Extend CUDACore.jl instead of CUDA.jl

.buildkite/pipeline.yml

@@ -12,7 +12,7 @@ steps:
       julia -e 'using Pkg
 
                 println("--- :julia: Instantiating environment")
-                Pkg.add("CUDA")
+                Pkg.add("CUDACore")
                 Pkg.develop(PackageSpec(name="Atomix", path="."))
 
                 println("+++ :julia: Running tests")

Project.toml

@@ -1,27 +1,29 @@
 name = "Atomix"
 uuid = "a9b6321e-bd34-4604-b9c9-b65b8de01458"
 authors = ["Takafumi Arakaki <aka.tkf@gmail.com> and contributors"]
-version = "1.2.1"
+version = "1.3.0"
 
 [deps]
 UnsafeAtomics = "013be700-e6cd-48c3-b4a1-df204f14c38f"
 
 [weakdeps]
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+CUDACore = "bd0ed864-bdfe-4181-a5ed-ce625a5fdea2"
 Metal = "dde4c033-4e86-420c-a63e-0dd931031962"
-oneAPI = "8f75cd03-7ff8-4ecb-9b8f-daf728133b1b"
 OpenCL = "08131aa3-fb12-5dee-8b74-c09406e224a2"
+oneAPI = "8f75cd03-7ff8-4ecb-9b8f-daf728133b1b"
 
 [extensions]
-AtomixCUDAExt = "CUDA"
+AtomixCUDACoreExt = "CUDACore"
 AtomixMetalExt = "Metal"
-AtomixoneAPIExt = "oneAPI"
 AtomixOpenCLExt = "OpenCL"
+AtomixoneAPIExt = "oneAPI"
 
 [compat]
-CUDA = "5, 6"
+CUDA = "6"
+CUDACore = "6"
 Metal = "1.7"
-oneAPI = "1, 2"
 OpenCL = "^0.10"
 UnsafeAtomics = "0.1, 0.2, 0.3"
 julia = "1.10"
+oneAPI = "1, 2"

ext/AtomixCUDAExt.jl → ext/AtomixCUDACoreExt.jl

@@ -1,7 +1,7 @@
-module AtomixCUDAExt
+module AtomixCUDACoreExt
 
 using Atomix: Atomix, IndexableRef, right
-using CUDA: CUDA, CuDeviceArray
+using CUDACore: CUDACore, CuDeviceArray
 using Core: LLVMPtr
 
 const CuIndexableRef{Indexable<:CuDeviceArray} = IndexableRef{Indexable}
@@ -26,7 +26,7 @@ const NativeFloat = Union{Float32,Float64}
     ptr = Atomix.pointer(ref)
     expected = convert(eltype(ref), expected)
     desired = convert(eltype(ref), desired)
-    old = CUDA.atomic_cas!(ptr, expected, desired)
+    old = CUDACore.atomic_cas!(ptr, expected, desired)
     return (; old = old, success = old === expected)
 end
 
@@ -42,14 +42,14 @@ for (op, fn) in [(+) => :atomic_add!, (-) => :atomic_sub!, (&) => :atomic_and!,
                  (|) => :atomic_or!, xor => :atomic_xor!, min => :atomic_min!,
                  max => :atomic_max!]
     @eval @inline modify_native!(ptr::LLVMPtr{<:NativeInt}, ::typeof($op), x) =
-        CUDA.$fn(ptr, x)
+        CUDACore.$fn(ptr, x)
 end
-@inline modify_native!(ptr::LLVMPtr{Float32}, ::typeof(+), x) = CUDA.atomic_add!(ptr, x)
-@inline modify_native!(ptr::LLVMPtr{Float32}, ::typeof(-), x) = CUDA.atomic_sub!(ptr, x)
+@inline modify_native!(ptr::LLVMPtr{Float32}, ::typeof(+), x) = CUDACore.atomic_add!(ptr, x)
+@inline modify_native!(ptr::LLVMPtr{Float32}, ::typeof(-), x) = CUDACore.atomic_sub!(ptr, x)
 # Float64 atomic add needs compute capability 6.0; use compare-and-swap below that.
 @inline function modify_native!(ptr::LLVMPtr{Float64}, ::typeof(+), x)
-    if CUDA.compute_capability().major >= 6
-        CUDA.atomic_add!(ptr, x)
+    if CUDACore.compute_capability().major >= 6
+        CUDACore.atomic_add!(ptr, x)
     else
         modify_cas!(ptr, +, x)
     end
@@ -58,10 +58,10 @@ end
 
 # swap: exchange floats through their integer representation
 @inline modify_native!(ptr::LLVMPtr{<:NativeInt}, ::typeof(right), x) =
-    CUDA.atomic_xchg!(ptr, x)
+    CUDACore.atomic_xchg!(ptr, x)
 for (T, I) in [Float32 => UInt32, Float64 => UInt64]
     @eval @inline function modify_native!(ptr::LLVMPtr{$T,A}, ::typeof(right), x) where {A}
-        old = CUDA.atomic_xchg!(reinterpret(LLVMPtr{$I,A}, ptr), reinterpret($I, x))
+        old = CUDACore.atomic_xchg!(reinterpret(LLVMPtr{$I,A}, ptr), reinterpret($I, x))
         return reinterpret($T, old)
     end
 end
@@ -73,11 +73,11 @@ end
     old = Base.unsafe_load(ptr)
     while true
         new = convert(T, op(old, x))
-        seen = CUDA.atomic_cas!(ptr, old, new)
+        seen = CUDACore.atomic_cas!(ptr, old, new)
         # bitwise comparison: `==` would spin forever on NaN
         seen === old && return old
         old = seen
     end
 end
 
