Switch to ParallelTestRunner.

.buildkite/pipeline.yml

@@ -7,13 +7,10 @@ steps:
           version: "1.10"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("CUDA")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--CUDA"])'
+                Pkg.test("AcceleratedKernels", test_args=["--cuda"])'
     agents:
       queue: "juliagpu"
       cuda: "*"
@@ -26,13 +23,10 @@ steps:
           version: "1.11"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("CUDA")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--CUDA"])'
+                Pkg.test("AcceleratedKernels", test_args=["--cuda"])'
     agents:
       queue: "juliagpu"
       cuda: "*"
@@ -47,13 +41,10 @@ steps:
           version: "1.10"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("AMDGPU")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--AMDGPU"])'
+                Pkg.test("AcceleratedKernels", test_args=["--amdgpu"])'
     agents:
       queue: "juliagpu"
       rocm: "*"
@@ -67,13 +58,10 @@ steps:
           version: "1.11"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("AMDGPU")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--AMDGPU"])'
+                Pkg.test("AcceleratedKernels", test_args=["--amdgpu"])'
     agents:
       queue: "juliagpu"
       rocm: "*"
@@ -89,13 +77,10 @@ steps:
           version: "1.10"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("oneAPI")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--oneAPI"])'
+                Pkg.test("AcceleratedKernels", test_args=["--oneapi"])'
     agents:
       queue: "juliagpu"
       intel: "*"
@@ -108,13 +93,10 @@ steps:
           version: "1.11"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("oneAPI")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--oneAPI"])'
+                Pkg.test("AcceleratedKernels", test_args=["--oneapi"])'
     agents:
       queue: "juliagpu"
       intel: "*"
@@ -129,13 +111,10 @@ steps:
           version: "1.10"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("Metal")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--Metal"])'
+                Pkg.test("AcceleratedKernels", test_args=["--metal"])'
     agents:
       queue: "juliaecosystem"
       os: "macos"
@@ -149,13 +128,10 @@ steps:
           version: "1.11"
     command: |
       julia -e 'using Pkg
-
                 println("--- :julia: Instantiating environment")
-                Pkg.add("Metal")
                 Pkg.develop(path=".")
-
                 println("+++ :julia: Running tests")
-                Pkg.test("AcceleratedKernels", test_args=["--Metal"])'
+                Pkg.test("AcceleratedKernels", test_args=["--metal"])'
     agents:
       queue: "juliaecosystem"
       os: "macos"

.github/workflows/CI-CPU.yml

@@ -82,7 +82,7 @@ jobs:
       - uses: julia-actions/julia-buildpkg@v1
       - uses: julia-actions/julia-runtest@v1
         with:
-          test_args: '--OpenCL'
+          test_args: '--opencl'
   # cpuKA:
   #   name: KA CPU Backend
   #   runs-on: ubuntu-latest
@@ -102,7 +102,7 @@ jobs:
   #     - uses: julia-actions/julia-buildpkg@v1
   #     - uses: julia-actions/julia-runtest@v1
   #       with:
-  #         test_args: '--cpuKA'
+  #         test_args: '--cpu-ka'
   docs:
     name: Documentation
     runs-on: ubuntu-latest

test/Project.toml

@@ -1,7 +1,17 @@
 [deps]
+AMDGPU = "21141c5a-9bdb-4563-92ae-f87d6854732e"
 Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
+Metal = "dde4c033-4e86-420c-a63e-0dd931031962"
+OpenCL = "08131aa3-fb12-5dee-8b74-c09406e224a2"
+ParallelTestRunner = "d3525ed8-44d0-4b2c-a655-542cee43accc"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+oneAPI = "8f75cd03-7ff8-4ecb-9b8f-daf728133b1b"
+pocl_jll = "627d6b7a-bbe6-5189-83e7-98cc0a5aeadd"
+
+[compat]
+ParallelTestRunner = "2"

