JuliaSIMD / JuliaSIMD/LoopVectorization.jl

Incorrect indexing when not using zero-based arrays

Offen
#413 0 Kommentare 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen
Vorherrschende Sprache
Julia
Sterne
789
Forks
73
PR-Merge-Kennzahlen
Keine gemergten PRs in 30 T.

Beschreibung

```julia
using LoopVectorization, Static
M,K,N = 3,4,5
A = view(fill(NaN,100,100), 11:10+M,11:10+K+1); A .= rand.();
B = view(fill(NaN,100,100), 11:10+N,11:10+K); B .= rand.();
C0 = view(zeros(100,100), 11:10+N, 11:10+M); C1 = deepcopy(C0);
function reference!(C,A,B)
K = lastindex(C,static(2))
for i_1 = indices((A, C), (1, 2))
for i_0 = indices((B, C), (1, 1))
begin
Caccum = zero(eltype(C))
for r_0 = axes(B, 2)
Caccum += A[i_1, r_0] * B[i_0, r_0]
end
C[i_0, i_1] = Caccum + ifelse(i_0 == firstindex(C, static(1)), A[i_1, K], zero(Caccum))
end
end
end
end
function turbo_fail!(C,A,B)
K = lastindex(C,static(2))
@turbo for i_1 = indices((A, C), (1, 2))
for i_0 = indices((B, C), (1, 1))
begin
Caccum = zero(eltype(C))
for r_0 = axes(B, 2)
Caccum += A[i_1, r_0] * B[i_0, r_0]
end
C[i_0, i_1] = Caccum + ifelse(i_0 == firstindex(C, static(1)), A[i_1, K], zero(Caccum))
end
end
end
end
reference!(C0,A,B); C0
turbo_fail!(C1,A,B); C1
C0 ≈ C1
C0 ≈ view(parent(C1), C1.indices[1] .+ 1, C1.indices[2])
using StrideArraysCore: PtrArray, zeroindex
turbo_fail!(zeroindex(PtrArray(C1)),zeroindex(PtrArray(A)),zeroindex(PtrArray(B))); C1
C0 ≈ C1
```
I get
```julia
julia> reference!(C0,A,B); C0
5×3 view(::Matrix{Float64}, 11:15, 11:13) with eltype Float64:
1.4703 1.87281 1.27726
0.870971 1.35742 0.975715
0.395836 1.33265 0.968053
0.219909 1.592 0.970622
0.719743 2.00412 1.42784

julia> turbo_fail!(C1,A,B); C1
5×3 view(::Matrix{Float64}, 11:15, 11:13) with eltype Float64:
0.0 0.0 0.0
1.4703 1.87281 1.27726
0.870971 1.35742 0.975715
0.395836 1.33265 0.968053
0.219909 1.592 0.970622

julia> C0 ≈ C1
false

julia> C0 ≈ view(parent(C1), C1.indices[1] .+ 1, C1.indices[2])
true

julia> using StrideArraysCore: PtrArray, zeroindex

julia> turbo_fail!(zeroindex(PtrArray(C1)),zeroindex(PtrArray(A)),zeroindex(PtrArray(B))); C1
5×3 view(::Matrix{Float64}, 11:15, 11:13) with eltype Float64:
1.4703 1.87281 1.27726
0.870971 1.35742 0.975715
0.395836 1.33265 0.968053
0.219909 1.592 0.970622
0.719743 2.00412 1.42784

julia> C0 ≈ C1
true
```
So with 1-based indexing, it is storing `C` at an offset of 1 in the dimension where we're using the index.
That is, LV converts `C` to use 0-based indexing, but uses 1-based indexing for `i_0` because `i_0` is used in an expression.

Beitragsleitfaden

Für dieses Repository ist kein Beitragsleitfaden indexiert

Bewertung

Dieses Issue wurde noch nicht bewertet.

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.