Abstract:

This live stream documents a significant refactoring effort within the Hazel game engine's rendering system, designated "Renderer 2025." The core objective is to transition Hazel from its custom Vulkan backend to NVIDIA's NVRHI, a graphics abstraction library that supports Vulkan, DirectX 11, and DirectX 12. The speaker systematically ports key rendering components, including index and vertex buffers, graphics pipelines, images, frame buffers, and textures, replacing Vulkan-specific implementations with NVRHI interfaces.

Challenges encountered include adapting Hazel's internal API to NVRHI's design, re-evaluating memory management (e.g., intrusive reference counting), and addressing specific rendering features like frame buffer load/store operations which require further investigation within NVRHI's model. Significant progress is made in reducing compilation errors as several Vulkan-specific modules are successfully ported and subsequently removed from the project. Future work identified includes a complete rewrite of the ImGui rendering backend and the compute shader infrastructure to align with the new NVRHI architecture.

Hazel Engine - Live Stream: Porting Renderer to NVRI (Renderer 2025)

• 0:00:44 New Renderer Initiative (Renderer 2025): The stream focuses on the ongoing development of Hazel's new renderer, "Renderer 2025," a critical step towards making the Hazel engine freely available.
• 0:03:34 Transition to NVRI: Hazel is migrating its rendering backend to NVIDIA's NVRHI library, a graphics abstraction layer for Vulkan, DirectX 11, and DirectX 12. This aims to simplify multi-API support and outsource backend implementation.
• 0:06:11 Simplified API Architecture: The new design routes Hazel's rendering API through NVRI, which then interfaces with specific graphics APIs, allowing for easier integration of future APIs like Metal.
• 0:08:44 Renderer 2025 Key Goals: Objectives include full NVRI support (initially Vulkan, then D3D12), shader cross-compilation (Spir-V, DXIL, DXBC), and leveraging NVRI's automatic resource tracking.
• 0:09:29 Shader Language Consideration: The speaker discusses potentially switching from GLSL to HLSL as the primary shading language for better cross-API compatibility in the new rendering pipeline.
• 0:14:22 Initial Porting & Debugging (Buffers): The process begins by addressing 1658 compilation errors. Focus is on IndexBuffer.cpp and VertexBuffer.cpp, making constructor access public to simplify Ref class usage, which acts as an intrusive reference counting system.
• 0:40:56 Pipeline Refactoring: The Vulkan-specific VulkanPipeline.cpp implementation is successfully ported and merged into the generic Pipeline.cpp class, demonstrating the benefits of NVRI's abstraction. VulkanPipeline.cpp is then deleted.
• 0:49:28 Image Component Porting: VulkanImage2D.cpp's logic is moved into Image.cpp, replacing Vulkan image and view creation with NVRI texture and TextureSubresourceSet equivalents. VulkanImage.h/.cpp is subsequently deleted.
• 1:22:20 Challenging Frame Buffer Port: FrameBuffer.cpp is identified as a complex porting task. Vulkan frame buffer creation, attachments, and render pass logic are adapted to NVRI's FramebufferDescription and FramebufferAttachment.
• 1:29:41 Unresolved Load/Store Operations: A key concern arises regarding how NVRI handles attachment load/store operations (e.g., clearing or loading existing data) and clear values, as direct equivalents are not immediately apparent in NVRI's FramebufferAttachment structure. This is marked for future investigation.
• 1:57:57 Significant Error Reduction: After porting FrameBuffer (and deleting VulkanFrameBuffer.h/.cpp), the compilation errors are significantly reduced to 36, indicating substantial progress.
• 2:00:36 Texture Component Porting: Texture.cpp (for Texture2D and TextureCube) is ported, replacing Vulkan texture and sampler creation with NVRI equivalents. VulkanTexture.h/.cpp is deleted.
• 2:46:19 Future Work: ImGui Backend Rewrite: The existing ImGui rendering backend (VulkanImGuiLayer.cpp) is heavily tied to Vulkan and will require a complete rewrite to utilize NVRI, identified as a major upcoming task.
• 2:50:01 Future Work: Compute Pipeline & Descriptor Sets: The VulkanComputePipeline.cpp and DescriptorSetManager (Vulkan-specific) are next in line for a significant rewrite, involving mapping to NVRI's ComputePipelineDesc, BindingLayout (descriptor set layout), and BindingSet (descriptor set data).
• 3:06:16 Stream Conclusion: The stream ends with core renderer components largely ported to NVRI, demonstrating the approach's effectiveness. Key remaining tasks involve the ImGui backend, compute shader infrastructure, and a deeper understanding of NVRI's resource binding and state management.