krnl

Cross-Platform GPU Computing Library (WebGPU / Dawn)

krnl is a lightweight, cross-platform GPU compute library designed on top of Google Dawn (WebGPU) to provide a unified compute abstraction over Vulkan, Metal, and DirectX. It exposes a clean, developer-friendly API for tensors, kernels, and multi-backend execution while providing modular buffer, pipeline, and shader management for efficient GPU compute workloads.

Status: development

---

Key Features

• Cross-backend compute: single API that targets Vulkan, Metal, and DirectX via the Dawn WebGPU backend.
• Modular resource management: reusable buffer, pipeline, and shader modules for low-friction GPU programming.
• Tensor & kernel API: high-level constructs for tensors and kernels to accelerate prototyping and experiments.
• Multi-backend execution: switch backends with minimal code changes for portability and performance testing.
• Developer-focused samples: samples/ and sandbox/ demonstrate common compute patterns and integration.

Repository Layout

• krnl/ — library headers and implementation
• shaders/ — WGSL / shader sources used by the library and samples
• samples/ — example applications demonstrating usage
• sandbox/ — experimental applications and quick tests
• external/ — third-party dependencies (Dawn and vendor projects)

Prerequisites

• Windows 10/11, macOS, or Linux (platform support depends on Dawn/tooling availability)
• CMake 3.20+
• Compiler/toolchain:
  • Windows: Visual Studio 2022 (C++ workload) or Ninja + appropriate MSVC/clang toolchain
  • macOS: Xcode or Ninja + clang
  • Linux: Ninja + clang/gcc
• Optional: Python for helper scripts in external/dawn

Build (Windows / PowerShell)

Using the Visual Studio generator:

mkdir build; cd build
cmake -S .. -B . -G "Visual Studio 17 2022" -A x64
cmake --build . --config Release

Using Ninja (cross-platform):

cmake -S . -B build -G "Ninja"
cmake --build build --config Release

Tip: The repository ships CMakePresets.json. Use cmake --preset <preset-name> to pick a configured toolchain or build profile.

Running Samples

After a successful build, binaries appear in the build folder (e.g. build/Release/ on Windows). Try the example apps in samples/ and sandbox/:

# from repo root
.
cd build\Release
./samples.exe   # or relevant binary name for your platform

Basic Usage (library overview)

• Create a krnl::Context configured for a target backend (Dawn will select the underlying API).
• Allocate buffers via krnl::Buffer and manage staging with krnl::StagingPool when needed.
• Build compute pipelines / kernels from WGSL shaders and dispatch workloads via krnl::Pipeline.
• Use the Tensor API for higher-level data structures and kernel bindings.

See samples/ for concrete usage examples and patterns.

Development Notes

• Shader sources live in shaders/ and are usually compiled/loaded at runtime as WGSL.
• Dawn integration and tools live under external/dawn/ — consult external/dawn/tools and README files there for dependency fetching and platform-specific setup.
• The codebase aims to keep backend-specific details behind the Dawn abstraction so that switching between Vulkan/Metal/DirectX remains straightforward.

Contributing

• Open issues to discuss features or bugs.
• Prefer small, focused pull requests and open an issue first for larger changes.

---