cuTile.jl → Triton backend prototype

A prototype backend that retargets cuTile.jl from NVIDIA Tile IR (tileiras) to Triton, for NVIDIA and AMD GPUs, without Reactant and without any Triton-specific native library:

Julia kernel  →  cuTile front end      (code_structured, unchanged)
              →  TritonEmitter.jl      (MLIR.jl, unregistered tt dialect → TTIR text)
              →  triton pip wheel      (in-process via PythonCall, CondaPkg-pinned 3.8.0)
              →  CUDA.jl / AMDGPU.jl   (module load + launch)

Kernels are written in the plain cuTile language; nothing kernel-side knows Triton exists. Prototype quality: this exists to inform the design of a real Triton target inside cuTile.jl, not to be a package. Internals, workarounds and upstream findings are in DEVDOCS.md.

Status (October 2026, triton 3.8.0, cuTile 1.0.0)

	H100 NVL (sm_90)	MI300A (gfx942, ROCm 7.2.4)
cuTile's device suite (pass / fail / error)	1,916 / 0 / 0	1,915 / 1 / 0 ¹
standalone suite (test/runtests.jl)	61 / 61	61 / 61

¹ core.jl's failing-assertion test runs native cuTile on CUDA in a child process; the standalone suite checks the same behaviour through this backend.

The whole cuTile 1.0 language is covered, including gather/scatter views, atomics through views and masked atomic_cas, RNG, asserts and runtime throws, general loops, and launch hints. On H100, 8 of cuTile's 11 example workloads run at or above native tileiras (fft 3.0×, layernorm 2–2.5×, softmax 1.3–1.7×), MoE at 0.98× and fmha at 0.90×; batched matmul is about half as fast, where tileiras warp-specializes automatically. On MI300A, tile reductions and fusions are 10–13× faster than AMDGPU.jl's GPUArrays paths and matmul matches or beats rocBLAS.

Usage

using TileTriton
k = TritonRun.triton_kernel(my_cutile_kernel, argtypes; name="k")
TritonRun.launch!(k, grid, arrays...)

TileTriton.install_shim!()     # or route all of cuTile (ct.launch, @cuda backend=cuTile) here

On AMD, load AMDGPU and call TileTriton.use_rocm!() first; the shim then runs unmodified cuTile code on ROCArrays.

Reflection works for either vendor from any machine:

TritonRun.code_triton(f, argtypes; stage=:ttgir)    # :ttir :ttgir :llir :ptx :sass :amdgcn
TritonRun.code_triton(f, argtypes; stage=:amdgcn, target=TritonRun.target(:amd))

Setup and tests

] activate .; instantiate
using CondaPkg; CondaPkg.resolve()     # pinned triton==3.8.0 (CondaPkg.toml)

julia --project=. test/runtests.jl                     # standalone suite
julia --project=. test/cutile_suite.jl [files...]      # cuTile's device suite through the shim
TILETRITON_TEST_BACKEND=rocm julia --project=. test/runtests.jl   # either one, on AMD