The NMODL Framework

WARNING

---

NMODL has been fully integrated into the NEURON repository. There will be no further development efforts on NMODL as an independent project.

All future development will happen at: https://github.com/neuronsimulator/nrn <https://github.com/neuronsimulator/nrn>_.

---

The NMODL Framework is a code generation engine for N\ EURON MOD\ eling L\ anguage (NMODL <https://www.neuron.yale.edu/neuron/static/py_doc/modelspec/programmatic/mechanisms/nmodl.html>__). It is designed with modern compiler and code generation techniques to:

• Provide modular tools for parsing, analysing and transforming NMODL
• Provide easy to use, high level Python API
• Generate optimised code for modern compute architectures including CPUs, GPUs
• Flexibility to implement new simulator backends
• Support for full NMODL specification

About NMODL

Simulators like NEURON <https://www.neuron.yale.edu/neuron/>__ use NMODL as a domain specific language (DSL) to describe a wide range of membrane and intracellular submodels. Here is an example of exponential synapse specified in NMODL:

.. code::

NEURON { POINT_PROCESS ExpSyn RANGE tau, e, i NONSPECIFIC_CURRENT i } UNITS { (nA) = (nanoamp) (mV) = (millivolt) (uS) = (microsiemens) } PARAMETER { tau = 0.1 (ms) <1e-9,1e9> e = 0 (mV) } ASSIGNED { v (mV) i (nA) } STATE { g (uS) } INITIAL { g = 0 } BREAKPOINT { SOLVE state METHOD cnexp i = g*(v - e) } DERIVATIVE state { g' = -g/tau } NET_RECEIVE(weight (uS)) { g = g + weight }

Installation

See INSTALL.rst <https://github.com/BlueBrain/nmodl/blob/master/INSTALL.rst>__ for detailed instructions to build the NMODL from source.

Try NMODL with Docker

To quickly test the NMODL Framework’s analysis capabilities we provide a docker <https://www.docker.com>__ image, which includes the NMODL Framework python library and a fully functional Jupyter notebook environment. After installing docker <https://docs.docker.com/compose/install/>__ and docker-compose <https://docs.docker.com/compose/install/>__ you can pull and run the NMODL image from your terminal.

To try Python interface directly from CLI, you can run docker image as:

::

docker run -it --entrypoint=/bin/sh bluebrain/nmodl

And try NMODL Python API discussed later in this README as:

::

$ python3 Python 3.6.8 (default, Apr 8 2019, 18:17:52)

from nmodl import dsl import os examples = dsl.list_examples() nmodl_string = dsl.load_example(examples[-1]) ...

To try Jupyter notebooks you can download docker compose file and run it as:

.. code:: sh

wget "https://raw.githubusercontent.com/BlueBrain/nmodl/master/docker/docker-compose.yml" DUID=$(id -u) DGID=$(id -g) HOSTNAME=$(hostname) docker-compose up

If all goes well you should see at the end status messages similar to these:

::

[I 09:49:53.923 NotebookApp] The Jupyter Notebook is running at: [I 09:49:53.923 NotebookApp] http://(4c8edabe52e1 or 127.0.0.1):8888/?token=a7902983bad430a11935 [I 09:49:53.923 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation). To access the notebook, open this file in a browser: file:///root/.local/share/jupyter/runtime/nbserver-1-open.html Or copy and paste one of these URLs: http://(4c8edabe52e1 or 127.0.0.1):8888/?token=a7902983bad430a11935

Based on the example above you should then open your browser and navigate to the URL http://127.0.0.1:8888/?token=a7902983bad430a11935.

You can open and run all example notebooks provided in the examples folder. You can also create new notebooks in my_notebooks, which will be stored in a subfolder notebooks at your current working directory.

Using the Python API

Once the NMODL Framework is installed, you can use the Python parsing API to load NMOD file as:

.. code:: python

from nmodl import dsl

examples = dsl.list_examples() nmodl_string = dsl.load_example(examples[-1]) driver = dsl.NmodlDriver() modast = driver.parse_string(nmodl_string)

The parse_file API returns Abstract Syntax Tree (AST <https://en.wikipedia.org/wiki/Abstract_syntax_tree>__) representation of input NMODL file. One can look at the AST by converting to JSON form as:

.. code:: python

print (dsl.to_json(modast)) { "Program": [ { "NeuronBlock": [ { "StatementBlock": [ { "Suffix": [ { "Name": [ { "String": [ { "name": "POINT_PROCESS" } ...

