Jan Petykiewicz
d5fd78d493
This reverts commit 60b70bb332d29fb5a2aecb801517be659cf4daf9. It appears to have minimal performance impact, and fails to compile on nvidia cards.
opencl_fdtd
opencl_fdtd is a python application for running 3D time-domain electromagnetic simulations on parallel compute hardware (mainly GPUs).
Performance highly depends on what hardware you have available:
- A 395x345x73 cell simulation (~10 million points, 8-cell absorbing boundaries) runs at around 91 iterations/sec. on my AMD RX480.
- On an Nvidia GTX 580, it runs at 66 iterations/sec
- On my laptop (Nvidia 940M) the same simulation achieves ~12 iterations/sec.
- An L3 photonic crystal cavity ringdown simulation (1550nm source, 40nm discretization, 8000 steps) takes about 3 minutes on my laptop.
Capabilities are currently pretty minimal:
- Absorbing boundaries (CPML)
- Perfect electrical conductors (PECs; to use set epsilon to inf)
- Anisotropic media (eps_xx, eps_yy, eps_zz, mu_xx, ...)
- Direct access to fields (eg., you can trivially add a soft or hard current source with just sim.E[ind] += sin(f0 * t), or save any portion of a field to a file)
Installation
Requirements:
- python 3 (written and tested with 3.5)
- numpy
- pyopencl
- jinja2
- fdfd_tools
Optional (used for examples):
To get the code, just clone this repository:
git clone https://mpxd.net/code/jan/opencl_fdtd.git
You can install the requirements and their dependencies easily with
pip install -r requirements.txt
Running
The root directory contains fdtd.py, which sets up and runs a sample simulation
(cavity ringdown).
python3 fdtd.py