NMODL¶
Name |
File extension |
Read |
Write |
---|---|---|---|
NMODL |
|
✓ |
✗ |
NMODL is a DSL for describing ion channel and synapse dynamics that is used by NEURON, which provides the mod2c compiler parses dynamics described in NMODL to generate C code that is called from NEURON.
Arbor has an NMODL compiler, modcc, that generates optimized code in C++ and CUDA, which is optimized for the target architecture. NMODL does not have a formal specification, and its semantics are often ambiguous. To manage this, Arbor uses its own dialect of NMODL that does not allow some constructions used in NEURON’s NMODL.
Note
We hope to replace NMODL with a DSL that is well defined, and easier for both users and the Arbor developers to work with in the long term. Until then, please write issues on our GitHub with any questions that you have about getting your NMODL files to work in Arbor.
This page is a collection of NMODL rules for Arbor. It assumes that the reader already has a working knowledge of NMODL.
Units¶
Arbor doesn’t support unit conversion in NMODL. This table lists the key NMODL quantities and their expected units.
quantity |
identifier |
unit |
---|---|---|
voltage |
v / v_peer |
mV |
temperature |
celsius |
°C |
diameter (cross-sectional) |
diam |
µm |
current_density (density mechanisms) |
identifier defined using |
mA/cm² |
conductivity (density mechanisms) |
identifier inferred from current_density equation
e.g. in |
S/cm² |
current (point and junction mechanisms) |
identifier defined using |
nA |
conductance (point and junction mechanisms) |
identifier inferred from current equation
e.g. in |
µS |
ion X current_density (density mechanisms) |
iX |
mA/cm² |
ion X current (point and junction mechanisms) |
iX |
nA |
ion X reversal potential |
eX |
mV |
ion X internal concentration |
Xi |
mmol/L |
ion X external concentration |
Xo |
mmol/L |
ion X diffusive concentration |
Xd |
mmol/L |
Ions¶
Arbor recognizes
na
,ca
andk
ions by default. Any new ions used in NMODL need to be explicitly added into Arbor along with their default properties and valence (this can be done in the recipe or on a single cell model). Simply specifying them in NMODL will not work.The parameters and variables of each ion referenced in a
USEION
statement are available automatically to the mechanism. The exposed variables are: internal concentrationXi
, external concentrationXo
, reversal potentialeX
and currentiX
. It is an error to also mark these asPARAMETER
,ASSIGNED
orCONSTANT
.READ
andWRITE
permissions ofXi
,Xo
,eX
andiX
can be set in NMODL in theNEURON
block. If a parameter is writable it is automatically readable and must not be specified as both.If
Xi
,Xo
,eX
,iX
,Xd
are used in aPROCEDURE
orFUNCTION
, they need to be passed as arguments.If
Xi
orXo
(internal and external concentrations) are written in the NMODL mechanism they need to be declared asSTATE
variables and their initial values have to be set in theINITIAL
block in the mechanism. This transfers all responsibility for handlingXi
/Xo
to the mechanism and will lead to painted initial values to be ignored. If these quantities are not madeSTATE
they may be written to, but their values will be reset to their initial values every time step.The diffusive concentration
Xd
does not share this semantics. It will not be reset, even if not inSTATE
, and may freely be written. This comes at the cost of awkward treatment of ODEs forXd
, see the includeddecay.mod
for an example.Xd
is present on all cables iff its associated diffusivity is set to a non-zero value.
Special variables¶
Arbor exposes some parameters from the simulation to the NMODL mechanisms. These include
v
,diam
, andcelsius
in addition to the previously mentioned ion parameters.These special variables should not be
ASSIGNED
orCONSTANT
, they arePARAMETER
. This is different from NEURON where a built-in variable is declaredASSIGNED
to make it accessible.diam
andcelsius
are set from the simulation side.v
is a reserved variable name and can be read but not written in NMODL.dt
,time
, andarea
are not exposed to NMODL mechanisms.NONSPECIFIC_CURRENTS
should not bePARAMETER
,ASSIGNED
orCONSTANT
. They just need to be declared in the NEURON block.
