OpenCL-based simulator for Nengo neural models
OpenCL-based Nengo Simulator
NengoOCL is an OpenCL-based simulator for brain models built using Nengo. It can be orders of magnitude faster than the reference simulator in nengo for large models.
To use the nengo_ocl project’s OpenCL simulator, build a Nengo model as usual, but use nengo_ocl.Simulator when creating a simulator for your model:
import numpy as np import matplotlib.pyplot as plt import nengo import nengo_ocl # define the model with nengo.Network() as model: stim = nengo.Node(np.sin) a = nengo.Ensemble(100, 1) b = nengo.Ensemble(100, 1) nengo.Connection(stim, a) nengo.Connection(a, b, function=lambda x: x**2) probe_a = nengo.Probe(a, synapse=0.01) probe_b = nengo.Probe(b, synapse=0.01) # build and run the model with nengo_ocl.Simulator(model) as sim: sim.run(10) # plot the results plt.plot(sim.trange(), sim.data[probe_a]) plt.plot(sim.trange(), sim.data[probe_b]) plt.show()
If you are running within nengo_gui make sure the PYOPENCL_CTX environment variable has been set. If this variable is not set it will open an interactive prompt which will cause nengo_gui to get stuck during build.
Dependencies and Installation
The requirements are the same as Nengo, with the additional Python packages mako and pyopencl (where the latter requires installing OpenCL).
- Python 2.7+ or Python 3.3+ (same as Nengo)
- One or more OpenCL implementations (test with e.g. PyOpenCl)
A working installation of OpenCL is the most difficult part of installing NengoOCL. See below for more details on how to install OpenCL.
In the ideal case, all of the Python dependencies will be automatically installed when installing nengo_ocl with
pip install nengo-ocl
If that doesn’t work, then do a developer install to figure out what’s going wrong.
First, pip install nengo. For best performance, first make sure a fast version of Numpy is installed by following the instructions in the Nengo README.
This repository can then be installed with:
git clone https://github.com/nengo/nengo-ocl.git cd nengo-ocl python setup.py develop --user
If you’re using a virtualenv (recommended!) then you can omit the --user flag. Check the output to make sure everything installed correctly. Some dependencies (e.g. pyopencl) may require manual installation.
How you install OpenCL is dependent on your hardware and operating system. A good resource for various cases is found in the PyOpenCL documentation:
- Installing PyOpenCL on Windows
- Installing PyOpenCL on Mac OS X
- Installing PyOpenCL on Linux, and a more detailed guide
Below are instructions that have worked for the NengoOCL developers at one point in time.
AMD OpenCL on Debian Unstable
On Debian unstable (sid) there are packages in non-free and contrib to install AMD’s OpenCL implementation easily. Actually, the easiest thing would be to apt-get install python-pyopencl. But if you’re using a virtual environment, you can sudo apt-get install opencl-headers libboost-python-dev amd-opencl-icd amd-libopencl1 and then pip install pyopencl.
Nvidia OpenCL on Debian/Ubuntu Linux
On Debian unstable (sid) there are packages for installing the Nvidia OpenCL implementation as well.
sudo apt-get install nvidia-opencl-common nvidia-libopencl1
Ensure that the Nvidia driver version matches the OpenCL library version. You can check the Nvidia driver version by running nvidia-smi in the command line. You can find the OpenCL library version by looking at the libnvidia-opencl.so.XXX.XX file in the /usr/lib/x86_64-linux-gnu/ folder.
From the nengo-ocl source directory, run:
py.test nengo_ocl/tests --pyargs nengo -v
This will run the tests using the default context. If you wish to use another context, configure it with the PYOPENCL_CTX environment variable (run the Python command pyopencl.create_some_context() for more info).
2.1.0 (Nov 23, 2020)
Compatible with Nengo 3.1.0
- Added remove_zero_incs and remove_unmodified_resets simplifications for the operator list. These are enabled by default, and remove unnecessary operators (e.g. that are multiplying by zero and adding that to a signal). This increases both build speed and run speed. These simplifications can be disabled by modifying nengo_ocl.operators.simplifications. (#183)
2.0.0 (Sept 4, 2020)
Compatible with Nengo 3.0.0
- The documentation is now online at https://labs.nengo.ai/nengo-ocl/ (#179)
- Sparse transforms are now supported. (#176)
- Added Simulator.clear_probes method to clear probe data stored in memory. (#179)
1.4.0 (July 4, 2018)
- Supports recent Nengo versions, up to 2.8.0.
- Supports the new SpikingRectifiedLinear neuron type.
1.3.0 (October 6, 2017)
- Supports recent Nengo versions, up to 2.6.0.
- Fixed an issue in which stochastic processes would not be fully reset on simulator reset.
- Fixed an issue in which building a model multiple times could result in old probe data persisting.
1.2.0 (February 23, 2017)
- Supports all Nengo versions from 2.1.2 to 2.3.1.
- nengo_ocl.Simulator is no longer a subclass of nengo.Simulator, reducing the chances that Nengo OCL will be affected by changes in Nengo.
1.1.0 (November 30, 2016)
- Added support for RectifiedLinear and Sigmoid neuron types.
- Added support for arbitrary Process subclasses. Unlike the processes that are explicitly supported like WhiteSignal, these processes may not fully utilize the OpenCL device.
- Added support for applying synaptic filters to matrices, which is commonly done when probing connection weights.
- Supports all Nengo versions from 2.1.2 to 2.3.0.
- The LIF model is now more accurate, and matches the implementation in Nengo (see Nengo#975).
- Several operators have been optimized and should now run faster.
- Fixed compatibility issues with Python 3, and certain versions of NumPy and Nengo.
1.0.0 (June 6, 2016)
Release in support of Nengo 2.1.0. Since Nengo no longer supports Python 2.6, we now support Python 2.7+ and 3.3+.
- Added support for Process class and subclasses, new in Nengo in 2.1.0. We specifically support the WhiteNoise, WhiteSignal, and PresentInput processes. We also support the Conv2d and Pool2d processes in nengo_extras.
- LinearFilter is now fully supported, allowing for general synapses.
- The Numpy simulator in this project (sim_npy) has been phased out and combined with the OCL simulator (sim_ocl). It is now called Simulator and resides in simulator.py.
- Operator scheduling (i.e. the planner) is much faster. We still have only one planner (greedy_planner), which now resides in planners.py.
- Many small speed improvements, including a number of cases where data was needlessly copied off the device to check sizes, dtypes, etc.
- Updated examples to use up-to-date Nengo syntax.
0.1.0 (June 8, 2015)
Initial release of Nengo OpenCL! Supports Nengo 2.0.x on Python 2.6+ and 3.3+.
Hashes for nengo_ocl-2.1.0-py3-none-any.whl