analysis.jl

# By default, Julia/LLVM does not use fused multiply-add operations (FMAs).
# Since these FMAs can increase the performance of many numerical algorithms,
# we need to opt-in explicitly.
# See https://ranocha.de/blog/Optimizing_EC_Trixi for further details.
@muladd begin
#! format: noindent

# TODO: Taal refactor
# - analysis_interval part as PeriodicCallback called after a certain amount of simulation time
"""
    AnalysisCallback(semi; interval=0,
                           save_analysis=false,
                           output_directory="out",
                           analysis_filename="analysis.dat",
                           extra_analysis_errors=Symbol[],
                           extra_analysis_integrals=())

Analyze a numerical solution every `interval` time steps and print the
results to the screen. If `save_analysis`, the results are also saved in
`joinpath(output_directory, analysis_filename)`.

Additional errors can be computed, e.g. by passing
`extra_analysis_errors = (:l2_error_primitive, :linf_error_primitive)`
or `extra_analysis_errors = (:conservation_error,)`.

If you want to omit the computation (to safe compute-time) of the [`default_analysis_errors`](@ref), specify
`analysis_errors = Symbol[]`.
Note: `default_analysis_errors` are `:l2_error` and `:linf_error` for all equations.
If you want to compute `extra_analysis_errors` such as `:conservation_error` solely, i.e.,
without `:l2_error, :linf_error` you need to specify
`analysis_errors = [:conservation_error]` instead of `extra_analysis_errors = [:conservation_error]`.

Further scalar functions `func` in `extra_analysis_integrals` are applied to the numerical
solution and integrated over the computational domain. Some examples for this are
[`entropy`](@ref), [`energy_kinetic`](@ref), [`energy_internal`](@ref), and [`energy_total`](@ref).
You can also write your own function with the same signature as the examples listed above and
pass it via `extra_analysis_integrals`.
See the developer comments about `Trixi.analyze`, `Trixi.pretty_form_utf`, and
`Trixi.pretty_form_ascii` for further information on how to create custom analysis quantities.

In addition, the analysis callback records and outputs a number of quantities that are useful for
evaluating the computational performance, such as the total runtime, the performance index
(time/DOF/rhs!), the time spent in garbage collection (GC), or the current memory usage (alloc'd
memory).
"""
mutable struct AnalysisCallback{Analyzer, AnalysisIntegrals, InitialStateIntegrals,
                                Cache}
    start_time::Float64
    start_time_last_analysis::Float64
    ncalls_rhs_last_analysis::Int
    start_gc_time::Float64
    const interval::Int
    const save_analysis::Bool
    const output_directory::String
    const analysis_filename::String
    const analyzer::Analyzer
    const analysis_errors::Vector{Symbol}
    const analysis_integrals::AnalysisIntegrals
    initial_state_integrals::InitialStateIntegrals
    const cache::Cache
end

# TODO: Taal bikeshedding, implement a method with less information and the signature
# function Base.show(io::IO, analysis_callback::AnalysisCallback)
# end
function Base.show(io::IO, ::MIME"text/plain",
                   cb::DiscreteCallback{<:Any, <:AnalysisCallback})
    @nospecialize cb # reduce precompilation time

    if get(io, :compact, false)
        show(io, cb)
    else
        analysis_callback = cb.affect!

        setup = Pair{String, Any}["interval" => analysis_callback.interval,
                                  "analyzer" => analysis_callback.analyzer]
        for (idx, error) in enumerate(analysis_callback.analysis_errors)
            push!(setup, "│ error " * string(idx) => error)
        end
        for (idx, integral) in enumerate(analysis_callback.analysis_integrals)
            push!(setup, "│ integral " * string(idx) => integral)
        end
        push!(setup,
              "save analysis to file" => analysis_callback.save_analysis ? "yes" : "no")
        if analysis_callback.save_analysis
            push!(setup, "│ filename" => analysis_callback.analysis_filename)
            push!(setup,
                  "│ output directory" => abspath(normpath(analysis_callback.output_directory)))
        end
        summary_box(io, "AnalysisCallback", setup)
    end
end

# This is the convenience constructor that gets called from the elixirs
function AnalysisCallback(semi::AbstractSemidiscretization; kwargs...)
    mesh, equations, solver, cache = mesh_equations_solver_cache(semi)
    return AnalysisCallback(mesh, equations, solver, cache; kwargs...)
end

