“Rendering Vector Displacement Mapped Surfaces in a GPU Ray Tracer” by Takahiro Harada, featured in GPU Pro 6:

---

Overview

The paper presents a method for direct ray tracing of vector displacement mapped (VDM) surfaces on the GPU, specifically using OpenCL and tested on AMD FirePro W9100 GPU. The technique avoids pre-tessellation and instead builds geometry on-the-fly, enabling high-detail rendering with low memory usage and competitive performance.

---

Key Concepts

1. Vector Displacement Mapping (VDM)

• Scalar Displacement Mapping displaces vertices along the surface normal using scalar values.
• Vector Displacement Mapping uses vector values to displace vertices in arbitrary directions, allowing complex geometry like overhangs.
• VDM provides greater modeling freedom but poses challenges for ray tracing, especially on GPUs.

---

2. Ray Tracing with VDM

Challenges

• VDM lacks directional constraints, making intersection tests expensive.
• Requires on-the-fly tessellation and displacement, followed by BVH construction for acceleration.

Solution

• Use Bounding Volume Hierarchies (BVHs) built dynamically per VD patch.
• Batch rays intersecting a VD patch to amortize tessellation and BVH build costs.

---

3. Quad BVH Construction

Structure

• Each VD patch is recursively subdivided into quads.
• Leaf nodes store:
  • Quantized AABBs (2-byte integers)
  • Texture coordinates and normals (4 bytes each)
• Nodes are stored in breadth-first order.
• Stackless traversal is used to reduce memory traffic and register pressure.

Compression

• AABBs are quantized to reduce memory footprint.
• Texture coordinates and normals are also compressed.

---

4. GPU Implementation with OpenCL

Parallelization Strategy

• Rays intersecting VD patches are grouped and sorted.
• A work group processes one VD patch at a time:
  • Computes LOD (Level of Detail)
  • Builds BVH
  • Performs ray intersection
• BVH data is stored in global memory due to LDS size limitations.

Atomic Operations

• Used to safely update hit information across work groups.
• 64-bit atomic min operations encode hit distance and other data.

---

5. Integration into Ray Tracer

Three-Level BVH Hierarchy

1. Top-Level: Stores meshes and transforms.
2. Middle-Level: Stores primitives (triangles, quads, VD patches).
3. Bottom-Level: Built on-the-fly for VD patches.

Traversal Strategy

• Top and middle-level BVHs are traversed first.
• VD patch intersections are deferred and processed in batches.

---

6. Performance and Results

Memory Usage

• Dramatically reduced compared to pre-tessellation:
  • “Party” scene: 52GB → 16MB
  • “Pumpkin” scene: 380MB → 0.12MB

Rendering Time

• Direct ray tracing is faster in most cases despite complexity.
• Pre-tessellation only faster for low-detail scenes.
• Bottom-level BVH build is the most time-consuming part.

Indirect Illumination

• Performance degrades with deeper ray bounces due to repeated BVH builds.
• Caching BVHs is suggested as a future optimization.

---

Conclusion

The proposed method enables efficient GPU ray tracing of vector displacement mapped surfaces with:

• Low memory footprint
• Competitive or superior performance
• Scalability for complex scenes

The paper emphasizes that optimizing bottom-level BVH construction and traversal is more critical than optimizing for simple primitives.

---