Hi.
I’d like to know if there is a way to convert a function to a different form without the actual evaluation.
For example, I’m trying to make an API of DifferentialEquations.jl. However, I encountered a serious performance degradation. The following code is an example of my API.
function DifferentialEquations.ODEProblem(env::AbstractEnv, f, x0, tspan; kwargs...)
warn_is_registered(env)
_x0 = raw(env, x0)
_f(_x, p, t) = raw(env, f(readable(env, _x), p, t))
ODEProblem(_f, _x0, tspan; kwargs...)
end
In the above code, x0
denotes NamedTuple-valued (i.e., readable) initial condition.
_x0
is a vector-valued (i.e., raw) initial condition.
From this line, _f(_x, p, t) = raw(env, f(readable(env, _x), p, t))
, you may notice that the argument f
of the extended function DifferentialEquations.ODEProblem(env::AbstractEnv, f, x0, tspan; kwargs...)
will receive a “readable” state (x
) and provide a “readable” output of its derivative.
For the compatibility with DifferentialEquations.jl, _f
, which is the injected dynamical equations into ODEProblem
provided by DifferentialEquations.jl, should provide a “raw” output, so the output of f(readable(env, _x), p, t)
is wrapped by raw
.
In this case, I realised that every call of _f
would require two processings, raw
and readable
, which drastically degrade the performance. So, I wonder if there is a way of parsing the “readable” dynamical equations, f
, to the “raw” dynamical equations, _f
, at the beginning, not every time it is called.