# This is the actual constructor
function AnalysisCallback(mesh, equations::AbstractEquations, solver, cache;
                          interval = 0,
                          save_analysis = false,
                          output_directory = "out",
                          analysis_filename = "analysis.dat",
                          extra_analysis_errors = Symbol[],
                          analysis_errors = union(default_analysis_errors(equations),
                                                  extra_analysis_errors),
                          extra_analysis_integrals = (),
                          analysis_integrals = union(default_analysis_integrals(equations),
                                                     extra_analysis_integrals),
                          RealT = real(solver),
                          uEltype = eltype(cache.elements),
                          kwargs...)
    # Decide when the callback is activated.
    # With error-based step size control, some steps can be rejected. Thus,
    #   `integrator.iter >= integrator.stats.naccept`
    #    (total #steps)       (#accepted steps)
    # We need to check the number of accepted steps since callbacks are not
    # activated after a rejected step.
    condition = (u, t, integrator) -> interval > 0 &&
        (integrator.stats.naccept % interval == 0 || isfinished(integrator))

    analyzer = SolutionAnalyzer(solver; kwargs...)
    cache_analysis = create_cache_analysis(analyzer, mesh, equations, solver, cache,
                                           RealT, uEltype)

    analysis_callback = AnalysisCallback(0.0, 0.0, 0, 0.0,
                                         interval, save_analysis, output_directory,
                                         analysis_filename,
                                         analyzer,
                                         analysis_errors, Tuple(analysis_integrals),
                                         SVector(ntuple(_ -> zero(uEltype),
                                                        Val(nvariables(equations)))),
                                         cache_analysis)

    return DiscreteCallback(condition, analysis_callback,
                            save_positions = (false, false),
                            initialize = initialize!)
end

# This method gets called from OrdinaryDiffEq's `solve(...)`
function initialize!(cb::DiscreteCallback{Condition, Affect!}, u_ode, t,
                     integrator) where {Condition, Affect! <: AnalysisCallback}
    semi = integrator.p
    du_ode = first(get_tmp_cache(integrator))
    return initialize!(cb, u_ode, du_ode, t, integrator, semi)
end

# This is the actual initialization method
# Note: we have this indirection to allow initializing a callback from the AnalysisCallbackCoupled
function initialize!(cb::DiscreteCallback{Condition, Affect!}, u_ode, du_ode, t,
                     integrator, semi) where {Condition, Affect! <: AnalysisCallback}
    initial_state_integrals = integrate(u_ode, semi)
    _, equations, _, _ = mesh_equations_solver_cache(semi)

    analysis_callback = cb.affect!
    analysis_callback.initial_state_integrals = initial_state_integrals
    @unpack save_analysis, output_directory, analysis_filename, analysis_errors, analysis_integrals = analysis_callback

    if save_analysis && mpi_isroot()
        mkpath(output_directory)

        # write header of output file
        open(joinpath(output_directory, analysis_filename), "w") do io
            print(io, "#timestep ")
            print(io, "time ")
            print(io, "dt ")
            if :l2_error in analysis_errors
                for v in varnames(cons2cons, equations)
                    print(io, "l2_" * v * " ")
                end
            end
            if :linf_error in analysis_errors
                for v in varnames(cons2cons, equations)
                    print(io, "linf_" * v * " ")
                end
            end
            if :conservation_error in analysis_errors
                for v in varnames(cons2cons, equations)
                    print(io, "cons_" * v * " ")
                end
            end
            if :residual in analysis_errors
                for v in varnames(cons2cons, equations)
                    print(io, "res_" * v * " ")
                end
            end
            if :l2_error_primitive in analysis_errors
                for v in varnames(cons2prim, equations)
                    print(io, "l2_" * v * " ")
                end
            end
            if :linf_error_primitive in analysis_errors
                for v in varnames(cons2prim, equations)
                    print(io, "linf_" * v * " ")
                end
            end

            for quantity in analysis_integrals
                print(io, pretty_form_ascii(quantity), " ")
            end

            println(io)
            return nothing
        end
    end

    # Record current time using a high-resolution clock
    analysis_callback.start_time = time_ns()

    # Record current time for performance index computation
    analysis_callback.start_time_last_analysis = time_ns()

    # Record current number of `rhs!` calls for performance index computation
    analysis_callback.ncalls_rhs_last_analysis = ncalls(semi.performance_counter)