-end  # module AtomixCUDAExt
+end  # module AtomixCUDACoreExt

test/runtests.jl

@@ -45,7 +45,7 @@ end
     mutable struct Atomic{T}
         @atomic x::T
     end
-    
+
     a = Atomic(123)
     @test (@atomic a.x) == 123
     @test (@atomic :monotonic a.x) == 123
@@ -143,7 +143,7 @@ if "--Metal" in ARGS
     include("test_atomix_metal.jl")
 elseif "--CUDA" in ARGS
     import Pkg
-    Pkg.add("CUDA")
+    Pkg.add("CUDACore")
     include("test_atomix_cuda.jl")
 elseif "--oneAPI" in ARGS
     import Pkg

test/test_atomix_cuda.jl

@@ -1,9 +1,9 @@
-using CUDA
-using CUDA: @allowscalar
+using CUDACore
+using CUDACore: @allowscalar
 
 
-@testset "AtomixCUDAExt:extension_found" begin
-    @test !isnothing(Base.get_extension(Atomix, :AtomixCUDAExt))
+@testset "AtomixCUDACoreExt:extension_found" begin
+    @test !isnothing(Base.get_extension(Atomix, :AtomixCUDACoreExt))
 end
 
 
@@ -12,13 +12,13 @@ function cuda(f)
         f()
         nothing
     end
-    CUDA.@cuda g()
+    @cuda g()
 end
 
 
 
 
-@testset "AtomixCUDAExt:test_cas" begin
+@testset "AtomixCUDACoreExt:test_cas" begin
     idx = (
         data = 1,
         cas1_ok = 2,
@@ -28,7 +28,7 @@ end
     @assert minimum(idx) >= 1
     @assert maximum(idx) == length(idx)
 
-    A = CUDA.zeros(Int, length(idx))
+    A = CUDACore.zeros(Int, length(idx))
     cuda() do
         GC.@preserve A begin
             ref = Atomix.IndexableRef(A, (1,))
@@ -42,8 +42,8 @@ end
 end
 
 
-@testset "AtomixCUDAExt:test_inc" begin
-    A = CUDA.CuVector(1:3)
+@testset "AtomixCUDACoreExt:test_inc" begin
+    A = CuVector(1:3)
     cuda() do
         GC.@preserve A begin
             ref = Atomix.IndexableRef(A, (1,))
@@ -56,8 +56,8 @@ end
 end
 
 
-@testset "AtomixCUDAExt:test_inc_sugar" begin
-    A = CUDA.ones(Int, 3)
+@testset "AtomixCUDACoreExt:test_inc_sugar" begin
+    A = CUDACore.ones(Int, 3)
     cuda() do
         GC.@preserve A begin
             @atomic A[begin] += 1
@@ -67,8 +67,8 @@ end
 end
 
 
-@testset "AtomixCUDAExt:test_get_set" begin
-    A = CUDA.ones(Int, 3)
+@testset "AtomixCUDACoreExt:test_get_set" begin
+    A = CUDACore.ones(Int, 3)
     cuda() do
         GC.@preserve A begin
             ref = Atomix.IndexableRef(A, (1,))
@@ -82,8 +82,8 @@ end
 end
 
 
-@testset "AtomixCUDAExt:test_swap" begin
-    A = CUDA.CuVector(Int[1, 0, 0])
+@testset "AtomixCUDACoreExt:test_swap" begin
+    A = CuVector(Int[1, 0, 0])
     cuda() do
         GC.@preserve A begin
             ref = Atomix.IndexableRef(A, (1,))
@@ -95,8 +95,8 @@ end
 end
 
 
-@testset "AtomixCUDAExt:test_ordering" begin
-    A = CUDA.ones(Int, 2)
+@testset "AtomixCUDACoreExt:test_ordering" begin
+    A = CUDACore.ones(Int, 2)
     cuda() do
         GC.@preserve A begin
             @atomic :monotonic A[1] += 1
@@ -107,8 +107,8 @@ end
 end
 
 
-@testset "AtomixCUDAExt:test_float" begin
-    A = CUDA.CuVector(Float32[1, 1, 1, 1, 1, 0, 0, 1])
+@testset "AtomixCUDACoreExt:test_float" begin
+    A = CuVector(Float32[1, 1, 1, 1, 1, 0, 0, 1])
     cuda() do
         GC.@preserve A begin
             @atomic A[1] += 1.5f0
@@ -125,8 +125,8 @@ end
 end
 
 
-@testset "AtomixCUDAExt:test_float64" begin
-    A = CUDA.CuVector(Float64[1, 1, 1, 1, 1, 0, 0, 1])
+@testset "AtomixCUDACoreExt:test_float64" begin
+    A = CuVector(Float64[1, 1, 1, 1, 1, 0, 0, 1])
     cuda() do
         GC.@preserve A begin
             @atomic A[1] += 1.5