Half-Signed Optical Dot-Product Multiplier

This repository accompanies “Free-Space Coherent Optical Dot-Product Multiplier with Lensless Fan-In” (Duque et al., 2026).

It provides simulation code for a coherent free-space optical dot-product system using DMD/SLM-style modulation and camera-like readout.

The code computes the dot product between two vectors (one signed, one non-negative) by modulating the amplitude of a CW monochromatic laser beam using a DMD, followed by phase modulation with an SLM. The output is recovered via interferometric detection on a camera.

Field propagation is based on Rafael Fuente’s Diffractsim library.

Features

• Configurable device parameters (DMD/SLM resolution, pixel pitch)
• Configurable propagation parameters (wavelength, focal length, Fourier-plane pupil)
• Configurable input sizes and distributions (random uncorrelated; images from CIFAR-10; images from CelebA)
• Configurable parallelism (number of simultaneous dot products via lenslet grid)
• Noise and nonideality modelling (shot/read noise, LC-SLM crosstalk, device misplacement)
• Per-run outputs: saved logs including linear-fit plots and error metrics

Installation

1. Clone this repository

git clone https://github.com/aliceduque/half_signed_opti_multiplier.git
cd half_signed_opti_multiplier

2. Create a conda environment using config file provided

conda env create -f environment.cpu.yml
conda activate half_signed_opti_multiplier

If running this code on CPU, this is the end of installation.

3. If using GPU (recommended), ensure torch packages are installed appropiately:

python -m pip uninstall -y torch torchvision
python -m pip install --index-url https://download.pytorch.org/whl/cu121 torch==2.4.1+cu121 torchvision==0.19.1+cu121

Data

In addition to random vectors, the code supports vectors derived from CIFAR-10 and CelebA.

• CIFAR-10 is downloaded automatically via torchvision the first time it is used.
• CelebA must be downloaded manually (e.g. from Kaggle) and the raw .jpg images placed under ./data/celeba/all/. (The dataset is not redistributed in this repository.)

Usage

opti_product.py is the main file. The simulation is configured via a set of command-line (CLI) arguments, which also makes it easy to build wrapper scripts to run parameter sweeps and batch experiments.

Examples

1. Single dot product, 8-bit, 2025-sized vectors

python opti_product.py --input_type figure --dataset cifar --find_linear_fit --seed 42 --pix_block_size 3 

2. Single dot product, 4-bit, 32400-sized vectors

python opti_product.py --input_type figure --dataset celeb --find_linear_fit --seed 42 --pix_block_size 3 --input_size 180 --cluster_size 4

3. 9 parallel dot products, 8-bit, 225-sized vectors

python opti_product.py --input_type figure --dataset cifar --find_linear_fit --seed 42 --pix_block_size 3 --lenslets 3 --input_size 15 --active_area_ratio 0.4982 --dmd_pixels 1020 --slm_pixels 1020 --focal_length 160e-3

4. Single dot product, 8-bit, 2025-sized vectors, shot noise considered

python opti_product.py --input_type figure --dataset celeb --find_linear_fit --seed 42 --pix_block_size 3 --shot_noise

When running on CPU, runtime for default resolution is severely affected. In that case, it's advised to reduce resolution by 2, by including the arguments --Nx 8192 --dx 2e-6.