cuTile FFT Optical Imaging Demo

A demonstration of using NVIDIA cuTile for 2D FFT convolution in an optical imaging application. This project compares cuTile FFT with CuPy FFT and validates correctness against NumPy.

Overview

This demo applies FFT-based convolution to simulate optical imaging:

• Input: Binary mask pattern (2048×2048)
• Kernels: 24 optical point-spread functions (35×35 each)
• Output: Aerial image showing light intensity after optical system

The Hopkins/SOCS model is used: Aerial = Σᵢ wᵢ |Mask ⊛ Kernelᵢ|²

Requirements

• Python 3.10+
• NVIDIA GPU with CUDA 13.3
• cuTile (cuda.tile)
• PyTorch (2.6+ recommended)
• CuPy
• NumPy
• Matplotlib

Install dependencies:

pip install torch cupy-cuda12x numpy matplotlib

For cuTile (cuda.tile), follow the NVIDIA cuTile installation instructions.

Setup

1. Clone this repository

git clone <repo-url>
cd project

2. Fetch OpenILT data

The optical kernels and sample masks come from the OpenILT project:

git clone https://github.com/OpenOPC/OpenILT.git openilt_data
cd openilt_data
git checkout dabb97c6ca3dfd159362e48273c436444c77353b
cd ..

Note: This demo was tested with OpenILT commit dabb97c. If the repository structure changes, use this specific commit.

This provides:

• openilt_data/kernel/kernels/focus.pt — 24 optical kernels
• openilt_data/kernel/scales/focus.pt — kernel weights
• openilt_data/tmp/CurvILT_target1.png — sample mask pattern

Usage

Generate comparison figure (CuPy vs cuTile)

python optics.py --compare --save comparison.png

Run with CuPy FFT only

python optics.py --cupy --save figure-cupy.png

Run with cuTile FFT only

python optics.py --cutile --save figure-cutile.png

Fast layout iteration (mock mode)

python optics.py --compare --mock --save comparison.png

Options

• --num-kernels N — Use N kernels (default: 24)
• --cutline Y — Set intensity profile cutline position
• --threshold T — Set print threshold for cutline plot

Running FFT Tests

Validate cuTile FFT against NumPy/CuPy/PyTorch:

python test_fft.py

This runs:

• 1D FFT tests (8 elements)
• 2D FFT tests (8×8)
• Benchmarks at 1024×1024 and 2048×2048

Project Structure

project/
├── FFT.py          # cuTile FFT kernel (from NVIDIA samples)
├── test_fft.py     # FFT correctness and benchmark tests
├── optics.py       # Optical imaging simulation
├── display.py      # Visualization utilities
├── openilt_data/   # Data directory (fetch separately)
└── README.md

Performance Notes

Size	CuPy (cuFFT)	cuTile
1024×1024	0.05 ms	92 ms
2048×2048	0.07 ms	324 ms

cuTile FFT produces correct results but is not yet performance-competitive with cuFFT. This demo focuses on correctness and integration.

License

MIT License — see LICENSE

Acknowledgments

• OpenILT — CUHK-EDA, optical kernels and benchmark data
• NVIDIA cuTile — Tile-based GPU programming