Functions, procedures and blocks¶
SOLVE
statements should be the first statement in theBREAKPOINT
block.The return variable of
FUNCTION
has to always be set.if
without associatedelse
can break that if users are not careful.Any non-
LOCAL
variables used in aPROCEDURE
orFUNCTION
need to be passed as arguments.
Unsupported features¶
Unit conversion is not supported in Arbor (there is limited support for parsing units, which are just ignored).
Unit declaration is not supported (ex:
FARADAY = (faraday) (10000 coulomb)
). They can be replaced by declaring them and setting their values inCONSTANT
.FROM
-TO
clamping of variables is not supported. The tokens are parsed, and reported through themechanism_info
, but otherwise ignored. However,CONSERVE
statements are supported.TABLE
is not supported, calculations are exact.derivimplicit
solving method is not supported, usecnexp
instead.VERBATIM
blocks are not supported.LOCAL
variables outside blocks are not supported.INDEPENDENT
variables are not supported.
Arbor-specific features¶
It is required to explicitly pass ‘magic’ variables like
v
into procedures. It makes things more explicit by eliding shared and implicit global state. However, this is only partially true, as having PARAMETER v brings it into scope, but only in BREAKPOINT.Arbor’s NMODL dialect supports the most widely used features of NEURON. It also has some features unavailable in NEURON such as the
POST_EVENT
procedure block. This procedure has a single argument representing the time since the last spike on the cell. In the event of multiple detectors on the cell, and multiple spikes on the detectors within the same integration period, the times of each of these spikes will be processed by thePOST_EVENT
block. Spikes are processed only once and then cleared.Example of a
POST_EVENT
procedure, whereg
is aSTATE
parameter representing the conductance:POST_EVENT(t) { g = g + (0.1*t) }
Arbor allows a gap-junction mechanism to access the membrane potential at the peer site of a gap-junction connection as well as the local site. The peer membrane potential is made available through the
v_peer
variable while the local membrane potential is available throughv
, as usual.Arbor offers a number of additional unary math functions which may offer improved performance compared to hand-rolled solutions (especially with the vectorized and GPU backends). All of the following functions take a single argument x and return a floating point value.
Function name
Description
Semantics
sqrt(x)
square root
\(\sqrt{x}\)
step_right(x)
right-continuous heaviside step
\(\begin{align*} 1 & ~~ \text{if} ~x \geq 0, \\ 0 & ~~ \text{otherwise}. \end{align*}\)
step_left(x)
left-continuous heaviside step
\(\begin{align*} 1 & ~~ \text{if} ~x \gt 0, \\ 0 & ~~ \text{otherwise}. \end{align*}\)
step(x)
heaviside step with half value
\(\begin{align*} 1 & ~~ \text{if} ~x \gt 0, \\ 0 & ~~ \text{if} ~x \lt 0, \\ 0.5 & ~~ \text{otherwise}. \end{align*}\)
signum(x)
sign of argument
\(\begin{align*} +1 & ~~ \text{if} ~x \gt 0, \\ -1 & ~~ \text{if} ~x \lt 0, \\ 0 & ~~ \text{otherwise}. \end{align*}\)
exprelr(x)
smooth continuation over \(x=0\) of
\(x/(1 - e^{-x})\)
sigmoid(x)
sigmoidal function
\(\frac{1}{1+e^{-x}}\)
relu(x)
rectified linear function
\(max(0, x)\)
tanh(x)
hyperbolic tangent
\(tanh(x)\)
Voltage Processes¶
Some cases require direct manipulation of the membrane voltage v
; which is
normally prohibited and for good reason so. For these limited application,
however, we offer mechanisms that are similar to density
mechanism, but are
tagged with VOLTAGE_PROCESS
where normally SUFFIX
would be used.
