Merge pull request #2911 from BenChung/improve-events

Support more of the SciMLBase events API

src/systems/callbacks.jl

@@ -76,24 +76,62 @@ end
 #################################### continuous events #####################################
 
 const NULL_AFFECT = Equation[]
+"""
+    SymbolicContinuousCallback(eqs::Vector{Equation}, affect, affect_neg, rootfind)
+
+A [`ContinuousCallback`](@ref SciMLBase.ContinuousCallback) specified symbolically. Takes a vector of equations `eq`
+as well as the positive-edge `affect` and negative-edge `affect_neg` that apply when *any* of `eq` are satisfied.
+By default `affect_neg = affect`; to only get rising edges specify `affect_neg = nothing`.
+
+Assume without loss of generality that the equation is of the form `c(u,p,t) ~ 0`; we denote the integrator state as `i.u`. 
+For compactness, we define `prev_sign = sign(c(u[t-1], p[t-1], t-1))` and `cur_sign = sign(c(u[t], p[t], t))`.
+A condition edge will be detected and the callback will be invoked iff `prev_sign * cur_sign <= 0`. 
+Inter-sample condition activation is not guaranteed; for example if we use the dirac delta function as `c` to insert a 
+sharp discontinuity between integrator steps (which in this example would not normally be identified by adaptivity) then the condition is not
+guaranteed to be triggered.
+
+Once detected the integrator will "wind back" through a root-finding process to identify the point when the condition became active; the method used 
+is specified by `rootfind` from [`SciMLBase.RootfindOpt`](@ref). Multiple callbacks in the same system with different `rootfind` operations will be resolved 
+into separate VectorContinuousCallbacks in the enumeration order of `SciMLBase.RootfindOpt`, which may cause some callbacks to not fire if several become
+active at the same instant. See the `SciMLBase` documentation for more information on the semantic rules.
+
+The positive edge `affect` will be triggered iff an edge is detected and if `prev_sign < 0`; similarly, `affect_neg` will be
+triggered iff an edge is detected `prev_sign > 0`. 
+
+Affects (i.e. `affect` and `affect_neg`) can be specified as either:
+* A list of equations that should be applied when the callback is triggered (e.g. `x ~ 3, y ~ 7`) which must be of the form `unknown ~ observed value` where each `unknown` appears only once. Equations will be applied in the order that they appear in the vector; parameters and state updates will become immediately visible to following equations.
+* A tuple `(f!, unknowns, read_parameters, modified_parameters, ctx)`, where:
+    + `f!` is a function with signature `(integ, u, p, ctx)` that is called with the integrator, a state *index* vector `u` derived from `unknowns`, a parameter *index* vector `p` derived from `read_parameters`, and the `ctx` that was given at construction time. Note that `ctx` is aliased between instances.
+    + `unknowns` is a vector of symbolic unknown variables and optionally their aliases (e.g. if the model was defined with `@variables x(t)` then a valid value for `unknowns` would be `[x]`). A variable can be aliased with a pair `x => :y`. The indices of these `unknowns` will be passed to `f!` in `u` in a named tuple; in the earlier example, if we pass `[x]` as `unknowns` then `f!` can access `x` as `integ.u[u.x]`. If no alias is specified the name of the index will be the symbol version of the variable name.
+    + `read_parameters` is a vector of the parameters that are *used* by `f!`. Their indices are passed to `f` in `p` similarly to the indices of `unknowns` passed in `u`.
+    + `modified_parameters` is a vector of the parameters that are *modified* by `f!`. Note that a parameter will not appear in `p` if it only appears in `modified_parameters`; it must appear in both `parameters` and `modified_parameters` if it is used in the affect definition.
+    + `ctx` is a user-defined context object passed to `f!` when invoked. This value is aliased for each problem.
+"""
 struct SymbolicContinuousCallback
     eqs::Vector{Equation}
     affect::Union{Vector{Equation}, FunctionalAffect}
-    function SymbolicContinuousCallback(eqs::Vector{Equation}, affect = NULL_AFFECT)
-        new(eqs, make_affect(affect))
+    affect_neg::Union{Vector{Equation}, FunctionalAffect, Nothing}
+    rootfind::SciMLBase.RootfindOpt
+    function SymbolicContinuousCallback(; eqs::Vector{Equation}, affect = NULL_AFFECT,
+            affect_neg = affect, rootfind = SciMLBase.LeftRootFind)
+        new(eqs, make_affect(affect), make_affect(affect_neg), rootfind)
     end # Default affect to nothing
 end
 make_affect(affect) = affect
 make_affect(affect::Tuple) = FunctionalAffect(affect...)
 make_affect(affect::NamedTuple) = FunctionalAffect(; affect...)
 