    # Record total time spent in garbage collection so far using a high-resolution clock
    # Note: For details see the actual callback function below
    analysis_callback.start_gc_time = Base.gc_time_ns()

    analysis_callback(u_ode, du_ode, integrator, semi)
    return nothing
end

# This method gets called from OrdinaryDiffEq's `solve(...)`
function (analysis_callback::AnalysisCallback)(integrator)
    semi = integrator.p
    du_ode = first(get_tmp_cache(integrator))
    u_ode = integrator.u
    return analysis_callback(u_ode, du_ode, integrator, semi)
end

# This method gets called internally as the main entry point to the AnalysiCallback
# TODO: Taal refactor, allow passing an IO object (which could be devnull to avoid cluttering the console)
function (analysis_callback::AnalysisCallback)(u_ode, du_ode, integrator, semi)
    mesh, equations, solver, cache = mesh_equations_solver_cache(semi)
    @unpack dt, t = integrator
    iter = integrator.stats.naccept

    # Compute the percentage of the simulation that is done
    t = integrator.t
    t_initial = first(integrator.sol.prob.tspan)
    t_final = last(integrator.sol.prob.tspan)
    sim_time_percentage = (t - t_initial) / (t_final - t_initial) * 100

    # Record performance measurements and compute performance index (PID)
    runtime_since_last_analysis = 1.0e-9 * (time_ns() -
                                   analysis_callback.start_time_last_analysis)
    # PID is an MPI-aware measure of how much time per global degree of freedom (i.e., over all ranks
    # and threads) and per `rhs!` evaluation is required. MPI-aware means that it essentially adds up
    # the time spent on each computing unit. Thus, in an ideally parallelized program, the PID should be constant
    # independent of the number of MPI ranks or threads used, since, e.g., using 4x the number of ranks should
    # divide the runtime on each rank by 4. See also the Trixi.jl docs ("Performance" section) for
    # more information.
    ncalls_rhs_since_last_analysis = (ncalls(semi.performance_counter)
                                      -
                                      analysis_callback.ncalls_rhs_last_analysis)
    # This assumes that the same number of threads is used on each MPI rank.
    performance_index = runtime_since_last_analysis * mpi_nranks() *
                        Threads.nthreads() /
                        (ndofsglobal(mesh, solver, cache)
                         *
                         ncalls_rhs_since_last_analysis)

    # Compute the total runtime since the analysis callback has been initialized, in seconds
    runtime_absolute = 1.0e-9 * (time_ns() - analysis_callback.start_time)

    # Compute the relative runtime per thread as time spent in `rhs!` divided by the number of calls
    # to `rhs!` and the number of local degrees of freedom
    # OBS! This computation must happen *after* the PID computation above, since `take!(...)`
    #      will reset the number of calls to `rhs!`
    runtime_relative = 1.0e-9 * take!(semi.performance_counter) * Threads.nthreads() /
                       ndofs(semi)

    # Compute the total time spent in garbage collection since the analysis callback has been
    # initialized, in seconds
    # Note: `Base.gc_time_ns()` is not part of the public Julia API but has been available at least
    #        since Julia 1.6. Should this function be removed without replacement in a future Julia
    #        release, just delete this analysis quantity from the callback.
    # Source: https://github.com/JuliaLang/julia/blob/b540315cb4bd91e6f3a3e4ab8129a58556947628/base/timing.jl#L83-L84
    gc_time_absolute = 1.0e-9 * (Base.gc_time_ns() - analysis_callback.start_gc_time)

    # Compute the percentage of total time that was spent in garbage collection
    gc_time_percentage = gc_time_absolute / runtime_absolute * 100

    # Obtain the current memory usage of the Julia garbage collector, in MiB, i.e., the total size of
    # objects in memory that have been allocated by the JIT compiler or the user code.
    # Note: `Base.gc_live_bytes()` is not part of the public Julia API but has been available at least
    #        since Julia 1.6. Should this function be removed without replacement in a future Julia
    #        release, just delete this analysis quantity from the callback.
    # Source: https://github.com/JuliaLang/julia/blob/b540315cb4bd91e6f3a3e4ab8129a58556947628/base/timing.jl#L86-L97
    memory_use = Base.gc_live_bytes() / 2^20 # bytes -> MiB

    @trixi_timeit timer() "analyze solution" begin
        # General information
        mpi_println()
        mpi_println("─"^100)
        mpi_println(" Simulation running '", get_name(equations), "' with ",
                    summary(solver))
        mpi_println("─"^100)
        mpi_println(" #timesteps:     " * @sprintf("% 14d", iter) *
                    "               " *
                    " run time:       " * @sprintf("%10.8e s", runtime_absolute))
        mpi_println(" Δt:             " * @sprintf("%10.8e", dt) *
                    "               " *
                    " └── GC time:    " *
                    @sprintf("%10.8e s (%5.3f%%)", gc_time_absolute, gc_time_percentage))
        mpi_println(rpad(" sim. time:      " *
                         @sprintf("%10.8e (%5.3f%%)", t, sim_time_percentage), 46) *
                    " time/DOF/rhs!:  " * @sprintf("%10.8e s", runtime_relative))
        mpi_println("                 " * "              " *
                    "               " *
                    " PID:            " * @sprintf("%10.8e s", performance_index))
        mpi_println(" #DOFs per field:" * @sprintf("% 14d", ndofsglobal(semi)) *
                    "               " *
                    " alloc'd memory: " * @sprintf("%14.3f MiB", memory_use))
        mpi_println(" #elements:      " *
                    @sprintf("% 14d", nelementsglobal(mesh, solver, cache)))

        # Level information (only for AMR and/or non-uniform `TreeMesh`es)
        print_level_information(integrator.opts.callback, mesh, solver, cache)
        mpi_println()

        # Open file for appending and store time step and time information
        if mpi_isroot() && analysis_callback.save_analysis
            io = open(joinpath(analysis_callback.output_directory,
                               analysis_callback.analysis_filename), "a")
            print(io, iter)
            print(io, " ", t)
            print(io, " ", dt)
        else
            io = devnull
        end

        # Calculate current time derivative (needed for semidiscrete entropy time derivative, residual, etc.)
        # `integrator.f` is usually just a call to `rhs!`
        # However, we want to allow users to modify the ODE RHS outside of Trixi.jl
        # and allow us to pass a combined ODE RHS to OrdinaryDiffEq, e.g., for
        # hyperbolic-parabolic systems.
        @notimeit timer() integrator.f(du_ode, u_ode, semi, t)
        u = wrap_array(u_ode, mesh, equations, solver, cache)
        du = wrap_array(du_ode, mesh, equations, solver, cache)
        # Compute l2_error, linf_error
        analysis_callback(io, du, u, u_ode, t, semi)

        mpi_println("─"^100)
        mpi_println()

        # Add line break and close analysis file if it was opened
        if mpi_isroot() && analysis_callback.save_analysis
            # This resolves a possible type instability introduced above, since `io`
            # can either be an `IOStream` or `devnull`, but we know that it must be
            # an `IOStream here`.
            println(io::IOStream)
            close(io::IOStream)
        end
    end

    # avoid re-evaluating possible FSAL stages
    u_modified!(integrator, false)

    # Reset performance measurements
    analysis_callback.start_time_last_analysis = time_ns()
    analysis_callback.ncalls_rhs_last_analysis = ncalls(semi.performance_counter)

    return nothing
end

# This method is just called internally from `(analysis_callback::AnalysisCallback)(integrator)`
# and serves as a function barrier. Additionally, it makes the code easier to profile and optimize.
function (analysis_callback::AnalysisCallback)(io, du, u, u_ode, t, semi)
    @unpack analyzer, analysis_errors, analysis_integrals = analysis_callback
    cache_analysis = analysis_callback.cache
    _, equations, _, _ = mesh_equations_solver_cache(semi)

