GeNN generates code according to the network model defined by the user as described in User Manual . Here we provide a guideline to setup GeNN and use generated functions. We recommend users to also .

Simulating a network model

Any variables marked as uninitialised using the uninitialisedVar() function or sparse connectivity not initialised using a snippet must be initialised by the user between calls to initialize() and initializeSparse().

Core functions generated by GeNN to be included in the user code include:

allocateMem()
initialize()
initializeSparse()
stepTime()
freeMem()
getFreeDeviceMemBytes()

In order to correctly access neuron state and spikes for the current timestep, correctly accounting for delay buffering etc, you can use the getCurrent<var name><neuron name>(), get<neuron name>CurrentSpikes() and get<neuron name>CurrentSpikeCount() functions. Additionally, custom update groups (see Defining custom updates) can be simulated by calling update<group name>().

By setting GENN_PREFERENCES::automaticCopy , GeNN can be used in a simple mode where CUDA automatically transfers data between the GPU and CPU when required (see https://devblogs.nvidia.com/unified-memory-cuda-beginners/). However, copying elements between the GPU and the host memory is costly in terms of performance and the automatic copying operates on a fairly coarse grain (pages are approximately 4 bytes). Therefore, in order to maximise performance, we recommend you do not use automatic copying and instead manually call the following functionsmethods when required:

You can use push<neuron or synapse name>StateToDevice() to copy from the host to the GPU. At the end of your simulation, if you want to access the variables you need to copy them back from the device using the pull<neuron or synapse name>StateFromDevice() function or one of the more fine-grained functions listed above.

Extra Global Parameters

If extra global parameters have a "scalar" type such as float they can be set directly from simulation code. For example the extra global parameter "reward" of population "Pop" can be set with:

However, if extra global parameters have a pointer type such as float*, GeNN generates additional functions to allocate, free and copy these extra global parameters between host and device:

allocate<egp name><neuron or synapse name>
free<egp name><neuron or synapse name>
push<egp name><neuron or synapse name>ToDevice
pull<egp name><neuron or synapse name>FromDevice These operate in much the same manner as the functions for interacting with standard variables described above but the allocate, push and pull functions all take a "count" parameter specifying how many entries the extra global parameter array should be.

Extra global parameters with a pointer type such as float* should be initialised and updated in the same manner but, if their value is changed after pygenn.GeNNModel.load is called, the updated values need to be pushed to the GPU:

pop.extra_global_params["reward"].view[:] = [1,2,3,4]

pop.push_extra_global_param_to_device("reward", 4)

Extra global parameters can also be used to provide additional data to snippets used for variable (see Variable initialisation) or sparse connectivity (see Sparse connectivity initialisation) initialisation.

Floating point precision

Double precision floating point numbers are supported by devices with compute capability 1.3 or higher. If you have an older GPU, you need to use single precision floating point in your models and simulation. Furthermore, GPUs are designed to work better with single precision while double precision is the standard for CPUs. This difference should be kept in mind while comparing performance.

Typically, variables in GeNN models are defined using the scalar type. This type is substituted with "float" or "double" during code generation, according to the model precision. This is specified .

There may be ambiguities in arithmetic operations using explicit numbers. Standard C compilers presume that any number defined as "X" is an integer and any number defined as "X.Y" is a double. Make sure to use the same precision in your operations in order to avoid performance loss.

Working with variables in GeNN

Model variables

User-defined model variables originate from classes derived off the NeuronModels::Base, WeightUpdateModels::Base or PostsynapticModels::Base classes. The name of model variable is defined in the model type, i.e. with a statement such as

When a neuron or synapse population using this model is added to the model, the full GeNN name of the variable will be obtained by concatenating the variable name with the name of the population. For example if we add a population called Pop using a model which contains our V variable, a variable VPop of type scalar* will be available in the global namespace of the simulation program. GeNN will pre-allocate this C array to the correct size of elements corresponding to the size of the neuron population. Users can otherwise manipulate these variable arrays as they wish.

For convenience, GeNN provides functions to copy each state variable from the device into host memory and vice versa e.g. pullVPopFromDevice() and pushVPoptoDevice() . Alternatively, all state variables associated with a population can be copied using a single call E.g. These conventions also apply to the the variables of postsynaptic and weight update models.

Built-in Variables in GeNN

GeNN has no explicitly hard-coded synapse and neuron variables. Users are free to name the variable of their models as they want. However, there are some reserved variables that are used for intermediary calculations and communication between different parts of the generated code. They can be used in the user defined code but no other variables should be defined with these names.

