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, HIP, OpenCL, Vulkan, Metal) without code changes, plus WGSL for the browser and two CPU backends.

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

• HIP backend (sarek-hip) — native ROCm/hiprtc backend for AMD GPUs, not going through ZLUDA. Measured against OpenCL and Vulkan on an RX 7900 XTX in docs/benchmarks/hip-vs-opencl-vulkan-2026-07-24.md

• float16 element type — float16 vectors and kernel values, with compute-in-f32 semantics and explicit narrowing. Codegen exists on the CUDA-C / HIP path only (both go through Sarek_ir_cuda), and the Interpreter also runs f16: it lists Float16 in Sarek_interp_capability.device_features and rounds through IEEE binary16 in Sarek_float16. The remaining targets refuse it — but not all for the same kind of reason, and the difference matters:

  • refused by measurement — OpenCL (rusticl/radeonsi) and GLSL/Vulkan (RADV's ACO): the driver absorbs the f32→f16 narrowing into the arithmetic feeding it, so a measured fraction of binary16 inputs disagrees with the interpreter — 620/63488 on OpenCL, 2912/63488 on GLSL — and the available remedy does not work (no affordable barrier on the OpenCL path; precise does not prevent it on GLSL). Devices, counts and reproducers in docs/fp-contraction-policy.md
  • refused by an emitter-internal invariant — PTX: Sarek_ir_ptx_types derives a value's register class from the register name prefix (%f/%fd/%rd), so adding a %h class requires auditing every such guard first
  • not implemented yet — Metal and WGSL: both arms raise unsupported_construct and say so ("not yet supported (#57 slice 2)"); these are deferrals, not measured refusals
  • Native does not list Float16 among its device features at all; it carries f16 vectors as a storage type only

  See docs/design/f16-dsl-element-type.md and the TFloat16 arms of each generator

• Structure-of-Arrays device layout — a record-typed vector parameter can be lowered as SoA (one coalesced device array per scalar leaf) instead of packed AoS; CUDA/PTX only, via Spoc_core.Soa_vector + Sarek.Soa_launch.run_soa. Design and measurements in docs/optimization/tier1b-emitter-soa-handoff.md

• In-tree kernel libraries, all written in pure Sarek rather than per-backend: Sarek_gemm (shared-memory tiled SGEMM; every backend with a block-shared-memory model — i.e. all but the sequential Interpreter), Sarek_worklist (dynamic parallelism over an atomic work queue, serving the use cases CUDA CDP targets without any device-side launch mechanism), Sarek_df64 / Sarek_real64 (double-float software extended precision for devices with no usable float64)

• Machine-checked formal models — 94 theorems are checked to agree with the code that actually ships, not merely with the Rocq model: 50 (the PTX layout theory) via its own extraction run directly against production, exhaustively over every record/variant shape up to a bound plus randomised beyond it; 44 (two type-safety modules) via QCheck differential testing against the real typer/unifier, sampled over 2000/1000 random cases — sampling, not exhaustive, so treat it as strong agreement evidence rather than a transported proof. scripts/check-production-link.py derives these numbers by mechanically verifying each production-link claim against its named test file's real content (a claim whose test regresses fails the gate) and reading theorem counts from the committed, drift-checked formal/*/proof-ledger.json files. Run scripts/check-production-link.py directly (no rebuild, no container, seconds not minutes) to reproduce them yourself; scripts/check-formal-proofs.sh runs the same check after a from-scratch coqchk-verified rebuild, which is the stronger guarantee but also destroys uncommitted work under formal/ on its way out, so prefer the standalone script unless you specifically need the rebuild. No gate compares this sentence's digits against either script's output, so treat the numbers above as a snapshot, not a live figure. The three Rocq projects under formal/ (convergence-safety, type-safety, codegen-ptx) additionally carry model-only proofs — machine-checked against the Rocq model itself, with no mechanical link to shipped code — listed separately in docs/formal/rocq-value-ledger.md, which explains what those do and do not cover

• Executable codegen gates in CI — the CI image ships CUDA 12.6 ptxas/nvdisasm/NVRTC, glslangValidator, naga and OpenCL-capable clang, and ci/assert-toolchain.sh fails the job if any of them is missing, so a self-skipping gate can no longer report success

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

Devices that appear in Device.init / sarek-device-info:

Not a device — a code-generation target only:

The CUDA family registers as two frameworks (CUDA/PTX and CUDA/C); use Device.filter_cuda () to match both.

