)](https://juliageometry.github.io/Raycore.jl/stable/)
 [![](https://img.shields.io/badge/docs-dev-blue.svg)](https://juliageometry.github.io/Raycore.jl/dev/)
 
-High-performance ray-triangle intersection engine with BVH acceleration for CPU and GPU.
+High-performance ray-triangle intersection engine with TLAS/BLAS acceleration for CPU and GPU.
 
 ## Features
 
-- **Fast BVH acceleration** for ray-triangle intersection
+- **Fast TLAS/BLAS acceleration** for ray-triangle intersection
 - **CPU and GPU support** via KernelAbstractions.jl
+- **MultiTypeSet**: GPU-safe heterogeneous collections with compile-time type-stable dispatch for materials, textures, lights, etc.
+- **GPU TLAS**: Two-level acceleration structure (BLAS/TLAS) with instanced geometry, per-instance transforms, and GPU-first design
 - **Analysis tools**: centroid calculation, illumination analysis, view factors for radiosity
 - **Makie integration** for visualization
 
@@ -26,14 +28,14 @@ Pkg.add(url="https://github.com/JuliaGeometry/Raycore.jl")
 using Raycore, GeometryBasics, LinearAlgebra
 
 # Create geometry
-sphere = Tesselation(Sphere(Point3f(0, 0, 2), 1.0f0), 20)
+mesh = normal_mesh(Sphere(Point3f(0, 0, 2), 1.0f0))
 
-# Build BVH acceleration structure
-bvh = BVH([sphere])
+# Build TLAS acceleration structure
+tlas = TLAS([mesh], (mi, ti) -> UInt32(mi))
 
 # Cast rays and find intersections
 ray = Ray(o=Point3f(0, 0, 0), d=Vec3f(0, 0, 1))
-hit_found, triangle, distance, bary_coords = closest_hit(bvh, ray)
+hit_found, triangle, distance, bary_coords, instance_id = closest_hit(tlas, ray)
 
 if hit_found
     hit_point = ray.o + ray.d * distance
@@ -46,13 +48,22 @@ end
 ```julia
 # Calculate scene centroid from a viewing direction
 viewdir = normalize(Vec3f(0, 0, -1))
-hitpoints, centroid = get_centroid(bvh, viewdir)
+hitpoints, centroid = get_centroid(tlas, viewdir)
 
 # Analyze illumination
-illumination = get_illumination(bvh, viewdir)
+illumination = get_illumination(tlas, viewdir)
 
 # Compute view factors for radiosity
-vf_matrix = view_factors(bvh; rays_per_triangle=1000)
+vf_matrix = view_factors(tlas; rays_per_triangle=1000)
+```
+
+## Testing
+
+Run tests with `--check-bounds=auto` (not the `Pkg.test` default of `--check-bounds=yes`), because GPU kernels compiled with bounds checking generate SPIR-V that crashes pocl:
+
+```julia
+using Pkg
+Pkg.test("Raycore"; julia_args=`--check-bounds=auto`)
 ```
 
