Switch u layout

src/delay_systems.jl

@@ -204,7 +204,7 @@ function _lsim(sys::DelayLtiSystem{T,S}, Base.@nospecialize(u!), t::AbstractArra
         y[:,k] .= sv.saveval[k][2]
     end
 
-    return y', t, x'
+    return y, t, x
 end
 
 # We have to default to something, look at the sys.P.P and delays
@@ -244,8 +244,8 @@ function Base.step(sys::DelayLtiSystem{T}, t::AbstractVector; kwargs...) where T
     if nu == 1
         y, tout, x = lsim(sys, u, t; x0=x0, kwargs...)
     else
-        x = Array{T}(undef, length(t), nstates(sys), nu)
-        y = Array{T}(undef, length(t), noutputs(sys), nu)
+        x = Array{T}(undef, nstates(sys), length(t), nu)
+        y = Array{T}(undef, noutputs(sys), length(t), nu)
         for i=1:nu
             y[:,:,i], tout, x[:,:,i] = lsim(sys[:,i], u, t; x0=x0, kwargs...)
         end
@@ -264,8 +264,8 @@ function impulse(sys::DelayLtiSystem{T}, t::AbstractVector; alg=MethodOfSteps(BS
     if nu == 1
         y, tout, x = lsim(sys, u, t; alg=alg, x0=sys.P.B[:,1], kwargs...)
     else
-        x = Array{T}(undef, length(t), nstates(sys), nu)
-        y = Array{T}(undef, length(t), noutputs(sys), nu)
+        x = Array{T}(undef, nstates(sys), length(t), nu)
+        y = Array{T}(undef, noutputs(sys), length(t), nu)
         for i=1:nu
             y[:,:,i], tout, x[:,:,i] = lsim(sys[:,i], u, t; alg=alg, x0=sys.P.B[:,i], kwargs...)
         end

src/plotting.jl

@@ -158,7 +158,7 @@ lsimplot
                 title   --> "System Response"
                 subplot --> s2i(1,i)
                 label     --> "\$G_{$(si)}\$"
-                t,  y[:, i]
+                t,  y[i, :]
             end
         end
     end
@@ -208,7 +208,7 @@ for (func, title, typ) = ((step, "Step Response", Stepplot), (impulse, "Impulse
             y,t = func(s, p.args[2:end]...)
             for i=1:ny
                 for j=1:nu
-                    ydata = reshape(y[:, i, j], size(t, 1))
+                    ydata = reshape(y[i, :, j], size(t, 1))
                     style = iscontinuous(s) ? :path : :steppost
                     ttext = (nu > 1 && i==1) ? title*" from: u($j) " : title
                     titles[s2i(i,j)] = ttext

src/synthesis.jl

@@ -30,7 +30,7 @@ u(x,t) = -L*x # Form control law,
 t=0:0.1:5
 x0 = [1,0]
 y, t, x, uout = lsim(sys,u,t,x0=x0)
-plot(t,x, lab=["Position" "Velocity"], xlabel="Time [s]")
+plot(t,x', lab=["Position" "Velocity"], xlabel="Time [s]")
 ```
 """
 function lqr(A, B, Q, R)
@@ -95,7 +95,7 @@ u(x,t) = -L*x # Form control law,
 t=0:Ts:5
 x0 = [1,0]
 y, t, x, uout = lsim(sys,u,t,x0=x0)
-plot(t,x, lab=["Position"  "Velocity"], xlabel="Time [s]")
+plot(t,x', lab=["Position"  "Velocity"], xlabel="Time [s]")
 ```
 """
 function dlqr(A, B, Q, R)

src/timeresp.jl

@@ -10,18 +10,20 @@ Calculate the step response of system `sys`. If the final time `tfinal` or time
 vector `t` is not provided, one is calculated based on the system pole
 locations.
 
