XTC

Links

Refer to documentation at https://corse.gitlabpages.inria.fr/XTC

Refer to installable python packages at: https://gitlab.inria.fr/corse/xtc/-/packages

Refer to tutorials here and to additionnal developers documentation here.

Overview

XTC is a domain-specific dataflow graph compiler for linear algebra operations. It provides:

• Operational DSL: Define computation graphs with tensors and operators
• Scheduling DSL: High-level transformations (tiling, parallelization, vectorization, etc.)
• Multiple backends: MLIR (linalg + transform), TVM (Tensor IR), JIR (INRIA internal)
• Autotuning: Definition and exploration of the optimization space

Build & Development

System requirements

Debian-like distributions:

sudo apt install python3 python3-dev build-essential libomp5 binutils binutils-aarch64-linux-gnu binutils-x86-64-linux-gnu

# Optionally if using PMU counters on CPU for evaluation
sudo apt install libpfm4-dev 
sudo sysctl kernel.perf_event_paranoid=1

Fedora:

sudo dnf install python3 python3-devel libomp binutils binutils-aarch64-linux-gnu binutils-x86_64-linux-gnu

# For Fedora 40+
sudo dnf group install c-development development-tools # For Fedora 40+

# Optionally if using PMU counters on CPU for evaluation
sudo dnf install libpfm-devel
sudo sysctl kernel.perf_event_paranoid=1

Installation

Python: 3.10 to 3.14 inclusive

python3 -m venv .venv && source .venv/bin/activate
pip3 install -e '.[dev]'
pip3 install -r mlir_requirements.txt  # Optional: MLIR backend
pip3 install -r tvm_requirements.txt   # Optional: TVM backend
make test                              # Run minimal unit tests

Code quality

Code quality requirements:

• Type annotations: Strict pyright mode, full annotations required
• Formatting: Ruff (line length 88)
• License headers: BSD-3-Clause required on all source files
• All checks must pass before merge

Type checking:

make check-type          # Run both pyright and mypy
pyright                  # Run pyright only
mypy                     # Run mypy only

Formatting:

make format              # Apply all formatting (license + ruff)
make check-format        # Check formatting without modifying files

Testing structure:

• tests/pytest/unit/: Core interface unit tests
• tests/pytest/{mlir,tvm}/: Backend-specific tests
• tests/filecheck/: Lit+FileCheck functional tests for code generation

Global test commands:

make test                # Run minimal unit tests
make check               # Run ALL acceptance tests (required for contributions)
make check-pytest        # Run pytest suite only
make check-lit           # Run LIT tests for LLVM IR target
make check-lit-c         # Run LIT tests for C target
pytest tests/pytest/unit # Run specific test directory

Running individual tests:

# Single pytest file
pytest tests/pytest/unit/test_specific.py -v

# Single lit test
lit -v tests/filecheck/backends/specific_test.py

# C target for lit tests
XTC_MLIR_TARGET=c lit -v tests/filecheck/backends/specific_test.py

Architecture

Core Abstractions (src/xtc/itf/)

Abstract interfaces defining the compilation pipeline:

• data/ - Tensor, DataType, ShapeType
• operator/ - Linear algebra operator interface
• graph/ - Graph, Node, Operation abstractions
• back/ - Backend interface
• schd/ - Scheduler and Schedule abstractions
• comp/ - Compiler interface
• exec/ - Executor and Evaluator interfaces
• search/ - Search space exploration interface

Backends (src/xtc/backends/)

Exposed backends:

• mlir/ - MLIR backend using linalg + transform dialects
• tvm/ - TVM backend using Tensor IR + Schedule APIs

XTC also supports multiple MLIR Targets for the code generation:

• llvmir (default)
• c
• nvgpu

To force the use of a specific target, you can set the env variable XTC_MLIR_TARGET=<mlir-target>.

The MLIR backend can be extended using the SDist extension, which provides distribution primitives. To install SDist, follow the instructions in docs/develop/optional_backends.md, in the "MLIR development version" section.

Note that the nvgpu target requires a recent version of Cuda (tested with Cuda 13.0). By default, it tries to find Cuda at /usr/local/cuda, but it can be overridden with the CUDA_INSTALL_DIR env variable. The performance counters can be accessed is the GPU has a compute capability >=7.5.

Compilation Pipeline

1. User defines Graph with Tensors and Operators
2. Backend created from Graph
3. Scheduler applies transformations and produces Schedule
4. Compiler generates executable Module
5. Executor/Evaluator runs and measures performance

CLI Tools (src/xtc/cli/)

• mlir-loop - High-level scheduling for MLIR linalg operators
• mlir-backend - MLIR backend wrapper
• loop-explore - Autotuning and space exploration
• loop-display - Visualization of exploration results

AI assistants

To create the CLAUDE.md file required to use Claude Code: make claude