 ## Documentation

diff --git a/benchmarks/implicitbvh_comparison.md b/benchmarks/implicitbvh_comparison.md
@@ -0,0 +1,65 @@
+# ImplicitBVH.jl vs Raycore.jl — GPU Benchmark
+
+**Date**: 2026-03-29
+**GPU**: AMD RX 7900 XTX (RDNA3)
+**Backend**: AMDGPU.jl (ROCArray)
+**Mesh**: xyzrgb_dragon.obj (249,882 triangles) + procedural random geometry
+
+## BVH Build
+
+| Triangles | ImplicitBVH | Raycore | Ratio |
+|-----------|-------------|---------|-------|
+| 250K      | 0.98 ms     | 4.93 ms | ImplicitBVH 5.0x faster |
+| 1M        | 2.25 ms     | 7.46 ms | ImplicitBVH 3.3x faster |
+| 4M        | 8.41 ms     | 16.16 ms | ImplicitBVH 1.9x faster |
+
+ImplicitBVH builds faster due to simpler construction (Morton sort + bottom-up aggregate).
+Raycore does more work: topology emission, parent pointers, leaf creation, atomic refit — all separate kernel launches.
+
+## Ray Tracing — Dragon Mesh (249K triangles)
+
+**Important**: ImplicitBVH `traverse_rays` returns bounding volume candidates (broad-phase only).
+Raycore `closest_hit` returns the actual closest triangle intersection (full narrow-phase).
+These are fundamentally different operations — ImplicitBVH does less work per ray but doesn't give a usable hit result.
+
+| Rays | ImplicitBVH (LVT) | Raycore | Speedup (Raycore) |
+|------|--------------------|---------|--------------------|
+| 100K | 4.60 ms            | 1.33 ms | 3.5x |
+| 500K | 11.06 ms           | 3.14 ms | 3.5x |
+| 1M   | 20.84 ms           | 3.00 ms | 6.9x |
+| 2M   | 41.52 ms           | 6.00 ms | 6.9x |
+| 4M   | 83.31 ms           | 5.91 ms | 14.1x |
+
+## Ray Tracing — Scaling with Triangle Count (1M rays)
+
+| Triangles | ImplicitBVH (BFS) | Raycore | Speedup (Raycore) |
+|-----------|--------------------|---------|--------------------|
+| 250K      | 43.89 ms           | 8.99 ms | 4.9x |
+| 1M        | 217.37 ms          | 11.08 ms | 19.6x |
+| 4M        | 313.0 ms           | 15.41 ms | 20.3x |
+
+## Why Raycore Is Faster for Ray Tracing
+
+| Factor | ImplicitBVH | Raycore |
+|--------|-------------|---------|
+| Output | All BV candidates (variable-size list) | Single closest hit (fixed) |
+| Triangle test | None (BSphere overlap only) | Moller-Trumbore per leaf |
+| Passes | Two-pass (count + write) | Single-pass |
+| Early termination | No — finds all overlaps | Yes — t_max shrinks on hit |
+| Node layout | Implicit tree + skip array | Inline leaves (BVH2IL) |
+| Allocations | Output buffer per trace | None |
+
+## What ImplicitBVH Does Better
+
+- **Build speed** (2-5x faster) — fewer kernel launches, implicit indexing
+- **Collision detection** — LVT (leaf-vs-tree) two-pass is designed for finding all contact pairs
+- **Two-BVH collision** — native support for inter-object contact detection
+- **Cache reuse** — `BVHTraversal` cache avoids re-allocation across frames
+- **Mixed bounding volumes** — BSphere leaves with BBox internal nodes
+
+## Reference: ImplicitBVH README Numbers (Nvidia A100)
+
+From the ImplicitBVH.jl README (249,882 triangles, BSphere/BBox):
+- Build: 410 us
+- Contact detection: 1.14 ms
+- 100K rays: 2.00 ms
diff --git a/docs/Project.toml b/docs/Project.toml
@@ -1,4 +1,5 @@
 [deps]
+Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
 Bonito = "824d6782-a2ef-11e9-3a09-e5662e0c26f8"
 BonitoBook = "b416d416-7a6e-4336-8c1a-1f8a8cd59518"
@@ -9,9 +10,13 @@ GeometryBasics = "5c1252a2-5f33-56bf-86c9-59e7332b4326"
 ImageShow = "4e3cecfd-b093-5904-9786-8bbb286a6a31"
 JET = "c3a54625-cd67-489e-a8e7-0a5a0ff4e31b"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Makie = "ee78f7c6-11fb-53f2-987a-cfe4a2b5a57a"
 MeshIO = "7269a6da-0436-5bbc-96c2-40638cbb6118"
+Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Raycore = "afc56b53-c9a9-482a-a956-d1d800e05559"
+Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
+Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 WGLMakie = "276b4fcb-3e11-5398-bf8b-a0c2d153d008"
 
 [sources]

diff --git a/docs/examples.jl b/docs/examples.jl
@@ -14,8 +14,11 @@ begin
     l = 0.5
     floor = Rect3f(-l, -l, -0.01, 2l, 2l, 0.01)
     cat = load(Makie.assetpath("cat.obj"))
-    bvh = Raycore.BVH([s1, s2, s3, s4, cat]);
-    world_mesh = GeometryBasics.Mesh(bvh)
+    bvh = Raycore.TLAS([normal_mesh(s1), normal_mesh(s2), normal_mesh(s3), normal_mesh(s4), cat], (mi, ti) -> UInt32(mi));
+    # TODO: examples.jl needs rewrite for TLAS API
+    # bvh.primitives → iterate tlas.blas_array[i].primitives
+    # GeometryBasics.Mesh(bvh) → use Makie extension: plot(tlas)
+    world_mesh = Makie.convert_arguments(Makie.Mesh, bvh)[1]
     f, ax, pl = Makie.mesh(world_mesh; color=:teal)
     center!(ax.scene)
     viewdir = normalize(ax.scene.camera.view_direction[])