 function Base.:(==)(e1::SymbolicContinuousCallback, e2::SymbolicContinuousCallback)
-    isequal(e1.eqs, e2.eqs) && isequal(e1.affect, e2.affect)
+    isequal(e1.eqs, e2.eqs) && isequal(e1.affect, e2.affect) &&
+        isequal(e1.affect_neg, e2.affect_neg) && isequal(e1.rootfind, e2.rootfind)
 end
 Base.isempty(cb::SymbolicContinuousCallback) = isempty(cb.eqs)
 function Base.hash(cb::SymbolicContinuousCallback, s::UInt)
     s = foldr(hash, cb.eqs, init = s)
-    cb.affect isa AbstractVector ? foldr(hash, cb.affect, init = s) : hash(cb.affect, s)
+    s = cb.affect isa AbstractVector ? foldr(hash, cb.affect, init = s) : hash(cb.affect, s)
+    s = cb.affect_neg isa AbstractVector ? foldr(hash, cb.affect_neg, init = s) :
+        hash(cb.affect_neg, s)
+    hash(cb.rootfind, s)
 end
 
 to_equation_vector(eq::Equation) = [eq]
@@ -108,6 +146,14 @@ function SymbolicContinuousCallback(args...)
 end # wrap eq in vector
 SymbolicContinuousCallback(p::Pair) = SymbolicContinuousCallback(p[1], p[2])
 SymbolicContinuousCallback(cb::SymbolicContinuousCallback) = cb # passthrough
+function SymbolicContinuousCallback(eqs::Equation, affect = NULL_AFFECT;
+        affect_neg = affect, rootfind = SciMLBase.LeftRootFind)
+    SymbolicContinuousCallback(eqs=[eqs], affect=affect, affect_neg=affect_neg, rootfind=rootfind)
+end
+function SymbolicContinuousCallback(eqs::Vector{Equation}, affect = NULL_AFFECT;
+        affect_neg = affect, rootfind = SciMLBase.LeftRootFind)
+    SymbolicContinuousCallback(eqs=eqs, affect=affect, affect_neg=affect_neg, rootfind=rootfind)
+end
 
 SymbolicContinuousCallbacks(cb::SymbolicContinuousCallback) = [cb]
 SymbolicContinuousCallbacks(cbs::Vector{<:SymbolicContinuousCallback}) = cbs
@@ -130,12 +176,20 @@ function affects(cbs::Vector{SymbolicContinuousCallback})
     mapreduce(affects, vcat, cbs, init = Equation[])
 end
 
+affect_negs(cb::SymbolicContinuousCallback) = cb.affect_neg
+function affect_negs(cbs::Vector{SymbolicContinuousCallback})
+    mapreduce(affect_negs, vcat, cbs, init = Equation[])
+end
+
 namespace_affects(af::Vector, s) = Equation[namespace_affect(a, s) for a in af]
 namespace_affects(af::FunctionalAffect, s) = namespace_affect(af, s)
+namespace_affects(::Nothing, s) = nothing
 
 function namespace_callback(cb::SymbolicContinuousCallback, s)::SymbolicContinuousCallback
-    SymbolicContinuousCallback(namespace_equation.(equations(cb), (s,)),
-        namespace_affects(affects(cb), s))
+    SymbolicContinuousCallback(
+        namespace_equation.(equations(cb), (s,)),
+        namespace_affects(affects(cb), s),
+        namespace_affects(affect_negs(cb), s))
 end
 
 """
@@ -159,7 +213,7 @@ function continuous_events(sys::AbstractSystem)
     filter(!isempty, cbs)
 end
 