    # Calculate and print derived quantities (error norms, entropy etc.)
    # Variable names required for L2 error, Linf error, and conservation error
    if any(q in analysis_errors
           for q in (:l2_error, :linf_error, :conservation_error, :residual)) &&
       mpi_isroot()
        print(" Variable:    ")
        for v in eachvariable(equations)
            @printf("   %-14s", varnames(cons2cons, equations)[v])
        end
        println()
    end

    if :l2_error in analysis_errors || :linf_error in analysis_errors
        # Calculate L2/Linf errors
        l2_error, linf_error = calc_error_norms(u_ode, t, analyzer, semi,
                                                cache_analysis)

        if mpi_isroot()
            # L2 error
            if :l2_error in analysis_errors
                print(" L2 error:    ")
                for v in eachvariable(equations)
                    @printf("  % 10.8e", l2_error[v])
                    print(io, " ", l2_error[v])
                end
                println()
            end

            # Linf error
            if :linf_error in analysis_errors
                print(" Linf error:  ")
                for v in eachvariable(equations)
                    @printf("  % 10.8e", linf_error[v])
                    print(io, " ", linf_error[v])
                end
                println()
            end
        end
    end

    # Conservation error
    if :conservation_error in analysis_errors
        @unpack initial_state_integrals = analysis_callback
        state_integrals = integrate(u_ode, semi)

        if mpi_isroot()
            print(" |∑U - ∑U₀|:  ")
            for v in eachvariable(equations)
                err = abs(state_integrals[v] - initial_state_integrals[v])
                @printf("  % 10.8e", err)
                print(io, " ", err)
            end
            println()
        end
    end

    # Residual (defined here as the vector maximum of the absolute values of the time derivatives)
    if :residual in analysis_errors
        mpi_print(" max(|Uₜ|):   ")
        for v in eachvariable(equations)
            # Calculate maximum absolute value of Uₜ
            res = maximum(abs, view(du, v, ..))
            if mpi_isparallel()
                # TODO: Debugging, here is a type instability
                # Base.max instead of max needed, see comment in src/auxiliary/math.jl
                global_res = MPI.Reduce!(Ref(res), Base.max, mpi_root(), mpi_comm())
                if mpi_isroot()
                    res::eltype(du) = global_res[]
                end
            end
            if mpi_isroot()
                @printf("  % 10.8e", res)
                print(io, " ", res)
            end
        end
        mpi_println()
    end

    # L2/L∞ errors of the primitive variables
    if :l2_error_primitive in analysis_errors ||
       :linf_error_primitive in analysis_errors
        l2_error_prim, linf_error_prim = calc_error_norms(cons2prim, u_ode, t, analyzer,
                                                          semi, cache_analysis)

        if mpi_isroot()
            print(" Variable:    ")
            for v in eachvariable(equations)
                @printf("   %-14s", varnames(cons2prim, equations)[v])
            end
            println()

            # L2 error
            if :l2_error_primitive in analysis_errors
                print(" L2 error prim.: ")
                for v in eachvariable(equations)
                    @printf("%10.8e   ", l2_error_prim[v])
                    print(io, " ", l2_error_prim[v])
                end
                println()
            end

            # L∞ error
            if :linf_error_primitive in analysis_errors
                print(" Linf error pri.:")
                for v in eachvariable(equations)
                    @printf("%10.8e   ", linf_error_prim[v])
                    print(io, " ", linf_error_prim[v])
                end
                println()
            end
        end
    end

    # additional integrals
    analyze_integrals(analysis_integrals, io, du, u, t, semi)

    return nothing
end

function print_level_information(mesh, solver, cache, min_level, max_level)
    # Get local element count per level
    elements_per_level = get_elements_per_level(min_level, max_level, mesh, solver,
                                                cache)

    # Sum up across all ranks
    MPI.Reduce!(elements_per_level, +, mpi_root(), mpi_comm())

    # Print
    for level in max_level:-1:(min_level + 1)
        mpi_println(" ├── level $level:    " *
                    @sprintf("% 14d", elements_per_level[level + 1 - min_level]))
    end
    mpi_println(" └── level $min_level:    " *
                @sprintf("% 14d", elements_per_level[1]))