This is both a very sharp tool and a somewhat experimental feature. Depending on
our experience, it might be changed or removed. Using a VOLTAGE_PROCESS
,
voltage clamping and limiting can be implemented, c.f. relevant examples in the
default
catalogue. Example: limiting membrane voltage from above and below
NEURON {
VOLTAGE_PROCESS v_limit
GLOBAL v_low, v_high
}
PARAMETER {
v_high = 20 (mV)
v_low = -70 (mV)
}
BREAKPOINT {
v = max(min(v, v_high), v_low)
}
As of the current implementation, we note the following details and constraints
only the
INITIAL
andBREAKPOINT
procedures are called.no
WRITE
access to ionic quantities is allowed.only one
VOLTAGE_PROCESS
maybe present on a single location, adding more results in an exception.the
BREAKPOINT
callback will execute _after_ the cable solver. A consequence of this is that if the initial membrane potential \(V_0\) is unequal to that of a potentially applied voltage clamp \(V_c\), the first timestep will observe \(V_0\).
Stochastic Processes¶
Arbor supports stochastic processes in the form of stochastic differential
equations. The white noise sources can be defined in the model files using a WHITE_NOISE
block:
WHITE_NOISE {
a b
c
}
Arbitrary white noise variables can be declared (a, b, c
in the example above). The
noise will be appropriately scaled with the numerical time step and can be considered unitless. In
order to influence the white noise generation, a seed value can be set at the level of the
simulation through the optional constructor argument seed
(see here or here).
If the state is updated by involving at least one of the declared white noise variables the system is considered to be stochastic:
DERIVATIVE state {
s' = f + g*a
}
The solver method must then accordingly set to stochastic
:
BREAKPOINT {
SOLVE state METHOD stochastic
}
Nernst¶
Many mechanisms make use of the reversal potential of an ion (eX
for ion X
).
A popular equation for determining the reversal potential during the simulation is
the Nernst equation.
Both Arbor and NEURON make use of nernst
. Arbor implements it as a mechanism and
NEURON implements it as a built-in method. However, the conditions for using the
nernst
equation to change the reversal potential of an ion differ between the
two simulators.
1. In Arbor, the reversal potential of an ion remains equal to its initial value (which
has to be set by the user) over the entire course of the simulation, unless another
mechanism which alters that reversal potential (such as nernst
) is explicitly selected
for the entire cell. (see Reversal potential dynamics for details).
2. In NEURON, there is a rule which is evaluated (under the hood) per section of a given
cell to determine whether or not the reversal potential of an ion remains constant or is
calculated using nernst
. The rule is documented
here
and can be summarized as follows:
Examining all mechansims on a given section, if the internal or external concentration of an ion is written, and its reversal potential is read but not written, then the nernst equation is used continuously during the simulation to update the reversal potential of the ion. And if the internal or external concentration of an ion is read, and its reversal potential is read but not written, then the nernst equation is used once at the beginning of the simulation to caluclate the reversal potential of the ion, and then remains constant. Otherwise, the reversal potential is set by the user and remains constant.
One of the main consequences of this difference in behavior is that in Arbor, a mechanism
modifying the reversal potential (for example nernst
) can only be applied (for a given ion)
at a global level on a given cell. While in Neuron, different mechanisms can be used for
calculating the reversal potential of an ion on different parts of the morphology.
This is due to the different methods Arbor and NEURON use for discretising the morphology.
(A region
in Arbor may include part of a CV, where as in NEURON, a section
can only
contain full segments
).
Modelers are encouraged to verify the expected behavior of the reversal potentials of ions as it can lead to vastly different model behavior.
Tips for Faster NMODL¶
Note
If you are looking for help with NMODL in the context of NEURON this guide might not help.
NMODL is a language without formal specification and many unexpected characteristics (many of which are not supported in Arbor), which results in existing NMODL files being treated as difficult to understand and best left as-is. This in turn leads to sub-optimal performance, especially since mechanisms take up a large amount of the simulations’ runtime budget. With some understanding of the subject matter, however, it is quite straightforward to obtain clean and performant NMODL files. We regularly have seen speed-ups factors of roughly three from optimising NMODL.