The published backend page is stale against the two tables above, and is deliberately not linked from them. gh-pages/docs/backends.md — which is what mathiasbourgoin.github.io/Sarek/ serves — still has a single CUDA row with no HIP row and no CUDA/PTX-vs-CUDA/C split, and its "CUDA Backend" section says the backend targets NVIDIA GPUs "using the CUDA Driver API and NVRTC", which is the CUDA/C path, not the PTX path that is now the default. Its WGSL section (WGSL is a transpiler target, not a device plugin) does agree with the table above. Bringing the published page in line is a website change with its own audience and its own row-by-row verification, so it is filed as follow-up work rather than folded in here — but it is named here so the divergence is not silent, and so this table is not read as endorsing that page.

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). The PTX emitter lowers 83 intrinsic names natively, enumerated from its own dispatch registry (Sarek_ir_ptx_expr.intrinsic_registry)
• Custom Types: Support for records and variants in kernels, laid out with the aligned C-ABI rule shared by the host PPX, the C-family backends and the PTX emitter
• Numeric widths: int32, int64, float32, float64; float16 on the CUDA-C/HIP path and in the Interpreter (refused elsewhere — see the float16 bullet above for which refusals are measured and which are unimplemented); Sarek_df64 / Sarek_real64 for software extended precision where a device has no usable float64
• Capability refusals over silent fallback: a kernel using a width or feature a target cannot honour is refused with a stated reason and provenance rather than quietly downgraded — see docs/design/capability-model.md
• 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 (incl. Soa_vector)
├── core_base/     Vector representation and host storage
├── codegen/       All code generators: PTX, CUDA C, OpenCL C, GLSL, MSL, WGSL
├── execute/       Launch pipeline (Execute, Soa_launch)
├── framework/     Framework integration
├── interp/        IR interpreter
├── ppx/           Sarek PPX compiler
├── sarek/         Unified execution dispatcher
├── transpile/     Standalone / in-browser transpiler
├── plugins/       Native and Interpreter backends (+ inert WebGPU stub)
└── Sarek_*/       Kernel libraries: stdlib, gemm, worklist, df64, real64,
                   float64, geometry, tuple_vec

GPU Backends:
├── sarek-cuda/    NVIDIA CUDA backend (CUDA/PTX default, CUDA/C via NVRTC)
├── sarek-hip/     AMD ROCm/HIP backend
├── sarek-opencl/  OpenCL backend (multi-vendor)
├── sarek-vulkan/  Vulkan/GLSL backend
└── sarek-metal/   Apple Metal backend

Formal models (Rocq):
├── formal/convergence-safety/   Barrier-safety analysis
├── formal/type-safety/          Sarek PPX type system
└── formal/codegen-ptx/          PTX emission + aggregate byte layout

The PTX Direct Emitter (Sarek_ir_ptx)

Assembled by ptxas in CI on every run; GPU execution in CI is still absent. The emitter is no longer an experiment sitting beside the CUDA backend — it is the CUDA backend's default device path. What CI cannot do is run a kernel: no runner has a GPU, so every device e2e test skips there and device validation remains manual (RX 7900 XTX under ZLUDA, GTX 1070).

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)

How the generated PTX is validated:

• ptxas assembly, on the whole intrinsic surface. sarek/tests/unit/test_ptx_intrinsic_sweep.ml builds one kernel per registered intrinsic name and width — 83 names today — enumerated from the emitter's own dispatch registry (Sarek_ir_ptx_expr.intrinsic_registry), not from a hand-picked list, and assembles each with ptxas. Adding an intrinsic to the emitter without adding a sweep recipe fails test_every_name_has_a_recipe, so the gate cannot drift behind the emitter. sarek/tests/unit/test_ptx_snapshot.ml assembles a smaller set of regression and SoA kernels on top of its PTX-text assertions.
• The gate cannot self-skip green in CI. Both gates skip cleanly when ptxas is absent, which is right on a developer machine and was a disaster in CI: the earlier image had no CUDA at all, so the gates printed SKIP and the job passed having validated nothing. ci/Dockerfile now installs CUDA 12.6 ptxas/nvdisasm/NVRTC + headers, and ci/assert-toolchain.sh runs before any test step and fails if ptxas is missing or cannot assemble a probe module — it runs each tool rather than command -v-ing it, because a present-but-broken binary skips just as silently as an absent one.
• nvdisasm SASS inspection for the f16 f32-discipline gate (sarek-cuda/test/test_cuda_f16_sass.ml), likewise asserted present.

