ffdas

ffdas is a CUDA-accelerated library of delay-and-sum and related primitives for image reconstruction in high frame rate ultrasound, photoacoustics, and similar domains. Python and MATLAB bindings integrate directly with common GPU array libraries (CuPy, PyTorch, MATLAB gpuArrays) via zero-copy interop.

import cupy as cp
import ffdas

# reconstruct a volume from acquired channel data
output = ffdas.das(
    channel_data,  # (batch, channels, sequence, samples)
    channel_pos,   # (channels, 3)
    voxel_pos,     # (..., 3), e.g., (nz, ny, nx, 3)
    delays,        # (sequence, ...)
    weights,       # (sequence, ...)
)

# remove tissue clutter for flow imaging
filtered = ffdas.truncate_rank(output, start=64)

# interpolate to arbitrary positions
interpolated = ffdas.interpolate(
    voxel_pos,     # (nz, ny, nx, 3)
    filtered,      # (batch, nz, ny, nx)
    interp_pos,    # (..., 3)
)

Installation

Python

pip install ffdas[cuda13]

Requires Python 3.12+ and a CUDA-capable GPU. CuPy or another DLPack-compatible GPU array library is needed to create the input arrays.

MATLAB

Download the prebuilt MEX binaries for your platform from the GitHub Releases page and add the bindings directory to your MATLAB path:

addpath("/path/to/ffdas/bindings/matlab")

Requires MATLAB R2018b+ with the Parallel Computing Toolbox.

Building from Source

See the installation guide for building from source, including platform-specific instructions and CMake options.

Operations

The library provides GPU-accelerated implementations of:

• Delay-and-sum (das, das_sparse) with multiple algorithm variants, sparse compounding, half-precision compute paths, and per-channel directivity masking.
• Green's function summation (greens) for frequency-domain wave propagation modeling. Requires SM 70+ (Volta or newer).
• Rank truncation (truncate_rank) for subspace-based clutter rejection.
• Tensor contraction (einsum) for arbitrary binary contractions on GPU arrays.
• Structured grid interpolation (interpolate) with nearest-neighbor and linear modes.
• Gather/scatter and contiguous-copy utilities for GPU tensor manipulation.

Python also provides spatial utility functions (cdist, rect_dist, angle) for computing transmit geometry. The MATLAB bindings include these plus take and take_along_axis.

Conventions

Positions and time offsets passed to das are in sampling wavelengths (c / f_s), making the kernel unitless. In practice this means scaling spatial coordinates by f_s / c and expressing time offsets in samples. greens and interpolate are unit-agnostic — positions just need to be consistent with the wavenumbers or query points respectively. The quickstart walks through this conversion for a concrete transmit setup.

The Python and MATLAB bindings follow the layout conventions of their respective languages. In Python, arrays are row-major with the coordinate dimension last: positions have shape (..., 3). In MATLAB, arrays are column-major with the coordinate dimension first: (3, ...). The rest of the API is analogous between the two. See the conventions reference for details.

Examples

The bindings/python/examples/ directory contains self-contained scripts that simulate data and demonstrate each part of the pipeline:

• reconstruct.py — End-to-end volume reconstruction: simulate diverging-wave channel data via Green's function propagation, compute transmit geometry, and reconstruct a 3D image with das.
• compound.py — Coherent plane-wave compounding: transmit at several steering angles and combine the resulting images with das.
• sparse_compounding.py — Focused-wave compounding with das_sparse: each voxel selects a per-target subset of transmit events.
• clutter_filter.py — Clutter filtering on a frame sequence: separate stationary tissue from moving flow using truncate_rank.
• interpolation.py — Reconstruct on a spherical grid, then interpolate back to Cartesian coordinates.

Equivalent MATLAB examples are in bindings/matlab/examples/.

Requirements

ffdas requires an NVIDIA GPU with compute capability 5.3 (Maxwell) or higher. The greens function uses tensor core instructions and requires compute capability 7.0 (Volta) or higher; calling it on an older GPU produces a runtime error.

Citation

If you use ffdas in your work, please cite:

L. Verhoef and P. Kruizinga, "ffdas: Volumetric ultrasound reconstruction at warp speed," arXiv:2606.13259, 2026. [paper]

@article{verhoef2026ffdas,
    title={ffdas: Volumetric ultrasound reconstruction at warp speed},
    author={Verhoef, Luuk and Kruizinga, Pieter},
    journal={arXiv preprint arXiv:2606.13259},
    year={2026},
}

License

MIT License (see LICENSE.txt)