Installation

cuTile.jl is installed like any other Julia package. From the Julia REPL, type ] to enter the Pkg REPL mode and run:

pkg> add cuTile

Or, equivalently, via the Pkg API:

julia> import Pkg; Pkg.add("cuTile")

Requirements

Julia 1.11 or newer.
An NVIDIA driver supporting CUDA 13 (version 580 or later). cuTile.jl generates Tile IR, which the CUDA 13 toolchain compiles to CUBIN.
A GPU with compute capability 8.0 or newer (Ampere and later). Individual features may require newer architectures; see Compatibility.
CUDA.jl, imported before launching a kernel. It provides the CuArray type and the @cuda backend=cuTile launch path.

You do not need to install the CUDA toolkit yourself: CUDA.jl downloads the appropriate artifacts automatically. See the CUDA.jl installation guide for driver, toolkit-selection and troubleshooting details; cuTile.jl does not duplicate that setup here.

Checking your setup

Kernel compilation goes through tileiras, the Tile IR assembler shipped with the CUDA compiler artifacts. To see which one is in use, and which Tile IR bytecode version cuTile will emit by default:

julia> using cuTile

julia> cuTile.versioninfo()
cuTile toolchain:
- tileiras 13.3.36, artifact installation
- bytecode v13.3, auto-detected

The bytecode version is probed from the tileiras binary in use, and can be overridden with the bytecode_version preference; cuTile.bytecode_version() returns the value on its own. Which features are available at which version is documented in Compatibility.

To run the test suite:

julia> using Pkg

julia> Pkg.test("cuTile")

The test suite requires a functional CUDA GPU and exits immediately when CUDA.functional() is false. Building the documentation and using ct.code_tiled with explicit argument types do not require a GPU.