-`y` has size `(length(t), ny, nu)`, `x` has size `(length(t), nx, nu)`"""
+`y` has size `(ny, length(t), nu)`, `x` has size `(nx, length(t), nu)`"""
 function Base.step(sys::AbstractStateSpace, t::AbstractVector; method=:cont)
     lt = length(t)
     ny, nu = size(sys)
     nx = sys.nx
-    u = (x,t)->[one(eltype(t))]
+    u = (x,t)->[one(eltype(t))] # 547.109 μs (5240 allocations: 748.75 KiB)
+    #tmp = [one(eltype(t))] # Slightly less allocations but maybe no real difference, 539.624 μs (4916 allocations: 721.61 KiB)
+    #u = (x,t)->tmp
     x0 = zeros(nx)
     if nu == 1
         y, tout, x, _ = lsim(sys, u, t, x0=x0, method=method)
     else
-        x = Array{Float64}(undef, lt, nx, nu)
-        y = Array{Float64}(undef, lt, ny, nu)
+        x = Array{Float64}(undef, nx, lt, nu)
+        y = Array{Float64}(undef, ny, lt, nu)
         for i=1:nu
             y[:,:,i], tout, x[:,:,i],_ = lsim(sys[:,i], u, t, x0=x0, method=method)
         end
@@ -41,7 +43,7 @@ Calculate the impulse response of system `sys`. If the final time `tfinal` or ti
 vector `t` is not provided, one is calculated based on the system pole
 locations.
 
-`y` has size `(length(t), ny, nu)`, `x` has size `(length(t), nx, nu)`"""
+`y` has size `(ny, length(t), nu)`, `x` has size `(nx, length(t), nu)`"""
 function impulse(sys::AbstractStateSpace, t::AbstractVector; method=:cont)
     T = promote_type(eltype(sys.A), Float64)
     lt = length(t)
@@ -61,8 +63,8 @@ function impulse(sys::AbstractStateSpace, t::AbstractVector; method=:cont)
     if nu == 1 # Why two cases # QUESTION: Not type stable?
         y, t, x,_ = lsim(sys, u, t, x0=x0s[:], method=method)
     else
-        x = Array{T}(undef, lt, nx, nu)
-        y = Array{T}(undef, lt, ny, nu)
+        x = Array{T}(undef, nx, lt, nu)
+        y = Array{T}(undef, ny, lt, nu)
         for i=1:nu
             y[:,:,i], t, x[:,:,i],_ = lsim(sys[:,i], u, t, x0=x0s[:,i], method=method)
         end
@@ -81,7 +83,7 @@ impulse(sys::TransferFunction, t::AbstractVector; kwags...) = impulse(ss(sys), t
 Calculate the time response of system `sys` to input `u`. If `x0` is ommitted,
 a zero vector is used.
 
-`y`, `x`, `uout` has time in the first dimension. Initial state `x0` defaults to zero.
+`y`, `x`, `uout` has time in the second dimension. Initial state `x0` defaults to zero.
 
 Continuous time systems are simulated using an ODE solver if `u` is a function. If `u` is an array, the system is discretized before simulation. For a lower level inteface, see `?Simulator` and `?solve`
 
