Supported Backends

Sarek is designed to be backend-agnostic. You write your kernel once in OCaml, and Sarek compiles it to the appropriate shading language (CUDA C, OpenCL C, GLSL, MSL) or executes it natively on the CPU.

Backend	Target	Status	Requirements
CUDA	NVIDIA GPUs	Stable	NVIDIA Driver + CUDA Toolkit
OpenCL	Multi-vendor (AMD, Intel, NVIDIA)	Stable	OpenCL Runtime / ICD
Vulkan	Cross-platform (Linux, Windows, Android)	Stable	Vulkan SDK + glslangValidator
Metal	Apple Silicon & Intel Macs	Stable	macOS 10.13+
Native	CPU (Multicore)	Stable	OCaml 5.4.0+
Interpreter	CPU (Debug)	Stable	None

CUDA Backend (sarek-cuda)

Targets NVIDIA GPUs using the CUDA Driver API and NVRTC (Runtime Compilation).

• Features: Shared memory, atomics, warp intrinsics, dynamic parallelism.
• Performance: Native CUDA performance; uses JIT compilation to optimize for the specific GPU architecture.
• Setup: Requires libcuda.so and libnvrtc.so in your library path.

Build it from the repository (dune build sarek-cuda), list sarek-cuda.plugin in your libraries, and call Sarek_cuda.Cuda_plugin.init () (CUDA/PTX, the default path) and, if you want the NVRTC path, Sarek_cuda.Cuda_c_plugin.init () (CUDA/C) before Device.init (see Getting Started).

OpenCL Backend (sarek-opencl)

Targets a wide range of devices including AMD GPUs, Intel integrated graphics, and FPGAs.

• Features: Work-group barriers, local memory, math intrinsics.
• Compatibility: Tested on NVIDIA, AMD, and Intel platforms.
• Setup: Requires an OpenCL ICD loader (e.g., ocl-icd-libopencl1 on Linux).

Build it from the repository (dune build sarek-opencl), list sarek-opencl.plugin in your libraries, and call Sarek_opencl.Opencl_plugin.init () before Device.init (see Getting Started).

Vulkan Backend (sarek-vulkan)

Uses GLSL compute shaders and SPIR-V for modern cross-platform GPU support.

• Pipeline: Generates GLSL -> Compiles to SPIR-V (via glslang or shaderc) -> Executes on Vulkan.
• Features: Push constants, SSBOs (Storage Buffers), specialization constants.
• Platform: Ideal for modern Linux desktops, Android devices, and Windows.

Build it from the repository (dune build sarek-vulkan), list sarek-vulkan.plugin in your libraries, and call Sarek_vulkan.Vulkan_plugin.init () before Device.init (see Getting Started).

Metal Backend (sarek-metal)

Native support for Apple hardware (M1/M2/M3 chips) using Metal Shading Language (MSL).

• Features: Threadgroup memory, SIMD-group functions.
• Limitations: No double precision (float64) support on most Apple hardware.
• Platform: macOS and iOS only.

Build it from the repository (dune build sarek-metal), list sarek-metal.plugin in your libraries, and call Sarek_metal.Metal_plugin.init () before Device.init (see Getting Started).

Native CPU Backend (sarek.native)

Executes kernels directly on the host CPU without GPU compilation.

• Mechanism: Uses OCaml 5 Domains for parallel execution.
• Use Case: High-performance fallback when no GPU is available, or for debugging logic with standard debuggers.

Interpreter Backend (sarek.interpreter)

Walks through the Sarek IR (Intermediate Representation) step-by-step.

• Use Case: Deep debugging of kernel logic, verification of IR transformations, and educational purposes.
• Performance: Slow (interpreted), but provides full visibility into execution.

WebGPU / WGSL Codegen Target

Sarek includes a WGSL (WebGPU Shading Language) code-generation backend that emits compute shaders for browser-side execution via the WebGPU API.

• Output: WGSL source (@compute @workgroup_size(...) entry points)
• Deployment: Runs in any modern browser with WebGPU support (Chrome/Chromium 113+)
• Note: WGSL is a transpiler target, not a runtime device plugin — it does not appear in the device enumeration table above. Use it to ship kernel logic to the browser, combined with a JavaScript WebGPU host.
• Try it: The live Playground lets you transpile any Sarek kernel to WGSL instantly, and the Learn course runs kernels directly on your GPU via WebGPU.