osmodi-solver 0.1.1

GPU-accelerated Poisson solver for pressure from PIV

OSMODI is a GPU-accelerated Poisson solver for pressure reconstruction from Particle Image Velocimetry (PIV) data. It provides a CUDA implementation for NVIDIA GPUs with a reliable CPU fallback.

Github Repository

Link

Features

• CUDA GPU solver (NVIDIA GPUs)
• CPU-only fallback (always available)
• 2D and 3D structured grids
• Float32 and Float64 support
• Designed for pressure-from-PIV workflows

---

Installation

pip install osmodi-solver

---

Quick Start

import osmodi
import numpy as np

print("GPU available:", osmodi.GPU_AVAILABLE)
print("CPU available:", osmodi.CPU_AVAILABLE)

# Create sample velocity gradients
Sx = np.random.randn(64, 64, 64).astype(np.float32)
Sy = np.random.randn(64, 64, 64).astype(np.float32)
Sz = np.random.randn(64, 64, 64).astype(np.float32)

delta = np.array([1.0])

if osmodi.GPU_AVAILABLE:
    P, progress = osmodi.solve_gpu(Sx, Sy, Sz, delta)
else:
    P, progress = osmodi.solve_cpu(Sx, Sy, Sz, delta)

print("Output shape:", P.shape)

---

API Overview

solve_cpu(Sx, Sy, Sz, delta, options=None) Runs the CPU-based Poisson solver.

solve_gpu(Sx, Sy, Sz, delta, options=None) Runs the CUDA GPU-based Poisson solver.

Parameters

Sx, Sy, Sz: NumPy arrays (2D or 3D) of velocity gradients

delta: grid spacing NumPy array

options: optional dictionary for solver configuration

Returns

P: pressure field (same shape as input)

progress: convergence history (NumPy array)

Solver Options

options = {
    "Verbose": True,               # Print solver progress
    "SolverToleranceRel": 1e-4,   # Relative tolerance
    "SolverToleranceAbs": 1e-6,   # Absolute tolerance
    "Kernel": "face-crossing",     # or 'cell-centered'
}

---

Supported Hardware

CPU: x86-64, any modern processor

GPU (optional): NVIDIA Pascal or newer

RAM: 4 GB minimum, 16+ GB recommended for large grids

---

Citation

If you use this solver in your work, please cite:

Zigunov, F., & Charonko, J. J. (2024). One-shot omnidirectional pressure integration through matrix-free multiscale methods. Measurement Science and Technology.

---

License

GNU General Public License v3.0 (GPL-3.0)