Type-Stable Function to Return Plots.Plot

General Usage Performance
1

The question below was already raised at This post.

This code

using Plots

function foo()

    #..Initialize the plot.. 
    p = plot()
        
    return p 
end

function gnu()

    #..Initialize the plot.. 
    p = plot()::Plots.Plot{Plots.GRBackend}
        
    # return p 
end

yields the output

@code_warntype foo()

MethodInstance for foo()
  from foo() @ Main In[14]:3
Arguments
  #self#::Core.Const(foo)
Locals
  p::Plots.Plot
Body::Plots.Plot
1 ─     (p = Main.plot())
└──     return p

and

@code_warntype gnu()

MethodInstance for gnu()
  from gnu() @ Main In[14]:11
Arguments
  #self#::Core.Const(gnu)
Locals
  p::Plots.Plot{Plots.GRBackend}
Body::Plots.Plot{Plots.GRBackend}
1 ─ %1 = Main.plot()::Plots.Plot
│   %2 = Plots.Plot::Core.Const(Plots.Plot)
│   %3 = Plots.GRBackend::Core.Const(Plots.GRBackend)
│   %4 = Core.apply_type(%2, %3)::Core.Const(Plots.Plot{Plots.GRBackend})
│   %5 = Core.typeassert(%1, %4)::Plots.Plot{Plots.GRBackend}
│        (p = %5)
└──      return %5

What is a type-stable definition of a function that returns a plot? Thx.

2

Plots.plot() is not type stable. Its return type depends on global variables:

julia> using Plots

julia> typeof(plot())
Plots.Plot{Plots.GRBackend}

julia> unicodeplots()
Plots.UnicodePlotsBackend()

julia> typeof(plot())
Plots.Plot{Plots.UnicodePlotsBackend}

For most purposes I’d guess the performance problems from this instability is negligible compared to the actual work done by the plot routines.

2 Likes
3

Many thx!

In the target application, Plots.plot is called inside a for-loop over elements in a array. See function plotMesh below. The array is typically large. The function appears to be slow.

Possibly the code design is faulty to start with. Not sure whether an in-place version, a plotting recipe or an alternative is more appropriate here.

Please advice.

function plotMesh(mesh::Mesh)

    #..Initialize the plot.. 
    p = Plots.plot(xlabel = "x", ylabel = "y",legend = false,aspect_ratio = :equal)
    
    #..loop over number of elements..
    for element in mesh.Elements 
        triangle = Shape([Tuple(element.p1),Tuple(element.p2),Tuple(element.p3)])
        p = Plots.plot!(p,triangle,fillcolor=:blue, alpha=0.7, linecolor=:black)        
    end

    Plots.plot(p)
    
    return p 
end
4

You can put the loop in a separate function to utilize a function barrier:

function plotloop!(p, Elements)
    for element in Elements
        triangle = ...
        Plots.plot!(p, triangle, ...) 
    end
end

function plotMesh(mesh::Mesh)
    p = Plots.plot(xlabel, ...)
    plotloop!(p, mesh.Elements)
    ...
end

In this way there is a single instability inside plotMesh, but inside plotloop! there is no instability since the type of p is fully known when plotloop! is called. (Note that it’s not necessary to reassign p in p = Plots.plot!(p, ...) since plot! returns the original p, but updated.)

Edit: on second thought, plot! is not specializing on the plot type, so this may not help a lot. It could be that the slowness are from something else, like garbage collection due to the many vectors created when you construct the Shape.

5

It might be worthwhile to create a separate plot receipe for your Mesh type, but it depends on how much you use such plots: Overview · Plots

6

Right!

The code below continues to be slow due to many allocations.

function plotloop!(p,mesh::Mesh)
    #..loop over number of elements..
    for element in mesh.Elements 
        triangle = Shape([Tuple(element.p1),Tuple(element.p2),Tuple(element.p3)])
        Plots.plot!(p,triangle,fillcolor=:blue, alpha=0.7, linecolor=:black)        
    end    
    return p 
end

function plotMesh(mesh::Mesh)
 
    #..Initialize the plot.. 
    p = Plots.plot(xlabel = "x", ylabel = "y",legend = false,aspect_ratio = :equal)

    plotloop!(p, mesh)

    Plots.plot(p)
    
    return p     
end

mesh = meshFromGmsh("data/square-10.msh"); @btime plotMesh(mesh)

results in 88.881 allocations.

7

I agree with @sgaure that type stability is not the bottleneck here. I’d try to make a vector of shapes and then call plot() just once. MWE:

julia> T = [Shape(rand(3), rand(3)) for i in 1:1000]
julia> plot(T, alpha=0.5, legend=false, aspect_ratio=:equal)

I don’t know how many triangles you’re gonna plot but this is pretty fast with the default GR backend:

Screenshot 2025-08-05 at 21.02.51
Screenshot 2025-08-05 at 21.02.512710×2154 2.15 MB

It takes less than a second to render and display the plot.

8

Correct! Many thanks! Great!

Assume next that each triangle has a value associated to it. How do I take this value into account in the plot?

Kind wishes, Domenico.

Screenshot 2025-08-06 at 15.15.40
Screenshot 2025-08-06 at 15.15.401338×800 191 KB

9

You can pass plot attributes as 1×n matrices. Let’s plot three triangles with different colors:

plot(T[1:3], alpha=0.5, color=[:blue :green :red], legend=false)

Just extract the colors from mesh as you did with the coordinates. Note that plot requires a 1×n matrix, not a vector (see here).

10

Thx! Requires more thought.