@@ -106,19 +108,19 @@ u(x,t) = -L*x # Form control law,
 t=0:0.1:5
 x0 = [1,0]
 y, t, x, uout = lsim(sys,u,t,x0=x0)
-plot(t,x, lab=["Position" "Velocity"], xlabel="Time [s]")
+plot(t,x', lab=["Position" "Velocity"], xlabel="Time [s]")
 ```
 """
 function lsim(sys::AbstractStateSpace, u::AbstractVecOrMat, t::AbstractVector;
-        x0::AbstractVector=zeros(Bool, sys.nx), method::Symbol=:unspecified)
+        x0::AbstractVecOrMat=fill!(similar(sys.B, nstates(sys), 1), zero(eltype(sys.B))), method::Symbol=:unspecified)
     ny, nu = size(sys)
     nx = sys.nx
 
     if length(x0) != nx
         error("size(x0) must match the number of states of sys")
     end
-    if !(size(u) in [(length(t), nu) (length(t),)])
-        error("u must be of size (length(t), nu)")
+    if size(u) != (nu, length(t))
+        error("u must be of size (nu, length(t))")
     end
 
     dt = Float64(t[2] - t[1])
@@ -135,7 +137,7 @@ function lsim(sys::AbstractStateSpace, u::AbstractVecOrMat, t::AbstractVector;
             dsys = c2d(sys, dt, :zoh)
         elseif method === :foh
             dsys, x0map = c2d_x0map(sys, dt, :foh)
-            x0 = x0map*[x0; transpose(u)[:,1]]
+            x0 = x0map*[x0; u[:,1]]
         else
             error("Unsupported discretization method")
         end
@@ -147,12 +149,12 @@ function lsim(sys::AbstractStateSpace, u::AbstractVecOrMat, t::AbstractVector;
     end
 
     x = ltitr(dsys.A, dsys.B, u, x0)
-    y = transpose(sys.C*transpose(x) + sys.D*transpose(u))
+    y = sys.C*x + sys.D*u
     return y, t, x
 end
 
 function lsim(sys::AbstractStateSpace{<:Discrete}, u::AbstractVecOrMat; kwargs...)
-    t = range(0, length=size(u, 1), step=sys.Ts)
+    t = range(0, length=size(u, 2), step=sys.Ts)
     lsim(sys, u, t; kwargs...)
 end
 
@@ -165,7 +167,7 @@ function lsim(sys::AbstractStateSpace, u::Function, tfinal::Real, args...; kwarg
 end
 
 function lsim(sys::AbstractStateSpace, u::Function, t::AbstractVector;
-        x0::VecOrMat=zeros(sys.nx), method::Symbol=:cont)
+        x0::AbstractVecOrMat=fill!(similar(sys.B, nstates(sys), 1), zero(eltype(sys.B))), method::Symbol=:cont)
     ny, nu = size(sys)
     nx = sys.nx
     u0 = u(x0,1)
@@ -189,15 +191,16 @@ function lsim(sys::AbstractStateSpace, u::Function, t::AbstractVector;
         end
         x,uout = ltitr(dsys.A, dsys.B, u, t, T.(x0))
     else
-        s = Simulator(sys, u)
-        sol = solve(s, T.(x0), (t[1],t[end]), Tsit5())
-        x = sol(t)'
-        uout = Array{eltype(x)}(undef, length(t), ninputs(sys))
-        for (i,ti) in enumerate(t)
-            uout[i,:] = u(x[i,:],ti)'
+        function f(dx,x,p,t) 
+            A, B, u = p
+            dx .= A*x .+ B*u(x,t)
         end
+        p = (sys.A, sys.B, u)
+        sol = solve(ODEProblem(f,x0,(t[1],t[end]),p), Tsit5(); saveat=t)
+        x = reduce(hcat, sol.u)
+        uout = reduce(hcat, u(x[:, i], t[i]) for i in eachindex(t))
     end
-    y = transpose(sys.C*transpose(x) + sys.D*transpose(uout))
+    y = sys.C*x + sys.D*uout
     return y, t, x, uout
 end
 
@@ -218,34 +221,29 @@ e.g, `x0` should prefereably not be a sparse vector.
 If `u` is a function, then `u(x,i)` is called to calculate the control signal every iteration. This can be used to provide a control law such as state feedback `u=-Lx` calculated by `lqr`. In this case, an integrer `iters` must be provided that indicates the number of iterations.
 """
 @views function ltitr(A::AbstractMatrix, B::AbstractMatrix, u::AbstractVecOrMat,
-        x0::AbstractVector=zeros(eltype(A), size(A, 1)))
+        x0::AbstractVecOrMat=zeros(eltype(A), size(A, 1)))
 
     T = promote_type(LinearAlgebra.promote_op(LinearAlgebra.matprod, eltype(A), eltype(x0)),
                       LinearAlgebra.promote_op(LinearAlgebra.matprod, eltype(B), eltype(u)))
 
-    n = size(u, 1)
-
-    # Transposing u allows column-wise operations, which apparently is faster.
-    ut = transpose(u)
+    n = size(u, 2)
 
