Takahiro HARADA

“HIPRT: A Ray Tracing Framework in HIP” by Daniel Meister, Paritosh Kulkarni, Aaryaman Vasishta, and Takahiro Harada (AMD, 2024):

1. Motivation & Background

Ray tracing has become increasingly feasible on GPUs thanks to modern hardware with large memory and dedicated acceleration units. While NVIDIA (OptiX) and Intel (Embree) provide professional-grade GPU ray tracing frameworks, AMD lacked an equivalent open-source framework tailored for professional rendering pipelines.

HIPRT is introduced as a cross-platform, GPU-based ray tracing framework in HIP (Heterogeneous-Compute Interface for Portability) designed specifically for AMD GPUs, but usable in broader environments. Unlike Vulkan/DirectX ray tracing APIs, HIPRT:

Provides ray tracing only, leaving shading and material evaluation to applications.
Targets professional offline rendering (e.g., Blender Cycles, PBRT-v4, Radeon ProRender).
Offers a lightweight, user-friendly API.

2. Core Contributions

Full-featured GPU ray tracing framework for AMD hardware, with support for:
- Multi-level instancing
- Motion blur with non-uniform time intervals
- Custom primitives & intersection filters
Novel massively-parallel SBVH (spatial split BVH) construction algorithm adapted to GPUs.
Flexible and minimal API design, avoiding complexities of shader binding tables.
Open-source release with ports of PBRT-v4 demonstrating integration.
Technical innovations:
- Component-wise motion blur representation (avoiding matrix singularities found in OptiX).
- Generic traversal stack supporting multiple strategies (private, global, dynamic).
- Support for importing custom BVHs (enabling research comparisons).

3. HIPRT API

Acceleration structures:
- BLAS (Geometry) – triangles or custom primitives.
- TLAS (Scene) – instancing with affine transformations & motion blur.
Build modes: Fast (LBVH), Balanced (PLOC), High Quality (SBVH).
User workflow:
- Create geometry/scene.
- Build BVH.
- Compile trace kernel.
Custom functions: user-defined intersection/filter functions indexed by (ray type × geometry type).

4. BVH Construction

HIPRT provides three GPU-based BVH builders:

LBVH (Linear BVH): Fastest, Morton code + radix sort.
PLOC (Parallel Locally Ordered Clustering): Balanced speed/quality.
SBVH (Spatial Split BVH): High-quality GPU adaptation of Stich et al.’s algorithm.

Key techniques:

Triangle pairing: merges neighboring triangles into pairs → ~30% fewer leaves.
Conversion to wide BVH (4-way): Optimized top-down pass for AMD hardware units.
Multi-level instancing: supports massive scenes (e.g., Disney’s Moana Island with 31B instances in-core).
Batch construction: efficient build for many small geometries (e.g., hair strands).

5. Ray Traversal

Hardware-accelerated ray-box and ray-triangle intersection (RDNA2/3).
General traversal loop supports multiple node types (internal, leaf, instance).
Traversal stack implementations:
- Private stack: simple, local memory.
- Global stack: shared+global memory, high efficiency.
- Dynamic stack: shared+global, allocated adaptively to save memory.
Custom function dispatch via compile-time generated function tables (avoids costly runtime indirections).
Supports multi-level hierarchies and motion blur with correct interpolation.

6. Technical Implementation

~17k lines of C++/HIP, leveraging templates and compile-time specialization.
Cross-vendor portability via Orochi library (AMD/NVIDIA).
Compiles into HIP binaries (no IR like SPIR-V/PTX, so architecture-specific builds required).
Fully open-sourced to encourage research/industry adoption.

7. Evaluation

Tested on AMD Radeon Pro W7900 (48GB VRAM) against Vulkan (Fast & HQ) and Embree (CPU high-quality BVHs imported).
Benchmarks on 10 complex scenes (0.8M–8.2M triangles).
Results:
- Build times:
  - LBVH = 1.4–3.4× faster than Vulkan Fast.
  - PLOC = 1.3–4.1× faster than Vulkan HQ.
  - SBVH = slowest (3.7–8.9× slower than Vulkan HQ), but gives highest-quality BVHs.
- Trace times:
  - HIPRT PLOC competitive with Vulkan; SBVH & Embree yield best performance.
  - Multi-level instancing shows HIPRT outperforming Vulkan (e.g., up to 1.8× faster for shadow rays).
- SAH (Surface Area Heuristic) cost:
  - SBVH ≈ Embree (lowest cost, best quality).
  - LBVH/PLOC higher cost but faster build.
Demonstrates trade-off between build speed vs. trace efficiency.

8. Conclusion & Impact

HIPRT provides:

A practical, research-friendly, and industry-ready GPU ray tracing framework for AMD GPUs.
Bridges a gap similar to Intel’s Embree and NVIDIA’s OptiX but tailored for HIP.
Open-source reference for ray tracing community.
Demonstrated integration in production renderers (Blender, PBRT-v4, Radeon ProRender).

The framework balances simplicity (API design), performance (optimized BVH build/traversal), and flexibility (custom primitives, motion blur, instancing), making it a strong candidate for both academic research and production rendering pipelines.