Sarek - GPU Computing for OCaml

SIMT Abstraction for Runtime Extensible Kernels

Sarek is a PPX-based DSL that lets you write GPU kernels directly in OCaml syntax. Kernels compile to multiple backends (CUDA, OpenCL, Vulkan, Metal) without code changes.

What is Sarek?

Sarek is the user-facing DSL and compiler. Write kernels in OCaml with [%kernel ...], and Sarek compiles them to GPU code at build time.

SPOC (SIMT Programming for OCaml) is the underlying runtime providing device abstraction, plugin architecture, and backend infrastructure.

Recent Development

This codebase has undergone significant modernization (2024-2026):

• OCaml 5.4 support with effect handlers and domains
• Code quality improvements across all GPU backends
• Structured error handling replacing untyped exceptions
• Plugin-based architecture for extensible backend support
• Test coverage with unit and end-to-end tests
• Documentation for all major components
• WGSL/WebGPU codegen — a 5th transpiler backend emitting WGSL for browser-side execution
• In-browser Playground — live kernel transpiler at mathiasbourgoin.github.io/Sarek/playground.html
• Interactive Learn course — edit and run Sarek kernels on your GPU in the browser at mathiasbourgoin.github.io/Sarek/learn/
• PTX direct emitter (experimental) — Sarek_ir_ptx emits NVIDIA PTX directly from Sarek IR, bypassing NVRTC; validated on real hardware; foundation for formal backend verification

The framework is actively maintained and uses modern OCaml features while preserving compatibility with existing SPOC code.

Note: This recent rework was completed with assistance from AI agents. Feedback, bug reports, and contributions are welcome via GitHub Issues.

Features

GPU Kernel Development

Write GPU kernels in OCaml syntax using the [%kernel ...] PPX extension:

let vector_add =
  [%kernel
    fun (a : float32 vector) (b : float32 vector) (c : float32 vector) (n : int32) ->
      let open Sarek_stdlib.Std in
      let tid = global_thread_id in
      if tid < n then c.(tid) <- a.(tid) + b.(tid)]

Kernels compile to multiple backends automatically without code changes.

Backend Support

Backend	Target	Status	Documentation
CUDA	NVIDIA GPUs	✓	sarek-cuda/
OpenCL	Multi-vendor GPUs/CPUs	✓	sarek-opencl/
Vulkan	Cross-platform GPUs	✓	sarek-vulkan/
Metal	Apple Silicon/Intel Macs	✓	sarek-metal/
Native	CPU (parallel)	✓	sarek/plugins/native/
Interpreter	CPU (debugging)	✓	sarek/plugins/interpreter/
PTX (direct)	NVIDIA GPUs	⚗️ Experimental	sarek/codegen/Sarek_ir_ptx.ml

Core Features

• Type Safety: GADTs and phantom types for compile-time guarantees
• Zero-Copy: Efficient memory sharing between host and device
• Automatic Selection: Runtime backend selection based on available hardware
• Intrinsics: Extensive library of GPU intrinsics (math, atomics, barriers)
• Custom Types: Support for records and variants in kernels
• Debug Logging: Controlled via SAREK_DEBUG environment variable

Framework Architecture

spoc/              Low-level SDK and plugin interface
├── framework/     Plugin registration and backend interface
├── ir/            Intermediate representation (IR)
└── registry/      Intrinsic function registry

sarek/             Runtime and PPX compiler
├── core/          Device abstraction and memory management
├── framework/     Framework integration
├── ppx/           Sarek PPX compiler
├── sarek/         Unified execution dispatcher
└── plugins/       Native and Interpreter backends

GPU Backends:
├── sarek-cuda/    NVIDIA CUDA backend
├── sarek-opencl/  OpenCL backend (multi-vendor)
├── sarek-vulkan/  Vulkan/GLSL backend
└── sarek-metal/   Apple Metal backend

Experimental:
└── sarek/codegen/Sarek_ir_ptx.ml   Direct PTX emitter (experimental)

Experimental Features

PTX Direct Emitter (Sarek_ir_ptx)

⚠️ Experimental — validated on real hardware, not yet exercised by CI on GPUs.

Sarek_ir_ptx emits NVIDIA PTX directly from Sarek IR, bypassing NVRTC entirely. It is the default device path of the CUDA backend (Cuda_ptx_plugin); the NVRTC/C path remains available as Cuda_c_plugin.

