cutile-basic

A BASIC to CUDA Tile IR compiler. Write GPU kernels in BASIC, compile them to .cubin files via cuTile bytecode and tileiras, and launch them on NVIDIA GPUs.

Documentation

Overview

cutile-basic extends classic BASIC with tile-based GPU operations (TILE, OUTPUT, BID) and built-in functions like MMA, enabling concise expression of GPU kernels such as vector addition and matrix multiplication.

Installation

pip install git+https://github.com/nvidia/cuda-tile.git@basic-experimental

Quick Start

Clone the repository and check out the basic-experimental branch:

git clone https://github.com/nvidia/cuda-tile.git
cd cuda-tile
git checkout basic-experimental

Compile a BASIC program to a .cubin:

python -m cutile_basic.cli examples/vector_add.bas -o vector_add.cubin

Run an end-to-end GPU demo:

python examples/vector_add.py

Or use the Python API:

from cutile_basic import compile_basic_to_cubin

source = """
10 INPUT N, A(), B()
20 DIM A(N), B(N), C(N)
30 TILE A(128), B(128), C(128)
40 LET C(BID) = A(BID) + B(BID)
50 OUTPUT C
60 END
"""

result = compile_basic_to_cubin(source)
print(result.cubin_path)   # path to the compiled .cubin
print(result.meta)         # kernel metadata (arrays, tile shapes, etc.)

Examples

Program	Description
examples/hello.bas	Variables, arithmetic, conditionals, loops
examples/vector_add.bas	GPU vector addition using BID
examples/gemm.bas	Tiled GPU matrix multiply (MMA)

End-to-end GPU demos:

python examples/vector_add.py
python examples/gemm.py
python examples/hello.py

Prerequisites

Hardware:

• NVIDIA GPU with Compute Capability 8.x (Ampere), 10.x, 11.x, or 12.x (Blackwell)

Software:

• NVIDIA Driver r580 or later
• Python 3.10 or later
• CUDA Toolkit 13.1 or later
• cuda-tile[tileiras], cuda-python, cuda-core[cu13], cupy-cuda13x[ctk]

License

Apache License 2.0 with LLVM Exceptions. See LICENSE for details.