Parallel Distances

Parallel versions of functions from Distances.jl .

parallel_pairwise(
    distance, X[, Y];
    dims::Integer,
    policy::Symbol = :greedy_sticky,
    progress::Maybe{AbstractProgress} = nothing
    progress_chunk::Maybe{Integer} = nothing,
)::AbstractMatrix

A parallel version of pairwise . This will use parallel_loop_wo_rng over the columns of Y , with the specified policy and progress . If policy is :serial , then the standard version of pairwise is called and progress is ignored.

using Distances
using Random

Random.seed!(123456)

m1 = rand(10, 20)
m2 = rand(10, 30)

d = pairwise(Euclidean(), m1; dims = 2)
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1; dims = 2, policy = :serial))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1; dims = 2, policy = :greedy))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1; dims = 2, policy = :dynamic))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1; dims = 2, policy = :static))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1; dims = 2, policy = :greedy_sticky))) < 1e-6

d = pairwise(Euclidean(), m1, m2; dims = 2)
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1, m2; dims = 2, policy = :serial))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1, m2; dims = 2, policy = :greedy))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1, m2; dims = 2, policy = :dynamic))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1, m2; dims = 2, policy = :static))) < 1e-6
@assert maximum(abs.(d .- parallel_pairwise(Euclidean(), m1, m2; dims = 2, policy = :greedy_sticky))) < 1e-6

# output

parallel_pairwise_closest(
    distance, X, Y;
    dims::Integer,
    closest_index::Maybe{AbstractVector{<:Integer}} = nothing,
    closest_distance::Maybe{AbstractVector} = nothing,
    policy::Symbol = :greedy,
    progress::Maybe{AbstractProgress} = nothing,
    progress_chunk::Maybe{Integer} = nothing,
)::Nothing

For each entry of Y , find the closest entry of X . This is equivalent to computing parallel_pairwise and reducing each column to its minimal value and its location, except that the full distances matrix is never stored. The distances are computed and reduced on the fly using parallel_loop_wo_rng over the entries of Y , with the specified policy and progress .

The results are written into the pre-allocated closest_index and/or closest_distance vectors (whose length must be the number of entries of Y ); at least one of them must be specified. This performs no allocations of its own.

using Distances
using Random

Random.seed!(123456)

m1 = rand(10, 20)
m2 = rand(10, 30)

d = pairwise(Euclidean(), m1, m2; dims = 2)

closest_index = Vector{Int}(undef, 30)
closest_distance = Vector{Float64}(undef, 30)
parallel_pairwise_closest(Euclidean(), m1, m2; dims = 2, closest_index, closest_distance)

@assert closest_index == [argmin(d[:, column]) for column in 1:30]
@assert maximum(abs.(closest_distance .- [minimum(d[:, column]) for column in 1:30])) < 1e-6

# output

parallel_colwise(
    distance, X, Y;
    policy::Symbol = :greedy_sticky,
    progress::Maybe{AbstractProgress} = nothing,
)::AbstractVector

A parallel version of colwise . This will use parallel_loop_wo_rng over the columns of X and Y , using the specified policy and progress . If policy is :serial , then the standard version of pairwise is called and progress is ignored.

using Distances
using Random

Random.seed!(123456)

m1 = rand(10, 20)
m2 = rand(10, 20)

d = colwise(Euclidean(), m1, m2)
@assert maximum(abs.(d .- parallel_colwise(Euclidean(), m1, m2; policy = :serial))) < 1e-6
@assert maximum(abs.(d .- parallel_colwise(Euclidean(), m1, m2; policy = :greedy))) < 1e-6
@assert maximum(abs.(d .- parallel_colwise(Euclidean(), m1, m2; policy = :dynamic))) < 1e-6
@assert maximum(abs.(d .- parallel_colwise(Euclidean(), m1, m2; policy = :static))) < 1e-6
@assert maximum(abs.(d .- parallel_colwise(Euclidean(), m1, m2; policy = :greedy_sticky))) < 1e-6

# output