# 2.12. Configuration¶

## 2.12.1. Setting the discretization step¶

An important value for the simulation is the discretization step dt. Its default value is 1 ms, which is usually fine for rate-coded networks, but may be too high for spiking networks, as the equations are stiffer. Taken too high, it can lead to high numerical errors. Too low, and the simulation will take an unnecessary amount of time.

To set the discretization step, just pass the desired value to setup() at the beginning of the script, or at any rate before the call to compile():

setup(dt=0.1)


Changing its value after calling compile() will not have any effect.

## 2.12.2. Setting the seed of the random number generators¶

By default, the random number generators are seeded with time(NULL), so each simulation will be different. If you want to have deterministic simulations, you simply need to provide a fixed seed to setup():

setup(seed=62756)


Note that this also sets the seed of Numpy, so you can also reproduce random initialization values prduced by numpy.random.

Note: Using the same seed with the OpenMP and CUDA backends will not lead to the same sequences of numbers!

## 2.12.3. Cleaning the compilation directory¶

When calling compile() for the first time, a subfolder annarchy/ will be created in the current directory, where the generated code will be compiled. The first compilation may last a couple of seconds, but further modifications to the script are much faster. If no modification to the network has been made except for parameter values, it will not be recompiled, sparing us this overhead.

ANNarchy tracks the changes in the script and re-generates the corresponding code. In some cases (a new version of ANNarchy has been installed, bugs), it may be necessary to perform a fresh compilation of the network. You can either delete the annarchy/ subfolder and restart the script:

$rm -rf annarchy/$ python MyNetwork.py


or pass the --clean flag to Python:

$python MyNetwork.py --clean  ## 2.12.4. Selecting the compiler¶ ANNarchy requires a C++ compiler. On GNU/Linux, the default choice is g++, while on MacOS it is clang++. You can change the compiler (and its flags) to use either during the call to compile() in your script: compile(compiler="clang++", compiler_flags="-march=native -O2")  or globally by modifying the configuration file located at ~/.config/ANNarchy/annarchy.json: { "openmp": { "compiler": "clang++", "flags": "-march=native -O2" } }  Be careful with the flags: for example, the optimization level -O3 does not obligatorily produce faster code. ## 2.12.5. Parallel computing with OpenMP¶ The default paradigm for an ANNarchy simulation is through openMP, which distributes automatically the computations over the available CPU cores. By default, OpenMP would use all the available cores for your simulation, even if it is not optimal: small networks in particular tend to run faster with a small amount of cores (for the provided example with Neural Fields, it is for example 2). For this reason, the OMP_NUM_THREADS environment variable has no effect in ANNarchy. You can control the number of cores by passing the -j flag to the Python command: user@machine:~$ python NeuralField.py -j2


It is the responsability of the user to find out which number of cores is optimal for his network, by comparing simulation times. When this optimal number is found, it can be hard-coded in the script by setting the num_threads argument to ANNarchy.setup():

from ANNarchy import *


## 2.12.6. Parallel computing with CUDA¶

To run your network on GPUs, you need to declare to ANNarchy that you want to use CUDA. One way to do so is to pass the --gpu flag to the command line:

user@machine:~$python NeuralField.py --gpu  You can also set the paradigm argument of ANNarchy.setup() to make it permanent: from ANNarchy import * setup(paradigm="cuda")  If there are multiple GPUs on your machine, you can select the ID of the device by specifying it to the --gpu flag on the command line: user@machine:~$ python NeuralField.py --gpu=2


You can also pass the cuda_config dictionary argument to compile():

compile(cuda_config={'device': 2})


The default GPU is defined in the configuration file ~/.config/ANNarchy/annarchy.json (0 unless you modify it).

{
"cuda": {
"device": 0,
"path": "/usr/local/cuda"
}
}


Hint

As the current implementation is a development version, some of the features provided by ANNarchy are not supported yet with CUDA:

• weight sharing,
• non-uniform synaptic delays,
• structural plasticity,
• spiking neurons: a) with mean firing rate and b) continous integration of inputs,
• SpikeSourceArray.