Standalone implementation

This – currently very incomplete – document describes some of the implementation details of Standalone code generation.

Array cache

As described in standalone variables, in standalone mode Python code does not usually have access to state variables and synaptic indices, since the code necessary to initialize/create them has not been run yet. Concretely, accessing a state variable (or other variables like synaptic indices), will call ArrayVariable.get_value which delegates to CPPStandaloneDevice.get_value. After a run, this will read the corresponding file from the disk and return the values. The user can therefore use the same code to analyze the results as for runtime mode. Before a run, this file does not exist, but CPPStandaloneDevice.get_value has another mechanism to return values: the “array cache”. This cache is a simple dictionary, stored in CPPStandaloneDevice.array_cache, mapping ArrayVariable objects to their respective values. If the requested object is present in this cache, its values can be accessed even before the simulation is run. Values are added to this cache, whenever simulation code sets variables with concrete values. Methods such as CPPStandaloneDevice.fill_with_array or CPPStandaloneDevice.init_with_zeros write the provided values into the array cache so that they can be retrieved later. Conversely, CPPStandaloneDevice.code_object will delete any existing information in array_cache for variables that are changed by a code object, i.e. invalidate any previously stored values:

>>> set_device('cpp_standalone')
>>> G = NeuronGroup(10, 'v : volt')
>>> v_var = G.variables['v']
>>> print(device.array_cache[v_var])  # CPPStandaloneDevice.init_with_zeros stored initial zero values
[ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.]
>>> G.v = -70*mV
>>> print(device.array_cache[v_var])  # CPPStandaloneDevice.fill_with_array updated the values
[-0.07 -0.07 -0.07 -0.07 -0.07 -0.07 -0.07 -0.07 -0.07 -0.07]
>>> G.v = '-70*mV + i*2*mV'
>>> print(device.array_cache[v_var])  # Array cache for v has been invalidated
None
>>> set_device('runtime')  # Reset device to avoid problems in other doctests

Command line arguments

The mechanisms described in Multiple full simulation runs are implemented via command-line arguments to the main binary. A call such as:

device.run(results_directory='results_1', run_args={group.tau : 10*ms})

will be executed by calling the compiled binary as follows:

./main --results_dir /full/path/to/results_1 neurongroup.tau=0.01

where neurongroup is group.name (either a default name, e.g. neurongroup, or the name set during construction with the name argument). The generated code applies the initialization for tau after the usual variable initializations, before the call of network.run (assuming that application has not been moved to a different place by using apply_run_args).

For initializations with a vector of values, the values are written to disk (in the same, simple binary format that is used elsewhere, e.g. for the results). The command line argument then specifies the file name instead of a value:

./main neurongroup.tau=static_arrays/init_neurongroup_v_aca4cd6a3f7e526a61bb5a07468b377e.dat

The hex string in the filename is an automatically generated MD5 hash of the array content. This makes it possible to assure that each array is only written to disk once, even for repeated and parallel executions with the same values (file locking is used to make sure only one process writes to each file).