LuisaGroup/LuisaCompute
High-Performance Rendering Framework on Stream Architectures
What it solves
LuisaCompute addresses the challenge of writing high-performance, cross-platform code for graphics and general-purpose GPU computation. It eliminates the need to write separate, platform-specific shading languages (like HLSL or GLSL) by allowing developers to write device-side kernels directly in modern C++.
How it works
The framework uses an embedded Domain-Specific Language (DSL) that leverages C++ meta-programming to trace code at runtime. It constructs an intermediate representation (AST) of the user's logic, which is then compiled into native shader code for various backends including CUDA, DirectX, Metal, Vulkan, and HIP. A unified runtime layer manages resources and command scheduling, hiding the complexities of different graphics APIs.
Who it’s for
It is designed for graphics engineers, researchers, and developers building high-performance rendering engines, simulation tools, or embodied AI systems who want to maintain a single, unified codebase across different hardware platforms.
Highlights
- Unified C++ DSL: Write host and device code in the same language without custom compiler extensions.
- Cross-Platform: Supports CUDA, DirectX, Metal, Vulkan, HIP, and a native C++ fallback.
- High Performance: Designed for stream-processing hardware, powering advanced rendering and robotics applications.
- Python Frontend: Offers a Python interface for easier prototyping and integration with AI workflows.
- Automatic Differentiation: Includes support for automatic differentiation, useful for optimization and learning tasks.
Related
- Project
- Project
- Project
- Project
- Project