KeyError: key Symbol not found

[SkyWorld117]

I did an update of LV from 0.12.50 to 0.12.59. The following code ran well on 0.12.50 but triggers KeyError on 0.12.59 and 0.12.60.

@avxt for i in axes(nn.voltage, 1)
    s = 0.0f0
    for j in axes(nn.weights, 2)
        s += nn.weights[i,j] * nn.output[j,t]
        nn.weights[i,j] += ifelse(i!=j, min(A₋*nn.M[i]*nn.weights[i,j], 1.0f0), 0.0f0) * ifelse(nn.output[j,t]!=0, 1.0f0, 0.0f0)
        nn.weights[i,j] = ifelse(nn.weights[i,j]>θ*Cₘ+gₗ*(θ-Eₗ), θ*Cₘ+gₗ*(θ-Eₗ), nn.weights[i,j])
        nn.weights[i,j] = ifelse(nn.weights[i,j]<-θ*Cₘ-gₗ*(θ-Eₗ), -θ*Cₘ-gₗ*(θ-Eₗ), nn.weights[i,j])
    end
    nn.voltage[i,1] += 1/Cₘ * (s - gₗ*(nn.voltage[i,1] - Eₗ))
    nn.voltage[i,1] = ifelse(nn.voltage[i,1]>θ, θ, nn.voltage[i,1])
    nn.voltage[i,1] = ifelse(nn.voltage[i,1]<Eₗ, Eₗ, nn.voltage[i,1])
end

However, if I deactivate some lines in the block, it can pass the compiling.

@avxt for i in axes(nn.voltage, 1)
    s = 0.0f0
    for j in axes(nn.weights, 2)
        s += nn.weights[i,j] * nn.output[j,t]
        nn.weights[i,j] += ifelse(i!=j, min(A₋*nn.M[i]*nn.weights[i,j], 1.0f0), 0.0f0) * ifelse(nn.output[j,t]!=0, 1.0f0, 0.0f0)
        #nn.weights[i,j] = ifelse(nn.weights[i,j]>θ*Cₘ+gₗ*(θ-Eₗ), θ*Cₘ+gₗ*(θ-Eₗ), nn.weights[i,j])
        #nn.weights[i,j] = ifelse(nn.weights[i,j]<-θ*Cₘ-gₗ*(θ-Eₗ), -θ*Cₘ-gₗ*(θ-Eₗ), nn.weights[i,j])
    end
    nn.voltage[i,1] += 1/Cₘ * (s - gₗ*(nn.voltage[i,1] - Eₗ))
    #nn.voltage[i,1] = ifelse(nn.voltage[i,1]>θ, θ, nn.voltage[i,1])
    #nn.voltage[i,1] = ifelse(nn.voltage[i,1]<Eₗ, Eₗ, nn.voltage[i,1])
end

I could not develop a simpler case because they work somehow fine with the similar principal.

[chriselrod]

I have a fix for this locally. Mind giving me something I can copy/paste and run (that errors on 0.12.60) to confirm it works correctly and add it to the tests?

[SkyWorld117]

Sounds great! You can find all the code here.

It is in the network.jl file, you can test the whole program via running test.jl.

[chriselrod]

I don't want all the code. Just something I can put in a test file that would fail a test on 0.12.60 with the error reported in this issue (but pass on 0.12.50).

[SkyWorld117]

I don't want all the code. Just something I can put in a test file that would fail a test on 0.12.60 with the error reported in this issue (but pass on 0.12.50).

OK, I'll try to get a short one.

[SkyWorld117]

Actually, if you don't mind a large file, you can run network.jl directly for a test. It will only be compiled to show if something goes wrong.

[SkyWorld117]

Here's the shorter version:

using LoopVectorization

mutable struct Network
    input_size::Int64
    output_size::Int64
    hidden_neurons::Int64

    max_runtime::Int64

    weights::Array{Float32}

    voltage::Array{Float32}
    M::Array{Float32}
    output::Array{Float32}

    activate::Any

    function Network(;input_size::Int64, output_size::Int64, hidden_neurons::Int64, max_runtime::Int64)
        s = input_size + output_size + hidden_neurons
        weights = rand(-sqrt(3/s):1.0e-5:sqrt(3/s), s,s)
        @avxt for i in 1:s
            weights[i,i] = 0.0f0
        end

        voltage = zeros(Float32, s,1)
        M = zeros(Float32, s,1)
        output = zeros(Float32, s,max_runtime)
        new(input_size, output_size, hidden_neurons, max_runtime, weights, voltage, M, output, activate)
    end

    function activate(nn::Network; Cₘ, gₗ, Eₗ, U, τᵣ, τᵤ, f, θ, A₊, A₋)
        t = 1
        @avxt for i in axes(nn.voltage, 1)
            s = 0.0f0
            for j in axes(nn.weights, 2)
                nn.weights[i,j] = ifelse(nn.weights[i,j]>θ*Cₘ+gₗ*(θ-Eₗ), θ*Cₘ+gₗ*(θ-Eₗ), nn.weights[i,j])
                nn.weights[i,j] = ifelse(nn.weights[i,j]<-θ*Cₘ-gₗ*(θ-Eₗ), -θ*Cₘ-gₗ*(θ-Eₗ), nn.weights[i,j])
            end
            nn.voltage[i,1] = ifelse(nn.voltage[i,1]>θ, θ, nn.voltage[i,1])
            nn.voltage[i,1] = ifelse(nn.voltage[i,1]<Eₗ, Eₗ, nn.voltage[i,1])
        end
    end
end

[chriselrod]

I wanted to run it as well to make sure the answer is correct before pushing it and tagging another release of LoopVectorization. I've already confirmed that it no longer throws that error.

[chriselrod]

I pushed my fix. I'll tag a new release after tests pass. Please file another issue if/when you encounter more problems!

[chriselrod]

Oops, definitely broke things. Should've tested locally first.