test/aqua.jl

@@ -0,0 +1,4 @@
+using Aqua
+@testset "Aqua" begin
+    Aqua.test_all(AK)
+end

test/runtests.jl

@@ -1,76 +1,143 @@
 import AcceleratedKernels as AK
-using KernelAbstractions
-using Test
-using Random
-import Pkg
-
-# Set to true when testing backends that support this
-const TEST_DL = Ref{Bool}(false)
-
-# Pass command-line argument to test suite to install the right backend, e.g.
-#   julia> import Pkg
-#   julia> Pkg.test(test_args=["--oneAPI"])
-if "--CUDA" in ARGS
-    Pkg.add("CUDA")
+using ParallelTestRunner
+
+const init_code = quote
+    import AcceleratedKernels as AK
+    using KernelAbstractions
+    using Test
+    using Random
+end
+
+# Discover root-level tests (aqua.jl, partition.jl) and generic tests
+const testsuite = find_tests(@__DIR__)
+const generic_tests = find_tests(joinpath(@__DIR__, "generic"))
+
+# Parse args with lowercase hyphenated backend flags
+args = parse_args(ARGS; custom=["cuda", "amdgpu", "metal", "oneapi", "opencl", "cpu-ka"])
+
+# Common helper code appended to every backend setup
+const _array_from_host_code = quote
+    global array_from_host
+    array_from_host(h_arr::AbstractArray, dtype=nothing) = array_from_host(BACKEND, h_arr, dtype)
+    function array_from_host(backend, h_arr::AbstractArray, dtype=nothing)
+        d_arr = KernelAbstractions.zeros(backend, isnothing(dtype) ? eltype(h_arr) : dtype, size(h_arr))
+        copyto!(d_arr, h_arr isa Array ? h_arr : Array(h_arr))
+        d_arr
+    end
+end
+
+# Build list of active backends, each with setup code
+backends = Pair{String, Expr}[]
+
+# CPU always active
+using InteractiveUtils
+@info "Julia information:\n" * sprint(InteractiveUtils.versioninfo)
+push!(backends, "cpu" => quote
+    global BACKEND = get_backend([])
+    global IS_CPU_BACKEND = true
+    global prefer_threads = true
+    global TEST_DL = Ref{Bool}(false)
+    $_array_from_host_code
+end)
+
+# cpu-ka only when --cpu-ka flag passed
+if args.custom["cpu-ka"] !== nothing
+    push!(backends, "cpu-ka" => quote
+        global BACKEND = get_backend([])
+        global IS_CPU_BACKEND = true
+        global prefer_threads = false
+        global TEST_DL = Ref{Bool}(false)
+        $_array_from_host_code
+    end)
+end
+
+# GPU backends: auto-detect functional ones, or enable explicitly via CLI flags.
+# Always assert functional() before proceeding, then print versioninfo() once.
+
+if try; using CUDA; CUDA.functional(); catch; false; end || args.custom["cuda"] !== nothing
     using CUDA
-    CUDA.versioninfo()
-    const BACKEND = CUDABackend()
-    TEST_DL[] = true
-elseif "--oneAPI" in ARGS
-    Pkg.add("oneAPI")
-    using oneAPI
-    oneAPI.versioninfo()
-    const BACKEND = oneAPIBackend()
+    @assert CUDA.functional()
+    @info "CUDA information:\n" * sprint(CUDA.versioninfo)
+    push!(backends, "cuda" => quote
+        using CUDA
+        global BACKEND = CUDABackend()
+        global IS_CPU_BACKEND = false
+        global prefer_threads = true
+        global TEST_DL = Ref{Bool}(true)
+        $_array_from_host_code
+    end)
+end
 
