tileir

FlagTree TileIR Backend

Source

Imported source

• Upstream repo: https://github.com/triton-lang/Triton-to-Tile-IR
• Upstream commit: a3befd959b02410cfbdac08d91d817b0ec0b3e33
• Synced path: third_party/tileir

Pinned dependencies

• cuda-tile submodule path: third_party/tileir/third_party/cuda-tile
• cuda-tile commit: 2e5ccba66fb3afdba34b26cf358418283027c248
• cuda-tile reference: v13.3.0
• tileiras version used for validation: CUDA 13.3, V13.3.36

Build

When FLAGTREE_BACKEND is not set, FlagTree builds the default in-tree backend set, including nvidia, amd, and tileir. Setting FLAGTREE_BACKEND=tileir selects TileIR explicitly while retaining the NVIDIA and AMD backends required for per-kernel fallback.

TileIR requires CUDA TileIR tooling from CTK 13.3. The validated toolchain is:

• tileiras from CUDA 13.3, version V13.3.36
• GCC or Clang 13 or newer
• the cuda-tile submodule at the pinned commit above

The TileIR path uses tileiras. Native NVIDIA fallback and comparison runs use FlagTree's normal NVIDIA assemblers, managed by setup.py at:

• third_party/nvidia/backend/bin/ptxas
• third_party/nvidia/backend/bin/ptxas-blackwell

pip install and pip install -e initialize the pinned cuda-tile submodule automatically through setup.py; no manual submodule command is needed.

At runtime, TileIR looks for tileiras in this order:

1. $TRITON_TILEIRAS_PATH/tileiras
2. triton/backends/nvidia/bin/tileiras in the installation
3. tileiras on PATH

Default FlagTree source install:

MAX_JOBS=32 python3 -m pip install . --no-build-isolation

For local editable development, the usual editable install also works when the build dependencies are already present:

MAX_JOBS=32 python3 -m pip install -e . --no-build-isolation

Runtime

Backend discovery

A normal installation registers the selected backends through the triton.backends entry-point group. No discovery environment variable is required.

The existing FlagTree option below bypasses entry points and scans the installed triton/backends/ directory instead:

TRITON_BACKENDS_IN_TREE=1

This option only changes where Triton discovers already-present backends. It does not build, install, enable, or route TileIR and is normally unnecessary.

Runtime routing

Routing is disabled by default. FLAGTREE_USE_TILEIR=1 enables per-kernel selection; it does not force every NVIDIA-targeted Triton kernel onto TileIR:

FLAGTREE_USE_TILEIR=1

On NVIDIA hardware, the initial Triton target is labeled cuda. The routing result is:

non-`cuda` Triton target                      -> unchanged
`cuda` target, Triton kernel uses no TLE      -> TileIR
`cuda` target, supported top-level `tle` subset only -> TileIR
`cuda` target, other or unknown TLE usage     -> native NVIDIA

For routing diagnostics:

FLAGTREE_TILEIR_VERBOSE=1

See the TileIR tutorials for functional checks and benchmark commands.

The two paths use these names:

Path	Backend name	Target label	Compiler	Kernel driver
Native NVIDIA	nvidia	cuda	CUDABackend	CudaDriver
TileIR	tileir	tileir	TileIRBackend	TileIRDriver

CudaDriver provides the initial cuda target. Routing either keeps it or changes it to tileir; the final target selects the compiler and kernel driver shown above.

Vendor Maintenance

FlagTree compatibility

• Target FlagTree/Triton base: main / Triton 3.6.0
• FlagTree's pinned LLVM is older than the LLVM expected by cuda-tile v13.3.0.
• scripts/patch_bytecode_utils.sh is a FlagTree-local compatibility patch for that LLVM skew. It rewrites cuda-tile source before build, including:
  • DenseTypedElementsAttr to DenseIntOrFPElementsAttr
  • llvm::scope_exit to llvm::make_scope_exit
  • DenseElementsAttr::isValidRawBuffer 2-argument calls to the 3-argument API

FlagTree-local changes

Compared with the imported upstream commit, FlagTree modifies 9 files and adds 4 entries:

• Build and provenance: Modified README.md, CMakeLists.txt, and scripts/patch_bytecode_utils.sh to record the vendor base, build the pinned cuda-tile source, and support FlagTree's LLVM. Added the third_party/cuda-tile submodule.
• Routing and compiler integration: Modified backend/compiler.py for backend hooks and lazy tileiras lookup. Added backend/router.py to choose TileIR or native NVIDIA per kernel.
• TLE frontend: Added backend/extend_core.py and backend/extend_semantic.py; CudaTileSemantic extends TritonSemantic with the view/token operations exposed as tle.<name> through tl.ext. Modified backend/code_generator.py to use it for TileIR kernels and triton_tileir.cc to expose the corresponding C++ builder methods.
• Lowering: Modified lib/TritonToTileIR/TritonToTileIRPass.cpp to add math.erf lowering and align flattened TensorDescriptor arguments with FlagTree's ABI.
• Release builds: Modified lib/TritonToTileIR/Utils.cpp so its unsupported reduction-type path remains unreachable when NDEBUG removes assertions.
• Runtime: Upstream activates TileIRDriver globally with ENABLE_TILE=1 and creates it during import. Modified backend/driver.py so it is not a global active driver and is created only for a kernel routed to TileIR. The file also adapts binary loading and TensorDescriptor arguments to FlagTree's runtime ABI.

TileIR-specific routing and frontend behavior must remain in this vendor directory. Changes under python/triton are limited to backend-neutral hooks so that non-TileIR backends do not import or depend on TileIR implementation code.

Update checklist

1. Sync third_party/tileir from the chosen Triton-to-Tile-IR commit.
2. Keep third_party/tileir/third_party/cuda-tile as an explicit submodule pin.
3. Reapply or verify the FlagTree-local changes above.
4. Recheck the LLVM compatibility patch in scripts/patch_bytecode_utils.sh.
5. Run TileIR routing, TLE/TKO, and tutorial validation before updating this file.