Add function get_device

.gitignore

@@ -1 +1,4 @@
 docs/build/
+
+.vscode
+

Project.toml

@@ -6,7 +6,9 @@ version = "0.8.0-dev"
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
+SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 UUIDs = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
 

examples/matmul.jl

@@ -24,11 +24,9 @@ function matmul!(a, b, c)
         println("Matrix size mismatch!")
         return nothing
     end
-    if isa(a, Array)
-        kernel! = matmul_kernel!(CPU(),4)
-    else
-        kernel! = matmul_kernel!(CUDADevice(),256)
-    end
+    device = KernelAbstractions.get_device(a)
+    n = device isa GPU ? 256 : 4
+    kernel! = matmul_kernel!(device, n)
     kernel!(a, b, c, ndrange=size(c)) 
 end
 

examples/naive_transpose.jl

@@ -17,17 +17,9 @@ function naive_transpose!(a, b)
         println("Matrix size mismatch!")
         return nothing
     end
-
-    if isa(a, Array)
-        kernel! = naive_transpose_kernel!(CPU(), 4)
-    elseif isa(a, CuArray)
-        kernel! = naive_transpose_kernel!(CUDADevice(), 256)
-    elseif isa(a, ROCArray)
-        kernel! = naive_transpose_kernel!(ROCDevice(), 256)
-    else
-        println("Unrecognized array type!")
-    end
-
+    device = KernelAbstractions.get_device(a)
+    n = device isa GPU ? 256 : 4
+    kernel! = naive_transpose_kernel!(device, n)
     kernel!(a, b, ndrange=size(a))
 end
 

lib/CUDAKernels/src/CUDAKernels.jl

@@ -8,6 +8,9 @@ import KernelAbstractions
 
 export CUDADevice
 
+KernelAbstractions.get_device(::CUDA.CuArray) = CUDADevice()
+KernelAbstractions.get_device(::CUDA.CUSPARSE.AbstractCuSparseArray) = CUDADevice()
+
 const FREE_STREAMS = CUDA.CuStream[]
 const STREAMS = CUDA.CuStream[]
 const STREAM_GC_THRESHOLD = Ref{Int}(16)

lib/CUDAKernels/test/Project.toml

@@ -4,8 +4,10 @@ Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 KernelGradients = "e5faadeb-7f6c-408e-9747-a7a26e81c66a"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MPI = "da04e1cc-30fd-572f-bb4f-1f8673147195"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
+SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

lib/CUDAKernels/test/runtests.jl

@@ -4,6 +4,7 @@ using Enzyme
 using CUDA
 using CUDAKernels
 using Test
+using SparseArrays
 
 include(joinpath(dirname(pathof(KernelAbstractions)), "..", "test", "testsuite.jl"))
 include(joinpath(dirname(pathof(KernelGradients)), "..", "test", "testsuite.jl"))

lib/ROCKernels/src/ROCKernels.jl

@@ -9,6 +9,9 @@ import KernelAbstractions
 
 export ROCDevice
 
+KernelAbstractions.get_device(::AMDGPU.ROCArray) = ROCDevice()
+
+
 const FREE_QUEUES = HSAQueue[]
 const QUEUES = HSAQueue[]
 const QUEUE_GC_THRESHOLD = Ref{Int}(16)

lib/ROCKernels/test/Project.toml

@@ -4,8 +4,10 @@ Enzyme = "7da242da-08ed-463a-9acd-ee780be4f1d9"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 KernelGradients = "e5faadeb-7f6c-408e-9747-a7a26e81c66a"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MPI = "da04e1cc-30fd-572f-bb4f-1f8673147195"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
+SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

src/KernelAbstractions.jl

@@ -6,7 +6,9 @@ export Device, GPU, CPU, Event, MultiEvent, NoneEvent
 export async_copy!
 
 
+using LinearAlgebra
 using MacroTools
+using SparseArrays
 using StaticArrays
 using Adapt
 
@@ -336,6 +338,23 @@ abstract type GPU <: Device end
 
 struct CPU <: Device end
 
+
+"""
+    KernelAbstractions.get_device(A::AbstractArray)::KernelAbstractions.Device
+
+Get a `KernelAbstractions.Device` instance suitable for array `A`.
+"""
+function get_device end
+
+# Should cover SubArray, ReshapedArray, ReinterpretArray, Hermitian, AbstractTriangular, etc.:
+get_device(A::AbstractArray) = get_device(parent(A))
+
+get_device(A::AbstractSparseArray) = get_device(rowvals(A))
+get_device(A::Diagonal) = get_device(A.diag)
+get_device(A::Tridiagonal) = get_device(A.d)
+
+get_device(::Array) = CPU()
+
 include("nditeration.jl")
 using .NDIteration
 import .NDIteration: get

test/Project.toml

@@ -1,7 +1,9 @@
 [deps]
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MPI = "da04e1cc-30fd-572f-bb4f-1f8673147195"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
+SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

test/test.jl

@@ -1,6 +1,8 @@
 using KernelAbstractions
 using KernelAbstractions.NDIteration
 using InteractiveUtils
+using LinearAlgebra
+using SparseArrays
 import SpecialFunctions
 
 identity(x) = x
@@ -64,6 +66,20 @@ end
        A[I] = i
 end
 
+@testset "get_device" begin
+    x = ArrayT(rand(Float32, 5))
+    A = ArrayT(rand(Float32, 5,5))
+    device = backend()
+    @test @inferred(KernelAbstractions.get_device(A)) == device
+    @test @inferred(KernelAbstractions.get_device(view(A, 2:4, 1:3))) == device
+    if !(isdefined(Main, :ROCKernels) && (device isa Main.ROCKernels.ROCDevice))
+        # Sparse arrays are not supported by the ROCm backend yet:
+        @test @inferred(KernelAbstractions.get_device(sparse(A))) == device
+    end
+    @test @inferred(KernelAbstractions.get_device(Diagonal(x))) == device
+    @test @inferred(KernelAbstractions.get_device(Tridiagonal(A))) == device
+end
+
 @testset "indextest" begin
     # TODO: add test for _group and _local_cartesian
     A = ArrayT{Int}(undef, 16, 16)