-    # FIXME: need atomic orderings for `DecoupledLookback` in oneAPI
-    # TEST_DL[] = true
-elseif "--AMDGPU" in ARGS
-    Pkg.add("AMDGPU")
+if try; using AMDGPU; AMDGPU.functional(); catch; false; end || args.custom["amdgpu"] !== nothing
     using AMDGPU
-    AMDGPU.versioninfo()
-    const BACKEND = ROCBackend()
-    TEST_DL[] = true
-elseif "--Metal" in ARGS
-    Pkg.add("Metal")
+    @assert AMDGPU.functional()
+    @info "AMDGPU information:\n" * sprint(AMDGPU.versioninfo)
+    push!(backends, "amdgpu" => quote
+        using AMDGPU
+        global BACKEND = ROCBackend()
+        global IS_CPU_BACKEND = false
+        global prefer_threads = true
+        global TEST_DL = Ref{Bool}(true)
+        $_array_from_host_code
+    end)
+end
+
+if try; using Metal; Metal.functional(); catch; false; end || args.custom["metal"] !== nothing
     using Metal
-    Metal.versioninfo()
-    const BACKEND = MetalBackend()
-elseif "--OpenCL" in ARGS
-    Pkg.add(name="OpenCL", rev="master")
-    Pkg.add(name="SPIRVIntrinsics", rev="master")
-    Pkg.add("pocl_jll")
-    using pocl_jll
-    using OpenCL
-    OpenCL.versioninfo()
-    const BACKEND = OpenCLBackend()
-elseif !@isdefined(BACKEND)
-    # Otherwise do CPU tests
-    using InteractiveUtils
-    InteractiveUtils.versioninfo()
-    const BACKEND = get_backend([])
+    @assert Metal.functional()
+    @info "Metal information:\n" * sprint(Metal.versioninfo)
+    push!(backends, "metal" => quote
+        using Metal
+        global BACKEND = MetalBackend()
+        global IS_CPU_BACKEND = false
+        global prefer_threads = true
+        global TEST_DL = Ref{Bool}(false)
+        $_array_from_host_code
+    end)
 end
 
-const IS_CPU_BACKEND = BACKEND == get_backend([])
+if try; using oneAPI; oneAPI.functional(); catch; false; end || args.custom["oneapi"] !== nothing
+    using oneAPI
+    @assert oneAPI.functional()
+    @info "oneAPI information:\n" * sprint(oneAPI.versioninfo)
+    push!(backends, "oneapi" => quote
+        using oneAPI
+        global BACKEND = oneAPIBackend()
+        global IS_CPU_BACKEND = false
+        global prefer_threads = true
+        global TEST_DL = Ref{Bool}(false)
+        $_array_from_host_code
+    end)
+end
 
-global prefer_threads::Bool = !(IS_CPU_BACKEND && "--cpuKA" in ARGS)
+if try; using pocl_jll, OpenCL; !isempty(OpenCL.cl.platforms()); catch; false; end || args.custom["opencl"] !== nothing
+    using pocl_jll, OpenCL
+    @assert !isempty(OpenCL.cl.platforms())
+    @info "OpenCL information:\n" * sprint(OpenCL.versioninfo)
+    push!(backends, "opencl" => quote
+        using pocl_jll
+        using OpenCL
+        global BACKEND = OpenCLBackend()
+        global IS_CPU_BACKEND = false
+        global prefer_threads = true
+        global TEST_DL = Ref{Bool}(false)
+        $_array_from_host_code
+    end)
+end
 
-array_from_host(h_arr::AbstractArray, dtype=nothing) = array_from_host(BACKEND, h_arr, dtype)
-function array_from_host(backend, h_arr::AbstractArray, dtype=nothing)
-    d_arr = KernelAbstractions.zeros(backend, isnothing(dtype) ? eltype(h_arr) : dtype, size(h_arr))
-    copyto!(d_arr, h_arr isa Array ? h_arr : Array(h_arr))      # Allow unmaterialised types, e.g. ranges
-    d_arr
+# Duplicate generic tests per active backend
+for (backend_name, setup_code) in backends
+    for (test_name, test_body) in generic_tests
+        testsuite["$backend_name/$test_name"] = quote
+            $setup_code
+            $test_body
+        end
+    end
 end
 
-@testset "Aqua" begin
-    using Aqua
-    Aqua.test_all(AK)
+# Filter tests by user-specified positional args; remove bare generic/ entries if no filter was specified
+if filter_tests!(testsuite, args)
+    filter!(((k,v),) -> !startswith(k, "generic/"), testsuite)
 end
 
-include("partition.jl")
-include("looping.jl")
-include("map.jl")
-include("sort.jl")
-include("reduce.jl")
-include("accumulate.jl")
-include("predicates.jl")
-include("binarysearch.jl")
+runtests(AK, args; init_code, testsuite)