First, let us discuss how NMODL becomes part of an Arbor simulation. NMODL
mechanisms are given in .mod
files, whose layout and syntax has been
discussed above. These are compiled by modcc
into a series of callbacks as
specified by the Mechanism ABI. These operate on data held in Arbor’s
internal storage. But, modcc
does not generate machine code, it goes through
C++ (and/or CUDA) as an intermediary which is processed by a standard C++
compiler like GCC (or nvcc) to produce either a shared object (for external
catalogues) and code directly linked into Arbor (the built-in catalogues).
Now, we turn to a series of tips we found helpful in producing fast NMODL mechanisms. In terms of performance of variable declaration, the hierarchy is from slowest to fastest:
RANGE ASSIGNED
– mutable arrayRANGE PARAMETER
– configurable arrayASSIGNED
– mutablePARAMETER
– configurableCONSTANT
– inlined constant
RANGE
¶
Parameters and ASSIGNED
variables marked as RANGE
will be stored as an
array with one entry per CV in Arbor. Reading and writing these incurs a memory
access and thus affects cache and memory utilisation metrics. It is often more
efficient to use LOCAL
variables instead, even if that means foregoing the
ability to re-use a computed value. Compute is so much faster than memory on
modern hardware that re-use at the expense of memory accesses is seldom
profitable, except for the most complex terms. LOCAL
variables become just
that in the generated code: a local variable that is likely residing in a
register and used only as long as needed.
PROCEDURE
¶
Prefer FUNCTION
over PROCEDURE
. The latter require ASSIGNED RANGE
variables to return values and thus stress the memory system, which, as noted
above, is not most efficient on current hardware. Also, they may not be inlined,
as opposed to a FUNCTION
.
PARAMETER
¶
PARAMETER
should only be used for values that must be set by the simulator.
All fixed values should be CONSTANT
instead. These will be inlined by
modcc
and propagated through the computations which can uncover more
optimisation potential.
Complex Expressions in Current Computation¶
modcc
, Arbor’s NMODL compiler, applies symbolic differentiation to the
current expression to find the conductance as g = d I/d U
which are then
used to compute the voltage update. g
is thus computed multiple times every
timestep and if the corresponding expression is inefficient, it will cost more
time than needed. The differentiation implementation quite naive and will not
optimise the resulting expressions. This is an internal detail of Arbor and
might change in the future, but for now this particular optimisation can help to
produce better performing code. Here is an example
: BAD, will compute m^4 * h every step
i = m^4 * h * (v - e)
: GOOD, will just use a constant value of g
LOCAL g
g = m^4 * h
i = g * (v - e)
Note that we do not lose accuracy here, since Arbor does not support
higher-order ODEs and thus will treat g
as a constant across
a single timestep even if g
actually depends on v
.
Using Memory versus Computation¶
Commonly ion channels need to correct for temperature differences, which yields a term similar to
q = 3^(0.1*celsius - 0.63)
Here, we find that the cost of the exponential when computing q
in the
DERIVATIVE
block is high enough to make pre-computing q
in INITIAL
and loading the value later an optimisation. Shown below is a simplified version
of this pattern from hh.mod
in the Arbor sources
NEURON {
...
RANGE ..., q
}
ASSIGNED { q }
PARAMETER {
...
celsius (degC)
}
STATE { ... }
BREAKPOINT {
SOLVE dS METHOD cnexp
...
}
INITIAL {
q = 3^(0.1*celsius - 0.63)
...
}
DERIVATIVE states {
... : uses q
}
Specialised Functions¶
Some extra cost can be saved by choosing Arbor-specific optimized math functions instead of hand-rolled versions. Please consult the table in this section. A common pattern is the use of a guarded exponential of the form
if (x != 0) {
r = a*x/(exp(-x) - 1)
} else {
r = a
}
However, it can be written in Arbor’s NMODL dialect as
exprelr(x)
which is more efficient and has the same guarantees. NMODL files originating
from NEURON often use this or related functions, e.g. vtrap(x, y) =
y*exprelr(x/y)
.
Small Tips and Micro-Optimisations¶
Divisions cost a bit more than multiplications and additions.
m * m
is more efficient thanm^2
. This holds for higher powers as well and if you want to squeeze out the utmost of performance use exponentiation-by-squaring. (Although GCC does this for you. Most of the time.)