    return nothing
end

function print_level_information(callbacks, mesh::TreeMesh, solver, cache)
    if uses_amr(callbacks)
        # Get global minimum and maximum level from the AMRController
        min_level = max_level = 0
        for cb in callbacks.discrete_callbacks
            if cb.affect! isa AMRCallback
                min_level = cb.affect!.controller.base_level
                max_level = cb.affect!.controller.max_level
            end
        end
        print_level_information(mesh, solver, cache, min_level, max_level)
        # Special check for `TreeMesh`es without AMR.
        # These meshes may still be non-uniform due to `refinement_patches`, see
        # `refine_box!`, `coarsen_box!`, and `refine_sphere!`.
    elseif minimum_level(mesh.tree) != maximum_level(mesh.tree)
        min_level = minimum_level(mesh.tree)
        max_level = maximum_level(mesh.tree)
        print_level_information(mesh, solver, cache, min_level, max_level)
    else # Uniform mesh
        return nothing
    end
end

function print_level_information(callbacks, mesh, solver, cache)
    if uses_amr(callbacks)
        # Get global minimum and maximum level from the AMRController
        min_level = max_level = 0
        for cb in callbacks.discrete_callbacks
            if cb.affect! isa AMRCallback
                min_level = cb.affect!.controller.base_level
                max_level = cb.affect!.controller.max_level
            end
        end
        print_level_information(mesh, solver, cache, min_level, max_level)
    else # Uniform mesh
        return nothing
    end
end

function get_elements_per_level(min_level, max_level, mesh::P4estMesh, solver, cache)
    elements_per_level = zeros(P4EST_MAXLEVEL + 1)

    for tree in unsafe_wrap_sc(p4est_tree_t, mesh.p4est.trees)
        elements_per_level .+= tree.quadrants_per_level
    end

    return @view(elements_per_level[(min_level + 1):(max_level + 1)])
end

function get_elements_per_level(min_level, max_level, mesh::T8codeMesh, solver, cache)
    levels = trixi_t8_get_local_element_levels(mesh.forest)

    return [count(==(l), levels) for l in min_level:max_level]
end

function get_elements_per_level(min_level, max_level, mesh::TreeMesh, solver, cache)
    levels = [mesh.tree.levels[cache.elements.cell_ids[element]]
              for element in eachelement(solver, cache)]
    return [count(==(l), levels) for l in min_level:max_level]
end

# Iterate over tuples of analysis integrals in a type-stable way using "lispy tuple programming".
function analyze_integrals(analysis_integrals::NTuple{N, Any}, io, du, u, t,
                           semi) where {N}

    # Extract the first analysis integral and process it; keep the remaining to be processed later
    quantity = first(analysis_integrals)
    remaining_quantities = Base.tail(analysis_integrals)

    res = analyze(quantity, du, u, t, semi)
    if mpi_isroot()
        @printf(" %-12s:", pretty_form_utf(quantity))
        @printf("  % 10.8e", res)
        print(io, " ", res)
    end
    mpi_println()

    # Recursively call this method with the unprocessed integrals
    analyze_integrals(remaining_quantities, io, du, u, t, semi)
    return nothing
end

# terminate the type-stable iteration over tuples
function analyze_integrals(analysis_integrals::Tuple{}, io, du, u, t, semi)
    return nothing
end

# used for error checks and EOC analysis
function (cb::DiscreteCallback{Condition, Affect!})(sol) where {Condition,
                                                                Affect! <:
                                                                AnalysisCallback}
    analysis_callback = cb.affect!
    semi = sol.prob.p
    @unpack analyzer = analysis_callback
    cache_analysis = analysis_callback.cache

    l2_error, linf_error = calc_error_norms(sol.u[end], sol.t[end], analyzer, semi,
                                            cache_analysis)
    return (; l2 = l2_error, linf = linf_error)
end

# some common analysis_integrals
# to support another analysis integral, you can overload
# Trixi.analyze, Trixi.pretty_form_utf, Trixi.pretty_form_ascii
function analyze(quantity, du, u, t, semi::AbstractSemidiscretization)
    mesh, equations, solver, cache = mesh_equations_solver_cache(semi)
    return analyze(quantity, du, u, t, mesh, equations, solver, cache)
end
function analyze(quantity, du, u, t, mesh, equations, solver, cache)
    return integrate(quantity, u, mesh, equations, solver, cache, normalize = true)
end
pretty_form_utf(quantity) = get_name(quantity)
pretty_form_ascii(quantity) = get_name(quantity)