-#################################### continuous events #####################################
+#################################### discrete events #####################################
 
 struct SymbolicDiscreteCallback
     # condition can be one of:
@@ -462,12 +516,38 @@ function generate_rootfinding_callback(sys::AbstractODESystem, dvs = unknowns(sy
     isempty(cbs) && return nothing
     generate_rootfinding_callback(cbs, sys, dvs, ps; kwargs...)
 end
-
-function generate_rootfinding_callback(cbs, sys::AbstractODESystem, dvs = unknowns(sys),
+"""
+Generate a single rootfinding callback; this happens if there is only one equation in `cbs` passed to 
+generate_rootfinding_callback and thus we can produce a ContinuousCallback instead of a VectorContinuousCallback.
+"""
+function generate_single_rootfinding_callback(
+        eq, cb, sys::AbstractODESystem, dvs = unknowns(sys),
         ps = full_parameters(sys); kwargs...)
+    if !isequal(eq.lhs, 0)
+        eq = 0 ~ eq.lhs - eq.rhs
+    end
+
+    rf_oop, rf_ip = generate_custom_function(
+        sys, [eq.rhs], dvs, ps; expression = Val{false}, kwargs...)
+    affect_function = compile_affect_fn(cb, sys, dvs, ps, kwargs)
+    cond = function (u, t, integ)
+        if DiffEqBase.isinplace(integ.sol.prob)
+            tmp, = DiffEqBase.get_tmp_cache(integ)
+            rf_ip(tmp, u, parameter_values(integ), t)
+            tmp[1]
+        else
+            rf_oop(u, parameter_values(integ), t)
+        end
+    end
+    return ContinuousCallback(
+        cond, affect_function.affect, affect_function.affect_neg, rootfind = cb.rootfind)
+end
+
+function generate_vector_rootfinding_callback(
+        cbs, sys::AbstractODESystem, dvs = unknowns(sys),
+        ps = full_parameters(sys); rootfind = SciMLBase.RightRootFind, kwargs...)
     eqs = map(cb -> flatten_equations(cb.eqs), cbs)
     num_eqs = length.(eqs)
-    (isempty(eqs) || sum(num_eqs) == 0) && return nothing
     # fuse equations to create VectorContinuousCallback
     eqs = reduce(vcat, eqs)
     # rewrite all equations as 0 ~ interesting stuff
@@ -477,45 +557,99 @@ function generate_rootfinding_callback(cbs, sys::AbstractODESystem, dvs = unknow
     end
 
     rhss = map(x -> x.rhs, eqs)
-    root_eq_vars = unique(collect(Iterators.flatten(map(ModelingToolkit.vars, rhss))))
+    _, rf_ip = generate_custom_function(
+        sys, rhss, dvs, ps; expression = Val{false}, kwargs...)
+
+    affect_functions = @NamedTuple{affect::Function, affect_neg::Union{Function, Nothing}}[compile_affect_fn(
+                                                                                               cb,
+                                                                                               sys,
+                                                                                               dvs,
+                                                                                               ps,
+                                                                                               kwargs)
+                                                                                           for cb in cbs]
+    cond = function (out, u, t, integ)
+        rf_ip(out, u, parameter_values(integ), t)
+    end
 
-    rf_oop, rf_ip = generate_custom_function(sys, rhss, dvs, ps; expression = Val{false},
-        kwargs...)
+    # since there may be different number of conditions and affects,
+    # we build a map that translates the condition eq. number to the affect number
+    eq_ind2affect = reduce(vcat,
+        [fill(i, num_eqs[i]) for i in eachindex(affect_functions)])
+    @assert length(eq_ind2affect) == length(eqs)
+    @assert maximum(eq_ind2affect) == length(affect_functions)
 
