Implement sorting with AK

Author: christiangnrd

Project.toml

@@ -6,6 +6,7 @@ version = "11.2.6"
 AcceleratedKernels = "6a4ca0a5-0e36-4168-a932-d9be78d558f1"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 GPUArraysCore = "46192b85-c4d5-4398-a991-12ede77f4527"
+GPUToolbox = "096a3bc2-3ced-46d0-87f4-dd12716f4bfc"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
@@ -26,6 +27,7 @@ JLD2Ext = "JLD2"
 AcceleratedKernels = "0.4.3"
 Adapt = "4.0"
 GPUArraysCore = "= 0.2.0"
+GPUToolbox = "0.2, 0.3, 1"
 JLD2 = "0.4, 0.5, 0.6"
 KernelAbstractions = "0.9.28, 0.10"
 LLVM = "3.9, 4, 5, 6, 7, 8, 9"

src/GPUArrays.jl

@@ -1,5 +1,6 @@
 module GPUArrays
 
+using GPUToolbox
 using KernelAbstractions
 using Serialization
 using Random
@@ -15,7 +16,7 @@ using LLVM.Interop
 using Reexport
 @reexport using GPUArraysCore
 
-using KernelAbstractions
+import KernelAbstractions as KA
 import AcceleratedKernels as AK
 
 # device functionality
@@ -32,6 +33,7 @@ include("host/mapreduce.jl")
 include("host/linalg.jl")
 include("host/math.jl")
 include("host/random.jl")
+include("host/sorting.jl")
 include("host/quirks.jl")
 include("host/uniformscaling.jl")
 include("host/statistics.jl")

src/host/sorting.jl

@@ -0,0 +1,62 @@
+
+abstract type SortingAlgorithm end
+struct MergeSortAlg <: SortingAlgorithm end
+
+const MergeSort = MergeSortAlg()
+
+
+function Base.sort!(c::AnyGPUVector, alg::MergeSortAlg; lt=isless, by=identity, rev=false)
+    # for reverse sorting, invert the less-than function
+    if rev
+        lt = !lt
+    end
+
+    AK.merge_sort!(c; lt, by)
+    return c
+end
+
+function Base.sort!(c::AnyGPUArray; alg::SortingAlgorithm = MergeSort, kwargs...)
+    return sort!(c, alg; kwargs...)
+end
+
+function Base.sort(c::AnyGPUArray; kwargs...)
+    return sort!(copy(c); kwargs...)
+end
+
+function Base.partialsort!(c::AnyGPUVector, k::Union{Integer, OrdinalRange},
+                           alg::MergeSortAlg; lt=isless, by=identity, rev=false)
+
+    sort!(c, alg; lt, by, rev)
+    return @allowscalar copy(c[k])
+end
+
+function Base.partialsort!(c::AnyGPUArray, k::Union{Integer, OrdinalRange};
+                           alg::SortingAlgorithm=MergeSort, kwargs...)
+    return partialsort!(c, k, alg; kwargs...)
+end
+
+function Base.partialsort(c::AnyGPUArray, k::Union{Integer, OrdinalRange}; kwargs...)
+    return partialsort!(copy(c), k; kwargs...)
+end
+
+function Base.sortperm!(ix::AnyGPUArray, A::AnyGPUArray; initialized=false, dims=nothing, kwargs...)
+    if axes(ix) != axes(A)
+        throw(ArgumentError("index array must have the same size/axes as the source array, $(axes(ix)) != $(axes(A))"))
+    end
+    if !isnothing(dims)
+        throw(ArgumentError("GPUArrays sort with `dims` kwarg not yet implemented."))
+    end
+
+    AK.merge_sortperm!(ix, A; kwargs...)
+    return ix
+end
+
+function Base.sortperm(c::AnyGPUVector; initialized=false, kwargs...)
+    AK.merge_sortperm!(KA.allocate(get_backend(c), Int, length(c)), c; kwargs...)
+end
+
+function Base.sortperm(c::AnyGPUArray; dims, kwargs...)
+    # Base errors for Matrices without dims arg, we should too
+    error("GPU sort with `dims` kwarg not yet implemented.")
+    # sortperm!(reshape(adapt(get_backend(c), collect(1:length(c))), size(c)), c; initialized=true, dims, kwargs...)
+end

test/testsuite.jl

@@ -94,6 +94,7 @@ include("testsuite/broadcasting.jl")
 include("testsuite/linalg.jl")
 include("testsuite/math.jl")
 include("testsuite/random.jl")
+include("testsuite/sorting.jl")
 include("testsuite/uniformscaling.jl")
 include("testsuite/statistics.jl")
 include("testsuite/alloc_cache.jl")

test/testsuite/sorting.jl

@@ -0,0 +1,51 @@
+@testsuite "sorting/sort" (AT, eltypes)->begin
+    # Fuzzy correctness testing
+    @testset "$ET" for ET in filter(x -> x <: Real, eltypes)
+        for _ in 1:10
+            num_elems = rand(1:100_000)
+            @test compare((A)->Base.sort!(A), AT, rand(ET, num_elems))
+        end
+        # Not yet implemented
+        # for _ in 1:5
+        #     size = rand(1:100, 2)
+        #     @test compare((A)->Base.sort!(A; dims=1), AT, rand(ET, size...))
+        #     @test compare((A)->Base.sort!(A; dims=2), AT, rand(ET, size...))
+        # end
+    end
+end
+
+@testsuite "sorting/sortperm" (AT, eltypes)->begin
+    # Fuzzy correctness testing
+    @testset "$ET" for ET in filter(x -> x <: Real, eltypes)
+        for _ in 1:10
+            num_elems = rand(1:100_000)
+            @test compare((ix, A)->Base.sortperm!(ix, A), AT, zeros(Int32, num_elems), rand(ET, num_elems))
+        end
+        # Not yet implemented
+        # for _ in 1:5
+        #     size = rand(1:100, 2)
+        #     @test compare((A)->Base.sort!(A; dims=1), AT, zeros(Int32, size...), rand(ET, size...))
+        #     @test compare((A)->Base.sort!(A; dims=2), AT, zeros(Int32, size...), rand(ET, size...))
+        # end
+    end
+end
+
+@testsuite "sorting/partialsort" (AT, eltypes)->begin
+    local N = 10000
+    @testset "$ET" for ET in filter(x -> x <: Real, eltypes)
+        @test compare((A)->Base.partialsort!(A, 1), AT, rand(ET, N))
+        @test compare((A)->Base.partialsort!(A, 1; rev=true), AT, rand(ET, N))
+
+        @test compare((A)->Base.partialsort!(A, N), AT, rand(ET, N))
+        @test compare((A)->Base.partialsort!(A, N; rev=true), AT, rand(ET, N))
+
+        @test compare((A)->Base.partialsort!(A, N÷2), AT, rand(ET, N))
+        @test compare((A)->Base.partialsort!(A, N÷2; rev=true), AT, rand(ET, N))
+
+        @test compare((A)->Base.partialsort!(A, (N÷10):(2N÷10)), AT, rand(ET, N))
+        @test compare((A)->Base.partialsort!(A, (N÷10):(2N÷10); rev=true), AT, rand(ET, N))
+
+        @test compare((A)->Base.partialsort!(A, 1:N), AT, rand(ET, N))
+        @test compare((A)->Base.partialsort!(A, 1:N; rev=true), AT, rand(ET, N))
+    end
+end