     # Using similar instead of Matrix{T} to allow for CuArrays to be used.
     # This approach is problematic if x0 is sparse for example, but was considered
     # to be good enough for now
     x = similar(x0, T, (length(x0), n))
 
     x[:,1] .= x0
-    mul!(x[:, 2:end], B, transpose(u[1:end-1, :])) # Do all multiplications B*u[:,k] to save view allocations
+    mul!(x[:, 2:end], B, u[:, 1:end-1]) # Do all multiplications B*u[:,k] to save view allocations
 
-    tmp = similar(x0, T) # temporary vector for storing A*x[:,k]
     for k=1:n-1
-        mul!(tmp, A, x[:,k])
-        x[:,k+1] .+= tmp
+        mul!(x[:, k+1], A, x[:,k], true, true)
     end
-    return transpose(x)
+    return x
 end
 
 function ltitr(A::AbstractMatrix{T}, B::AbstractMatrix{T}, u::Function, t,
-    x0::VecOrMat=zeros(T, size(A, 1))) where T
+    x0::AbstractVecOrMat=zeros(T, size(A, 1))) where T
     iters = length(t)
     x = similar(A, size(A, 1), iters)
     uout = similar(A, size(B, 2), iters)
@@ -255,7 +253,7 @@ function ltitr(A::AbstractMatrix{T}, B::AbstractMatrix{T}, u::Function, t,
         uout[:,i] .= u(x0,t[i])
         x0 = A * x0 + B * uout[:,i]
     end
-    return transpose(x), transpose(uout)
+    return x, uout
 end
 
 # HELPERS:

src/utilities.jl

@@ -27,7 +27,7 @@ function to_abstract_matrix(A::AbstractArray)
     end
     return A
 end
-to_abstract_matrix(A::Vector) = reshape(A, length(A), 1)
+to_abstract_matrix(A::AbstractVector) = reshape(A, length(A), 1)
 to_abstract_matrix(A::Number) = fill(A, 1, 1)
 
 # Do no sorting of eigenvalues

test/test_delayed_systems.jl

@@ -166,21 +166,21 @@ println("Simulating first delay system:")
 @time step(delay(1)*tf(1,[1,1]))
 
 @time y1, t1, x1 = step([s11;s12], 10)
-@time @test y1[:,2] ≈ step(s12, t1)[1] rtol = 1e-14
+@time @test y1[2:2,:] ≈ step(s12, t1)[1] rtol = 1e-14
 
 t = 0.0:0.1:10
 y2, t2, x2 = step(s1, t)
 # TODO Figure out which is inexact here
-@test y2[:,1,1:1] + y2[:,1,2:2] ≈ step(s11, t)[1] + step(s12, t)[1] rtol=1e-14
+@test y2[1:1,:,1:1] + y2[1:1,:,2:2] ≈ step(s11, t)[1] + step(s12, t)[1] rtol=1e-14
 
 y3, t3, x3 = step([s11; s12], t)
-@test y3[:,1,1] ≈ step(s11, t)[1] rtol = 1e-14
-@test y3[:,2,1] ≈ step(s12, t)[1] rtol = 1e-14
+@test y3[1:1,:,1] ≈ step(s11, t)[1] rtol = 1e-14
+@test y3[2:2,:,1] ≈ step(s12, t)[1] rtol = 1e-14
 
 y1, t1, x1 = step(DelayLtiSystem([1.0/s 2/s; 3/s 4/s]), t)
 y2, t2, x2 = step([1.0/s 2/s; 3/s 4/s], t)
 @test y1 ≈ y2 rtol=1e-14
-@test size(x1,1) == length(t)
+@test size(x1,2) == length(t)
 @test size(x1,3) == 2
 
 
@@ -202,7 +202,7 @@ function y_expected(t, K)
       end
 end
 
-@test ystep ≈ y_expected.(t, K) atol = 1e-12
+@test ystep' ≈ y_expected.(t, K) atol = 1e-12
 
 function dy_expected(t, K)
       if t < 1
@@ -219,13 +219,13 @@ end
 y_impulse, t, _ = impulse(sys_known, 3)
 
 # TODO Better accuracy for impulse
-@test y_impulse ≈ dy_expected.(t, K) rtol=1e-13
-@test maximum(abs, y_impulse - dy_expected.(t, K)) < 1e-12
+@test y_impulse' ≈ dy_expected.(t, K) rtol=1e-13
+@test maximum(abs, y_impulse' - dy_expected.(t, K)) < 1e-12
 
 y_impulse, t, _ = impulse(sys_known, 3, alg=MethodOfSteps(Tsit5()))
 # Two orders of magnitude better with BS3 in this case, which is default for impulse
-@test y_impulse ≈ dy_expected.(t, K) rtol=1e-5
-@test maximum(abs, y_impulse - dy_expected.(t, K)) < 1e-5
+@test y_impulse' ≈ dy_expected.(t, K) rtol=1e-5
+@test maximum(abs, y_impulse' - dy_expected.(t, K)) < 1e-5
 
 ## Test delay with D22 term
 t = 0:0.01:4
@@ -234,16 +234,16 @@ sys = delay(1)
 
 y, t, x = step(sys, t)
 @test y[:] ≈ [zeros(100); ones(301)] atol = 1e-12
-@test size(x) == (401,0)
+@test size(x) == (0,401)
 
 sys = delay(1)*delay(1)*1/s
 
 y, t, x = step(sys, t)
 
-y_sol = [zeros(200);0:0.01:2]
+y_sol = [zeros(200);0:0.01:2]'
 
 @test maximum(abs,y-y_sol) < 1e-13
-@test maximum(abs,x-collect(0:0.01:4)) < 1e-15
+@test maximum(abs,x-collect(0:0.01:4)') < 1e-15
 
 # TODO For some reason really bad accuracy here
 # Looks like a lag in time