DT : Time step (typically in ms) for simulation; Neuron integration can be done in multiple sub-steps inside the neuron model for numerical stability (see Traub-Miles and Izhikevich neuron model variations in Neuron models).

Note: DT exceptionally does not require bracketing with $(.)

t : The current time as a floating point value, typically interpreted as in units of ms
id : The index of a neuron in a neuron population; can be used in the context of neurons, current sources, postsynaptic models, and pre- and postsynaptic weight update models
id_syn : index of a synapse in a synapse population
id_pre : Used in a synapse context. The index of the pre-synaptic neuron in its population.
id_post : Used in a synapse context. The index of the post-synaptic neuron in its population.
inSyn : Used in the context of post_synaptic models. This is an intermediary synapse variable which contains the summed input into a postsynaptic neuron (originating from the $(addToInSyn, X) or $(addToInSynDelay, X, Y) functions of the weight update model used by incoming synapses).
Isyn : This is a local variable which contains the (summed) input current to a neuron. It is typically the sum of any explicit current input and all synaptic inputs. The way its value is calculated during the update of the postsynaptic neuron is defined by the code provided in the postsynaptic model. For example, the standard PostsynapticModels::ExpCond postsynaptic model defines
SET_APPLY_INPUT_CODE("$(Isyn) += $(inSyn)*($(E)-$(V));");

apply_input_code="$(Isyn) += $(inSyn)*($(E)-$(V));"

which implements a conductance based synapse in which the postsynaptic current is given by $I_{\rm syn}= g*s*(V_{\rm rev}-V_{\rm post})$ . The value of $(Isyn) resulting from the apply input code can then be used in neuron sim code like so:
$(V)+= (-$(V)+$(Isyn))*DT
sT : This is a neuron variable containing the spike time of each neuron and is automatically generated for pre and postsynaptic neuron groups if they are connected using a synapse population with a weight update model that has set.
prev_sT: This is a neuron variable containing the previous spike time of each neuron and is automatically generated for pre and postsynaptic neuron groups if they are connected using a synapse population with a weight update model that has set.

In addition to these variables, neuron variables can be referred to in the synapse models by calling $(<neuronVarName>_pre) for the presynaptic neuron population, and $(<neuronVarName>_post) for the postsynaptic population. For example, $(sT_pre), $(sT_post), $(V_pre), etc.

Spike Recording

Especially in models simulated with small timesteps, very few spikes may be emitted every timestep, making calling every timestep very inefficient. Instead, the spike recording system allows spikes and spike-like events emitted over a number of timesteps to be collected in GPU memory before transferring to the host. Spike recording can be enabled on chosen neuron groups with the . Remaining GPU memory can then be allocated at runtime for spike recording by . The data structures can then be copied from the GPU to the host using the and the spikes emitted by a population can be accessed Similarly, spike-like events emitted by a population can be accessed via the recordSpkEvent<neuron name> variable .

Debugging suggestions

In Linux, users can call cuda-gdb to debug on the GPU. Example projects in the userproject directory come with a flag to enable debugging (–debug). genn-buildmodel.sh has a debug flag (-d) to generate debugging data. If you are executing a project with debugging on, the code will be compiled with -g -G flags. In CPU mode the executable will be run in gdb, and in GPU mode it will be run in cuda-gdb in tui mode.

Note: Do not forget to switch debugging flags -g and -G off after debugging is complete as they may negatively affect performance.

On Mac, some versions of clang aren't supported by the CUDA toolkit. This is a recurring problem on Fedora as well, where CUDA doesn't keep up with GCC releases. You can either hack the CUDA header which checks compiler versions - cuda/include/host_config.h - or just use an older XCode version (6.4 works fine).

On Windows models can also be debugged and developed by opening the sln file used to build the model in Visual Studio. From here files can be added to the project, build settings can be adjusted and the full suite of Visual Studio debugging and profiling tools can be used.

Note: When opening the models in the userproject directory in Visual Studio, right-click on the project in the solution explorer, select 'Properties'. Then, making sure the desired configuration is selected, navigate to 'Debugging' under 'Configuration Properties', set the 'Working Directory' to '..' and the 'Command Arguments' to match those passed to genn-buildmodel e.g. 'outdir' to use an output directory called outdir.

Previous | Top | Next