What works:

• Scalar and vector kernels (int32/int64, float32/float64), global loads/stores, barriers
• Records and variants with the aligned C-ABI aggregate layout (proven in formal/codegen-ptx/theories/PtxLayout.v), match expressions, static tag erasure
• Helper functions via EApp inlining (sarek.inline budget-controlled)
• Static and dynamic shared memory (.shared, module-scope extern .shared), per-thread .local arrays
• Atomics: add/min/max/and/or/xor/exch, CAS, wrapping inc/dec, 64-bit add/exch
• float64 softmath library (trig, exp/log family, hypot, fma, …) with an interpreter oracle
• Parameterised SM target (?sm_target, default sm_86); Cuda_api.Kernel.load_from_ptx adapts .target to the device's actual SM (tested: GTX 1070, sm_61; AMD RX 7900 XTX via ZLUDA)

Known gaps:

• The formal proofs cover the aggregate byte layout and a scalar statement/expression fragment of a Rocq-side model; the production emitter is conformance-tested against that model, not extracted from it
• No ptxas validation gate or GPU execution in CI yet — GPU validation is manual
• Warp-level primitives are modeled in the IR but not emitted by any backend yet

Intended purpose: foundation for formal verification of the CUDA backend, developed in formal/codegen-ptx/ alongside specs/ptx-records-variants.md and specs/ptx-dshared-formal.md.

Installation

Prerequisites

• OCaml 5.4.0+ (local opam switch included in repository)
• dune 3.15+
• GPU backends (optional):
  • CUDA: NVIDIA driver + CUDA toolkit (see CUDA requirements below)
  • OpenCL: OpenCL implementation for your device
  • Vulkan: Vulkan SDK + glslangValidator or Shaderc
  • Metal: macOS 10.13+ (included with Xcode)

The Native (CPU parallel) and Interpreter (CPU sequential) backends work without any GPU drivers.

CUDA Requirements

For NVIDIA GPUs, especially newer architectures:

• CUDA Toolkit: 12.9 or later recommended
• Driver Version:
  • CUDA 12.9 requires driver 575+
  • CUDA 13.1+ requires driver 580+
• Blackwell GPUs (RTX 5000 series, compute capability 12.0):
  • Minimum: CUDA 12.9 + driver 575
  • Recommended: CUDA 13.1 + driver 580+

Note: The "CUDA Version" shown by nvidia-smi indicates the maximum CUDA runtime API version your driver supports. This may differ from your installed CUDA toolkit version, which is normal. For example, driver 575 with CUDA toolkit 12.9 will show "CUDA Version: 12.9" in nvidia-smi.

AMD GPUs via ZLUDA

The CUDA backend also runs on AMD GPUs through ZLUDA, a CUDA implementation on top of ROCm. ZLUDA ships the CUDA driver API but not NVRTC, so only the CUDA/PTX backend (the default) is available. Records, variants, match expressions and shared memory are all supported by the PTX emitter, so typical Sarek kernels run unmodified; only kernels that explicitly select the CUDA/C (NVRTC) backend will not run.

# Prerequisites: ROCm (tested with 7.2) and a supported AMD GPU (e.g. RDNA3)
# Download a ZLUDA release and point the dynamic loader at it:
LD_LIBRARY_PATH=/path/to/zluda dune exec -- sarek-device-info
# → AMD Radeon RX 7900 XTX [ZLUDA] (CUDA/PTX)

Tested on an RX 7900 XTX with ZLUDA v7-preview.3: the CUDA/PTX backend matches or exceeds the OpenCL and Vulkan backends on memory-bound benchmarks.

Installing via OPAM

SPOC is not yet published to the OPAM repository, but you can use OPAM to install from source with all dependencies:

# Clone repository
git clone https://github.com/mathiasbourgoin/Sarek.git
cd Sarek

# Install dependencies via OPAM (OCaml 5.4+)
opam update
opam install . --deps-only --working-dir

# Build all backends
dune build

# Or build only specific backends you need
dune build sarek sarek-cuda
dune build sarek sarek-opencl

Backends detect compatible drivers at runtime. You can install backends even without corresponding GPU drivers - they will simply not be available for use.

Building from Source

# Clone and use local opam switch
cd SPOC
opam install . --deps-only

# Build all packages
dune build

# Build specific backend
dune build sarek-cuda
dune build sarek-opencl

The framework uses dynamic linking, so you can build without GPU drivers installed. GPU support is detected at runtime.

Verifying Installation

# List all available devices
dune exec -- sarek-device-info

# Run unit tests
dune runtest

# Run fast benchmarks (Native + OpenCL if available)
make benchmarks-fast

# Run full benchmark suite on all available devices
make benchmarks

The fast benchmarks use small problem sizes and complete in ~20 seconds, while the full benchmark suite exercises all backends with larger datasets.

Benchmark Suite: 6 comprehensive benchmarks covering compute-bound (matrix multiplication, Mandelbrot), memory-bound (vector addition, reduction), and optimization patterns (transpose naive vs tiled). Results are published to an interactive web viewer with multiple visualization modes.

