Hardware Acceleration Systems

Hardware Acceleration Systems

Audience: Programmers

Thread Context: GPU work is submitted from the main render thread. Async compute is D3D11 immediate-dispatch today, designed for D3D12/Vulkan async queues later. DXR TLAS is rebuilt per frame inside GraphicsEngine::RenderScene().

Platform/Backend Scope: D3D11/D3D12 primary on Windows. All D3D code is behind #ifdef SPARK_PLATFORM_WINDOWS with stub fallbacks. Mesh shaders are D3D12/Vulkan only (D3D11 falls back to traditional rendering). DirectStorage uses an async-I/O fallback on all platforms (SDK not linked). DXR is gated behind SPARK_HARDWARE_RT.

Overview

SparkEngine implements GPU hardware acceleration across eight systems: GPU compute particles, GPU skinning, async compute scheduling, GPU-driven indirect rendering, DXR ray tracing, GPU clustered light culling, DirectStorage I/O, and a mesh-shader pipeline. Much of the infrastructure had been built but not wired; this work activated existing systems and added new GPU-driven capabilities.

New systems implemented

All files confirmed present 2026-06-08.

System	Header	Impl	Shaders	Tests
GPU Compute Particles	Graphics/GPUParticleSystem.h	GPUParticleSystem.cpp	ParticleSimulate/Emit/BitonicSort.hlsl	TestGPUParticleSystem.cpp (11)
GPU Skinning	Graphics/GPUSkinning.h	GPUSkinning.cpp	SkinningCS.hlsl	TestGPUSkinning.cpp (9)
Async Compute	Graphics/AsyncComputeScheduler.h	AsyncComputeScheduler.cpp	—	TestAsyncComputeScheduler.cpp (9)
GPU-Driven Renderer	Graphics/GPUDrivenRenderer.h	GPUDrivenRenderer.cpp	GPUCull.hlsl, HiZBuild.hlsl	TestGPUDrivenRenderer.cpp (12)
GPU Cluster Culling	Graphics/GPUClusterCulling.h	GPUClusterCulling.cpp	ClusterCull.hlsl	TestGPUClusterCulling.cpp (11)
DirectStorage	Engine/Streaming/DirectStorageLoader.h	DirectStorageLoader.cpp	—	TestDirectStorageLoader.cpp (11)
Mesh Shader Pipeline	Graphics/MeshShaderPipeline.h	MeshShaderPipeline.cpp	MeshletMS/AS.hlsl	TestMeshShaderPipeline.cpp (9)

Existing systems wired

• DXR — wired into GraphicsEngine::RenderScene(). Verified 2026-06-08 in SparkEngine/Source/Graphics/GraphicsEngineWindows.cpp: under #ifdef SPARK_HARDWARE_RT, the renderer rebuilds the TLAS each frame (dxr.BuildTLAS(m_dxrInstances)) and dispatches enabled RT effects (reflections, shadows, AO, GI). A HybridRT path selects RayTracingBackend::HardwareDXR when available.

Note on file layout: the original audit placed the DXR hook in GraphicsEngine.cpp. The renderer has since been split per-platform; the DXR wiring now lives in GraphicsEngineWindows.cpp (with GraphicsRenderPipelinesWindows.cpp also referencing SPARK_HARDWARE_RT). Supporting RT code: Graphics/RHI/DXRSupport.{h,cpp}, Graphics/HybridRT/HybridRTManager.cpp, and Graphics/RHI/Metal/MetalRayTracing.mm for the Metal backend.

Architecture decisions

1. All D3D11 code behind #ifdef SPARK_PLATFORM_WINDOWS with stub fallbacks.
2. GPU particles use append/consume structured buffers + bitonic sort.
3. GPU skinning stores bones as compact 4×3 matrices (48 bytes vs 64 for 4×4).
4. Async compute is D3D11 immediate-dispatch now, designed for D3D12/Vulkan async queues later.
5. GPU-driven renderer: HiZ + frustum-cull compute shader → visibility buffer → indirect draw.
6. DirectStorage exposes the full API with a transparent async-file-I/O fallback.
7. Mesh-shader pipeline includes a meshlet builder with greedy triangle packing.

Key file locations

• Compute shaders: Shaders/HLSL/Compute/
• Mesh shaders: Shaders/HLSL/MeshShaders/
• C++ systems: SparkEngine/Source/Graphics/
• DirectStorage: SparkEngine/Source/Engine/Streaming/
• Tests: Tests/ (7 files, ~72 tests)

Notes

• GPUParticleSystem.cpp exceeds the ~500-line guideline but is one cohesive unit.
• The mesh-shader pipeline is D3D12/Vulkan only; D3D11 falls back to traditional rendering.
• DirectStorage currently uses the fallback path on all platforms — verified 2026-06-08 in DirectStorageLoader.cpp: comments note "async I/O fallback since full queue integration requires a D3D12 device" and "DirectStorage is Windows-only; Linux/macOS use async I/O fallback." The SDK is not linked.

Source & Freshness

• Original entry: .claude/knowledge/hardware-acceleration-systems.md (last updated 2026-03-28).
• Verified against codebase 2026-06-08.

Status changes / verifications found during freshening:

• All 7 new-system header/impl pairs and all 7 test files confirmed present.
• DXR wiring moved file: now in GraphicsEngineWindows.cpp (the renderer was split per-platform), not GraphicsEngine.cpp. BuildTLAS per-frame call and SPARK_HARDWARE_RT gate confirmed.
• DirectStorage still on the fallback path (SDK not linked) — confirmed by in-file comments.
• Metal ray-tracing backend (MetalRayTracing.mm) now present alongside the DXR/HybridRT paths.
• No regressions; all named systems still present.

Related Pages

• Rendering and Graphics
• Performance Profiling Guide
• Performance Tips
• Benchmark Framework
• Stub & Abandoned Features