Known gaps:

• Proof-to-production link is uneven. The aggregate byte layout half is extracted: formal/codegen-ptx/extraction/LayoutExtract.v extracts formal/codegen-ptx/theories/PtxLayout.v to OCaml, and the conformance suite runs the theory's own definitions against Sarek_ir_layout. The expression/statement/kernel half is not: formal/codegen-ptx/test/test_codegen_ptx_conformance.ml still checks a hand-written OCaml mirror of the Rocq definitions, and the five .ml files Extract.v produces are regenerated by every proof run, are not tracked (backlog-188), and are linked by nothing. On that half the production emitter is not compared to the model at all — the mirror is compared to the mirror's own spec. The only cases in that suite that invoke production code are the 7 in the ptx-dshared group, which call the real Sarek_codegen.Sarek_ir_ptx_kernel.generate and inspect its PTX text against the shared-memory acceptance criteria of specs/ptx-dshared-formal.md — a string/exception check on the emitter's output, still not a check against the Rocq model. So: extraction links theory to production on the layout half; on the expr/stmt/kernel half neither extraction nor conformance reaches the production emitter.
• No GPU execution in CI. See the note above; device validation is manual.
• Warp primitives are emitted but not reachable from kernel source. SWarpBarrier and SMemFence are lowered by all six generators — bar.warp.sync on PTX, __syncwarp() on CUDA, sub_group_barrier on OpenCL, simdgroup_barrier on Metal, subgroupBarrier() on GLSL and WGSL, each pinned by sarek/tests/unit/test_sync_stmt_emission.ml. What is missing is the front end: no PPX surface syntax constructs either statement, so today they are reachable only from hand-built IR. The front end declares three synchronising intrinsic names — block_barrier, memory_fence_block, memory_fence_device — as entries in Sarek_core_primitives' primitive table, which Sarek_env.with_stdlib binds into kernel scope; lowering leaves all three as EIntrinsic expression calls and never builds SWarpBarrier or SMemFence from them, which is why those two statements have no surface spelling. warp_barrier is not among them: no Sarek_core_primitives entry and no let%sarek_intrinsic in Sarek_stdlib.Gpu declares it, so it cannot be named from Sarek source and cannot reach the IR at all — pinned by sarek/tests/unit/test_sync_stmt_emission.ml ("warp_barrier is not a declared core primitive", with block_barrier as the positive control). The four-name list in Sarek_lower_ir.synchronising_intrinsics is not a declaration of surface names: it is the refusal guard for the initializer-prefixing pass, and it carries warp_barrier ahead of any declaration on purpose, so that adding the primitive cannot silently escape the refusal. Warp shuffle/vote/ballot are a separate, larger gap — not modelled in the IR and emitted by no backend.
• f16 is refused by this emitter. Sarek_ir_ptx_types has no TFloat16 register type; f16 codegen exists on the CUDA-C/HIP path only.

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

Two paths, both real:

• sarek-hip — the native ROCm path (hiprtc → gfx code object → hipModuleLaunchKernel), no ZLUDA involved. It reuses the CUDA-C generator verbatim, because HIP C++ is source-compatible with the CUDA-C subset Sarek emits. Measured against OpenCL and Vulkan on an RX 7900 XTX in docs/benchmarks/hip-vs-opencl-vulkan-2026-07-24.md.
• CUDA/PTX under ZLUDA — described below; this is the path the PTX emitter's device evidence comes from.

CUDA/PTX 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: 22 benchmarks (benchmarks/bench_*.ml) covering compute-bound (matrix multiplication naive/tiled, Mandelbrot, n-body, conv2d), memory-bound (vector add/copy, stream triad, reduction, stencil, gather/scatter), irregular (histogram, bitonic and radix sort, scan), and layout/transfer optimisation patterns (transpose naive vs tiled, AoS vs SoA, pinned transfer). 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