Usage

Basic Example

open Sarek
module Device = Spoc_core.Device
module Vector = Spoc_core.Vector

(* Define a kernel *)
let saxpy =
  [%kernel
    fun (a : float32 vector) (x : float32 vector)
        (y : float32 vector) (alpha : float32) (n : int32) ->
      let open Sarek_stdlib.Std in
      let i = global_thread_id in
      if i < n then y.(i) <- alpha *. x.(i) +. a.(i)]

let () =
  (* Initialize framework *)
  let devs = Device.init ~frameworks:["CUDA"; "OpenCL"; "Native"; "Interpreter"] () in
  let dev = devs.(0) in

  (* Get IR from kernel *)
  let _, kirc = saxpy in
  let ir = match kirc.Sarek.Kirc_types.body_ir with
    | Some ir -> ir | None -> failwith "No IR" in

  (* Create vectors *)
  let n = 1024 in
  let a = Vector.create Vector.float32 n in
  let x = Vector.create Vector.float32 n in
  let y = Vector.create Vector.float32 n in

  (* Execute kernel *)
  let block = Execute.dims1d 256 in
  let grid  = Execute.dims1d ((n + 255) / 256) in
  Execute.run_vectors ~device:dev ~ir ~args:[Vec a; Vec x; Vec y; Float 2.5; Int n]
    ~block ~grid ()

Backend Selection

(* List available devices *)
let devices = Device.all () in
Array.iter (fun dev ->
  Printf.printf "%s (%s)\n"
    dev.Device.name
    dev.Device.framework
) devices

(* Select specific backend. CUDA backends register as "CUDA/PTX" and
   "CUDA/C"; use filter_cuda to match the whole family. *)
let cuda_devices = Device.filter_cuda () in
let opencl_devices = Device.by_framework "OpenCL" in

See sarek/sarek/README.md for comprehensive usage documentation.

Testing

# Run all tests
dune runtest

# Run specific backend tests
dune test sarek-cuda
dune test sarek-opencl

# Run with specific backend
SAREK_BACKEND=cuda dune runtest

See COVERAGE.md for coverage measurement instructions.

Troubleshooting

CUDA Issues

Error: CUDA_ERROR_UNKNOWN(222) when loading PTX on new GPUs

This error typically occurs on newer GPU architectures (e.g., Blackwell/RTX 5000 series) with mismatched CUDA versions:

• Solution: Ensure you have CUDA 12.9+ installed with driver 575+
• Check versions:

  nvidia-smi                    # Shows driver version and API level
  nvcc --version                # Shows installed CUDA toolkit version

• Common cause: CUDA 13.1 requires driver 580+. If you have driver 575, use CUDA 12.9 instead.

PTX compilation succeeds but module loading fails

Sarek automatically handles forward compatibility by compiling PTX for compute_90 on compute capability 9.0+ devices. The CUDA driver then JIT-compiles for your actual hardware (e.g., sm_120 for RTX 5070 Ti). This requires:

• CUDA toolkit 12.9+ (for Blackwell GPU support)
• Compatible driver version (see requirements above)

Verifying CUDA setup

# Check if CUDA devices are detected
nvidia-smi

# Verify Sarek can find devices
dune exec -- sarek-device-info

# Check driver API compatibility
cat /proc/driver/nvidia/version

OpenCL Issues

If OpenCL is not detecting your device, ensure you have the appropriate ICD (Installable Client Driver) installed:

• NVIDIA: Install NVIDIA driver with OpenCL support
• AMD: Install ROCm or AMDGPU-PRO driver
• Intel: Install Intel OpenCL runtime

Documentation

• GitHub Pages - User guides, tutorials, and API docs
• CONTRIBUTING.md - Contribution guidelines
• CHANGES.md - Changelog
• Backend Documentation - Individual backend READMEs

For API documentation, see inline comments and README files in each package directory.

Contributing to Documentation

Documentation sources are in gh-pages/ directory:

• User guides: gh-pages/docs/*.md
• Jekyll layouts: gh-pages/_layouts/
• API docs: Auto-generated from code comments via odoc

Changes merged to main branch automatically deploy to GitHub Pages via CI.

Requirements

• OCaml: 5.4.0+ (uses domains, effects)
• System: 64-bit Linux, macOS, Windows (limited testing)
• GPU: Optional - Native and Interpreter backends work on any system

Project History

This work originates from Mathias Bourgoin's PhD thesis at UPMC-LIP6 laboratory (Paris) and was partially funded by the OpenGPU project. Development continued at Verimag laboratory (Grenoble, 2014-2015) and LIFO laboratory (Orléans, 2015-2018).

Current maintainer: Mathias Bourgoin (Nomadic Labs)

License

See LICENSE.md for license information.

Resources

• GitHub Pages: http://mathiasbourgoin.github.io/Sarek/
• GitHub Actions: Build status and CI
• Issues: Bug reports and feature requests