-    affect_functions = map(cbs) do cb # Keep affect function separate
-        eq_aff = affects(cb)
-        affect = compile_affect(eq_aff, sys, dvs, ps; expression = Val{false}, kwargs...)
+    affect = let affect_functions = affect_functions, eq_ind2affect = eq_ind2affect
+        function (integ, eq_ind) # eq_ind refers to the equation index that triggered the event, each event has num_eqs[i] equations
+            affect_functions[eq_ind2affect[eq_ind]].affect(integ)
+        end
     end
-
-    if length(eqs) == 1
-        cond = function (u, t, integ)
-            if DiffEqBase.isinplace(integ.sol.prob)
-                tmp, = DiffEqBase.get_tmp_cache(integ)
-                rf_ip(tmp, u, parameter_values(integ), t)
-                tmp[1]
-            else
-                rf_oop(u, parameter_values(integ), t)
+    affect_neg = let affect_functions = affect_functions, eq_ind2affect = eq_ind2affect
+        function (integ, eq_ind) # eq_ind refers to the equation index that triggered the event, each event has num_eqs[i] equations
+            affect_neg = affect_functions[eq_ind2affect[eq_ind]].affect_neg
+            if isnothing(affect_neg)
+                return # skip if the neg function doesn't exist - don't want to split this into a separate VCC because that'd break ordering
             end
+            affect_neg(integ)
         end
-        ContinuousCallback(cond, affect_functions[])
+    end
+    return VectorContinuousCallback(
+        cond, affect, affect_neg, length(eqs), rootfind = rootfind)
+end
+
+"""
+Compile a single continuous callback affect function(s).
+"""
+function compile_affect_fn(cb, sys::AbstractODESystem, dvs, ps, kwargs)
+    eq_aff = affects(cb)
+    eq_neg_aff = affect_negs(cb)
+    affect = compile_affect(eq_aff, sys, dvs, ps; expression = Val{false}, kwargs...)
+    if eq_neg_aff === eq_aff
+        affect_neg = affect
+    elseif isnothing(eq_neg_aff)
+        affect_neg = nothing
     else
-        cond = function (out, u, t, integ)
-            rf_ip(out, u, parameter_values(integ), t)
+        affect_neg = compile_affect(
+            eq_neg_aff, sys, dvs, ps; expression = Val{false}, kwargs...)
+    end
+    (affect = affect, affect_neg = affect_neg)
+end
+
+function generate_rootfinding_callback(cbs, sys::AbstractODESystem, dvs = unknowns(sys),
+        ps = full_parameters(sys); kwargs...)
+    eqs = map(cb -> flatten_equations(cb.eqs), cbs)
+    num_eqs = length.(eqs)
+    total_eqs = sum(num_eqs)
+    (isempty(eqs) || total_eqs == 0) && return nothing
+    if total_eqs == 1
+        # find the callback with only one eq
+        cb_ind = findfirst(>(0), num_eqs)
+        if isnothing(cb_ind)
+            error("Inconsistent state in affect compilation; one equation but no callback with equations?")
         end
+        cb = cbs[cb_ind]
+        return generate_single_rootfinding_callback(cb.eqs[], cb, sys, dvs, ps; kwargs...)
+    end
 
-        # since there may be different number of conditions and affects,
-        # we build a map that translates the condition eq. number to the affect number
-        eq_ind2affect = reduce(vcat,
-            [fill(i, num_eqs[i]) for i in eachindex(affect_functions)])
-        @assert length(eq_ind2affect) == length(eqs)
-        @assert maximum(eq_ind2affect) == length(affect_functions)
+    # group the cbs by what rootfind op they use
+    # groupby would be very useful here, but alas
+    cb_classes = Dict{
+        @NamedTuple{rootfind::SciMLBase.RootfindOpt}, Vector{SymbolicContinuousCallback}}()
+    for cb in cbs
+        push!(
+            get!(() -> SymbolicContinuousCallback[], cb_classes, (rootfind = cb.rootfind,)),
+            cb)
+    end
 
-        affect = let affect_functions = affect_functions, eq_ind2affect = eq_ind2affect
-            function (integ, eq_ind) # eq_ind refers to the equation index that triggered the event, each event has num_eqs[i] equations
-                affect_functions[eq_ind2affect[eq_ind]](integ)
-            end
-        end
-        VectorContinuousCallback(cond, affect, length(eqs))
+    # generate the callbacks out; we sort by the equivalence class to ensure a deterministic preference order
+    compiled_callbacks = map(collect(pairs(sort!(
+        OrderedDict(cb_classes); by = p -> p.rootfind)))) do (equiv_class, cbs_in_class)
+        return generate_vector_rootfinding_callback(
+            cbs_in_class, sys, dvs, ps; rootfind = equiv_class.rootfind, kwargs...)
+    end
+    if length(compiled_callbacks) == 1
+        return compiled_callbacks[]
+    else
+        return CallbackSet(compiled_callbacks...)
     end
 end
 
@@ -529,7 +663,6 @@ function compile_user_affect(affect::FunctionalAffect, sys, dvs, ps; kwargs...)
         ps_ind = Dict(reverse(en) for en in enumerate(ps))
         p_inds = map(sym -> ps_ind[sym], parameters(affect))
     end
-
     # HACK: filter out eliminated symbols. Not clear this is the right thing to do
     # (MTK should keep these symbols)
     u = filter(x -> !isnothing(x[2]), collect(zip(unknowns_syms(affect), v_inds))) |>