Support GPUArrays 12

Project.toml

@@ -32,7 +32,7 @@ SPIRVIntrinsics = {path = "lib/intrinsics"}
 
 [compat]
 Adapt = "4"
-GPUArrays = "11.2.1"
+GPUArrays = "12"
 GPUCompiler = "2.13"
 GPUToolbox = "3.3.2"
 KernelAbstractions = "0.9.38"

src/OpenCL.jl

@@ -45,7 +45,6 @@ include("compiler/precompile.jl")
 # integrations and specialized functionality
 include("util.jl")
 include("broadcast.jl")
-include("mapreduce.jl")
 include("gpuarrays.jl")
 include("random.jl")
 

src/gpuarrays.jl

@@ -5,19 +5,3 @@ function GPUArrays.derive(::Type{T}, a::CLArray, dims::Dims{N}, offset::Int) whe
     offset = a.offset + offset * sizeof(T)
     CLArray{T,N}(ref, dims; offset)
 end
-
-const GLOBAL_RNGs = Dict{cl.Device,GPUArrays.RNG}()
-const global_rngs_lock = ReentrantLock()
-
-function GPUArrays.default_rng(::Type{<:CLArray})
-    dev = cl.device()
-    return Base.@lock global_rngs_lock begin
-        get!(GLOBAL_RNGs, dev) do
-            N = dev.max_work_group_size
-            state = CLArray{NTuple{4, UInt32}}(undef, N)
-            rng = GPUArrays.RNG(state)
-            Random.seed!(rng)
-            rng
-        end
-    end
-end

src/random.jl

@@ -1,6 +1,15 @@
 using Random
 
-gpuarrays_rng() = GPUArrays.default_rng(CLArray)
+const GLOBAL_RNGs = Dict{cl.Device,GPUArrays.RNG{CLArray}}()
+const global_rngs_lock = ReentrantLock()
+
+# one RNG per device, used by the RNG-less `rand!`/`randn!` methods and `seed!`
+function gpuarrays_rng()
+    dev = cl.device()
+    return Base.@lock global_rngs_lock begin
+        get!(() -> GPUArrays.RNG{CLArray}(), GLOBAL_RNGs, dev)
+    end
+end
 
 # GPUArrays in-place
 Random.rand!(A::WrappedCLArray) = Random.rand!(gpuarrays_rng(), A)

test/array.jl

@@ -239,12 +239,6 @@ end
     @test length(b) == 1
 end
 
-@testset "mapreducedim! returning same type" begin
-    R = transpose(OpenCL.zeros(Float32, 2, 3))
-    A = CLArray(rand(Float32, 3, 2, 10))
-    @test @inferred(OpenCL.GPUArrays.mapreducedim!(identity, +, R, A)) === R
-end
-
 # finalizers run in no particular order, e.g. at exit (JuliaGPU/OpenCL.jl#279), so memory
 # has to remain freeable after the queue and context it was allocated with are finalized
 @testset "freeing after finalizing its queue and context" begin

test/device/random.jl

@@ -167,7 +167,7 @@ if Float16 in GPUArraysTestSuite.supported_eltypes(CLArray)
     end
 
     @testset "randn!(Complex{Float16}) is finite" begin
-        rng = OpenCL.GPUArrays.default_rng(CLArray)
+        rng = OpenCL.GPUArrays.RNG{CLArray}()
         Random.seed!(rng, 1)
         A = CLArray{Complex{Float16}}(undef, 4096)
         randn!(rng, A)