TAIDL is a domain-specific language designed to define instruction set architectures (ISAs) for tensor accelerators. It provides a flexible and extensible framework for specifying the operations, data types, and memory models of tensor accelerators.

TAIDL is an integral part of the ACT Ecosystem, enabling automated generation of:

• Test Oracles, a.k.a. functional simulators (act-compiler/act-oracle)
• Compiler Backends (act-compiler/act-backend)

TAIDL is published in MICRO 2025. For detailed evaluations, please refer to our MICRO 2025 Artifact.

pip install git+https://github.com/act-compiler/taidl.git@v2.2

For development:

git clone https://github.com/act-compiler/taidl.git
cd taidl
pip install -e .

• QKV.py — QKV attention accelerator
• QKV_new.py — variant of QKV

pip install -e .
pip install pytest
pytest tests/

TAIDL is a pure data model + language parser. It stores:

• Accelerator metadata (name, constants, state)
• Data model definitions (memory spaces with dimensions and types)
• Instruction definitions (parameters, constraints, inputs/outputs)
• Parsed semantics AST (ANTLR4 parse tree)

Generators (TAIDL-TO, ACT-Backend) consume the Accelerator object and its pre-parsed ASTs to produce target-specific software tools.