# Special analyze for `SemidiscretizationHyperbolicParabolic` such that
# precomputed gradients are available.
function analyze(quantity::typeof(enstrophy), du, u, t,
                 semi::SemidiscretizationHyperbolicParabolic)
    mesh, equations, solver, cache = mesh_equations_solver_cache(semi)
    equations_parabolic = semi.equations_parabolic
    cache_parabolic = semi.cache_parabolic
    # We do not apply `enstrophy` directly here because we might later have different `quantity`s
    # that we wish to integrate, which can share this routine.
    return analyze(quantity, du, u, t, mesh, equations, equations_parabolic, solver,
                   cache, cache_parabolic)
end
function analyze(quantity, du, u, t, mesh, equations, equations_parabolic, solver,
                 cache, cache_parabolic)
    return integrate(quantity, u, mesh, equations, equations_parabolic, solver, cache,
                     cache_parabolic, normalize = true)
end

function entropy_timederivative end
pretty_form_utf(::typeof(entropy_timederivative)) = "∑∂S/∂U ⋅ Uₜ"
pretty_form_ascii(::typeof(entropy_timederivative)) = "dsdu_ut"

pretty_form_utf(::typeof(entropy)) = "∑S"

pretty_form_utf(::typeof(energy_total)) = "∑e_total"
pretty_form_ascii(::typeof(energy_total)) = "e_total"

pretty_form_utf(::typeof(energy_kinetic)) = "∑e_kinetic"
pretty_form_ascii(::typeof(energy_kinetic)) = "e_kinetic"

pretty_form_utf(::typeof(energy_kinetic_nondimensional)) = "∑e_kinetic*"
pretty_form_ascii(::typeof(energy_kinetic_nondimensional)) = "e_kinetic*"

pretty_form_utf(::typeof(energy_internal)) = "∑e_internal"
pretty_form_ascii(::typeof(energy_internal)) = "e_internal"

pretty_form_utf(::typeof(energy_magnetic)) = "∑e_magnetic"
pretty_form_ascii(::typeof(energy_magnetic)) = "e_magnetic"

pretty_form_utf(::typeof(cross_helicity)) = "∑v⋅B"
pretty_form_ascii(::typeof(cross_helicity)) = "v_dot_B"

pretty_form_utf(::typeof(enstrophy)) = "∑enstrophy"
pretty_form_ascii(::typeof(enstrophy)) = "enstrophy"

pretty_form_utf(::Val{:l2_divb}) = "L2 ∇⋅B"
pretty_form_ascii(::Val{:l2_divb}) = "l2_divb"

pretty_form_utf(::Val{:linf_divb}) = "L∞ ∇⋅B"
pretty_form_ascii(::Val{:linf_divb}) = "linf_divb"

pretty_form_utf(::typeof(lake_at_rest_error)) = "∑|H₀-(h+b)|"
pretty_form_ascii(::typeof(lake_at_rest_error)) = "|H0-(h+b)|"
end # @muladd

# specialized implementations specific to some solvers
include("analysis_dg1d.jl")
include("analysis_dg2d.jl")
include("analysis_surface_integral.jl")
include("analysis_dg2d_parallel.jl")
include("analysis_dg3d.jl")
include("analysis_dg3d_parallel.jl")

# This version of `analyze` is used for [`AnalysisSurfaceIntegral`](@ref) which requires
# `semi` to be passed along to retrieve the current boundary indices, which are non-static
# in the case of AMR.
function analyze(quantity::AnalysisSurfaceIntegral, du, u, t,
                 semi::AbstractSemidiscretization)
    mesh, equations, solver, cache = mesh_equations_solver_cache(semi)
    return analyze(quantity, du, u, t, mesh, equations, solver, cache, semi)
end

# Special analyze for `SemidiscretizationHyperbolicParabolic` such that
# precomputed gradients are available. Required for `enstrophy` (see above) and viscous forces.
# Note that this needs to be included after `analysis_surface_integral_2d.jl` to
# have `VariableParabolic` available.
function analyze(quantity::AnalysisSurfaceIntegral{Variable},
                 du, u, t,
                 semi::SemidiscretizationHyperbolicParabolic) where {Variable <:
                                                                     VariableParabolic}
    mesh, equations, solver, cache = mesh_equations_solver_cache(semi)
    equations_parabolic = semi.equations_parabolic
    cache_parabolic = semi.cache_parabolic
    return analyze(quantity, du, u, t, mesh, equations, equations_parabolic, solver, cache,
                   semi, cache_parabolic)
end