Every key in the JSON form represent a node in the AST. You can also use visualization API to look at the details of AST as:

::

from nmodl import ast ast.view(modast)

which will open AST view in web browser:

.. figure:: https://user-images.githubusercontent.com/666852/57329449-12c9a400-7114-11e9-8da5-0042590044ec.gif :alt: ast_viz

Vizualisation of the AST in the NMODL Framework

The central Program node represents the whole MOD file and each of it’s children represent the block in the input NMODL file. Note that this requires X-forwarding if you are using the Docker image.

Once the AST is created, one can use exisiting visitors to perform various analysis/optimisations. One can also easily write his own custom visitor using Python Visitor API. See Python API tutorial <docs/notebooks/nmodl-python-tutorial.ipynb>__ for details.

The NMODL Framework also allows us to transform the AST representation back to NMODL form as:

.. code:: python

print (dsl.to_nmodl(modast)) NEURON { POINT_PROCESS ExpSyn RANGE tau, e, i NONSPECIFIC_CURRENT i }

UNITS { (nA) = (nanoamp) (mV) = (millivolt) (uS) = (microsiemens) }

PARAMETER { tau = 0.1 (ms) <1e-09,1000000000> e = 0 (mV) } ...

High Level Analysis and Code Generation

The NMODL Framework provides rich model introspection and analysis capabilities using various visitors <https://bluebrain.github.io/nmodl/html/doxygen/group__visitor__classes.html>. Here is an example of theoretical performance characterisation of channels and synapses from rat neocortical column microcircuit published in 2015 <https://www.cell.com/cell/fulltext/S0092-8674%2815%2901191-5>:

.. figure:: https://user-images.githubusercontent.com/666852/57336711-2cc0b200-7127-11e9-8053-8f662e2ec191.png :alt: nmodl-perf-stats

Performance results of the NMODL Framework

To understand how you can write your own introspection and analysis tool, see this tutorial <docs/notebooks/nmodl-python-tutorial.ipynb>__.

Once analysis and optimization passes are performed, the NMODL Framework can generate optimised code for modern compute architectures including CPUs (Intel, AMD, ARM) and GPUs (NVIDIA, AMD) platforms. For example, C++, OpenACC and OpenMP backends are implemented and one can choose these backends on command line as:

::

$ nmodl expsyn.mod sympy --analytic

To know more about code generation backends, see here <https://bluebrain.github.io/nmodl/html/doxygen/group__codegen__backends.html>__. NMODL Framework provides number of options (for code generation, optimization passes and ODE solver) which can be listed as:

::

$ nmodl -H NMODL : Source-to-Source Code Generation Framework [version] Usage: /path/<>/nmodl [OPTIONS] file... [SUBCOMMAND]

Positionals: file TEXT:FILE ... REQUIRED One or more MOD files to process

Options: -h,--help Print this help message and exit -H,--help-all Print this help message including all sub-commands --verbose=info Verbose logger output (trace, debug, info, warning, error, critical, off) -o,--output TEXT=. Directory for backend code output --scratch TEXT=tmp Directory for intermediate code output --units TEXT=/path/<>/nrnunits.lib Directory of units lib file

Subcommands: host HOST/CPU code backends Options: --c C/C++ backend (true)

acc Accelerator code backends Options: --oacc C/C++ backend with OpenACC (false)

sympy SymPy based analysis and optimizations Options: --analytic Solve ODEs using SymPy analytic integration (false) --pade Pade approximation in SymPy analytic integration (false) --cse CSE (Common Subexpression Elimination) in SymPy analytic integration (false) --conductance Add CONDUCTANCE keyword in BREAKPOINT (false)

passes Analyse/Optimization passes Options: --inline Perform inlining at NMODL level (false) --unroll Perform loop unroll at NMODL level (false) --const-folding Perform constant folding at NMODL level (false) --localize Convert RANGE variables to LOCAL (false) --global-to-range Convert GLOBAL variables to RANGE (false) --localize-verbatim Convert RANGE variables to LOCAL even if verbatim block exist (false) --local-rename Rename L