Skip to content

Problem/error in execution order #531

@chengxo

Description

@chengxo

I observe that the following MWE code works differently with vs without the tturbo.

using LoopVectorization, Random
function weird_update(a, c, d)
    @tturbo for j in axes(a, 2)
        tmp = 0.0
        for i in axes(a, 1)
            a_ij = a[i, j]
            b_ij = a_ij + 1.0
            tmp += b_ij^2
        end
        c[j] = tmp
        d[j] = exp(2.0*tmp)
    end
    return nothing
end
m = 100
n = 5
rng = MersenneTwister(0)
a = zeros(n, m)
randn!(rng, a)
c = zeros(m)
d = zeros(m)
weird_update(a, c, d)
println(c)
println(d)

In fact, with tturbo, d is always 1, as if the code for updating d is written after tmp=0.0. On the other hand, the update for c works well.

Could you help me figure out a way to do this correctly with tturbo?

UPDATE1: it is weird that if I change d[j] = exp(2.0*tmp) to d[j]=exp(2.0*(0.0+tmp)), then it works correctly.

UPDATE2: The following code works even more weirdly

function weird_update(a, c, d)
    @tturbo for j in axes(a, 2)
        tmp = 0.0
        for i in axes(a, 1)
            a_ij = a[i, j]
            b_ij = a_ij + 1.0
            tmp += b_ij^2
        end
        c[j] = tmp
        d[j] = exp(2.0*c[j])
        
    end
    return nothing
end
m = 100
n = 5
rng = MersenneTwister(0)
a = zeros(n, m)
randn!(rng, a)
c = ones(m)
d = zeros(m)
weird_update(a, c, d)
println(c)
println(d)

The output of d is all one's -- not even exp(2). I am totally confused.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions