From 15d75b9461e095c6a49e82d201dc78e09caa9249 Mon Sep 17 00:00:00 2001
From: Valentin Churavy <v.churavy@gmail.com>
Date: Tue, 13 May 2025 11:36:04 +0200
Subject: [PATCH 1/4] CUDA: Fallback to UnsafeAtomics

---
 ext/AtomixCUDAExt.jl | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/ext/AtomixCUDAExt.jl b/ext/AtomixCUDAExt.jl
index fc92895..838e130 100644
--- a/ext/AtomixCUDAExt.jl
+++ b/ext/AtomixCUDAExt.jl
@@ -1,17 +1,19 @@
 # TODO: respect ordering
 module AtomixCUDAExt
 
-using Atomix: Atomix, IndexableRef
+using Atomix: Atomix, IndexableRef, UnsafeAtomics
 using CUDA: CUDA, CuDeviceArray
 
 const CuIndexableRef{Indexable<:CuDeviceArray} = IndexableRef{Indexable}
 
 function Atomix.get(ref::CuIndexableRef, order)
-    error("not implemented")
+    ptr = Atomix.pointer(ref)
+    return UnsafeAtomics.load(ptr, order)
 end
 
 function Atomix.set!(ref::CuIndexableRef, v, order)
-    error("not implemented")
+    ptr = Atomix.pointer(ref)
+    return UnsafeAtomics.store!(ptr, v, order)
 end
 
 @inline function Atomix.replace!(
@@ -51,7 +53,7 @@ end
         elseif op === Atomix.right
             CUDA.atomic_xchg!(ptr, x)
         else
-            error("not implemented")
+            return UnsafeAtomics.modify(ptr, op, x)
         end
     end
     return old => op(old, x)

From 19bd807579184554c8570862bf4a02a8b6cd5576 Mon Sep 17 00:00:00 2001
From: Valentin Churavy <v.churavy@gmail.com>
Date: Tue, 13 May 2025 11:57:50 +0200
Subject: [PATCH 2/4] use monotonic order

---
 ext/AtomixCUDAExt.jl | 54 +++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 51 insertions(+), 3 deletions(-)

diff --git a/ext/AtomixCUDAExt.jl b/ext/AtomixCUDAExt.jl
index 838e130..48d6d5e 100644
--- a/ext/AtomixCUDAExt.jl
+++ b/ext/AtomixCUDAExt.jl
@@ -6,14 +6,62 @@ using CUDA: CUDA, CuDeviceArray
 
 const CuIndexableRef{Indexable<:CuDeviceArray} = IndexableRef{Indexable}
 
+# from https://github.com/JuliaGPU/CUDA.jl/pull/1644
+# function atomic_load(ptr::LLVMPtr{T}, order, scope::System=System()) where T
+#     if order == Acq_Rel() || order == Release()
+#         assert(false)
+#     end
+#     if compute_capability() >= sv"7.0"
+#         if order == Relaxed()
+#             val = __load(ptr, Relaxed(), scope)
+#             return val
+#         end
+#         if order == Seq_Cst()
+#             atomic_thread_fence(Seq_Cst(), scope)
+#         end
+#         val = __load(ptr, Acquire(), scope)
+#         return val
+#     else
+#         if order == Seq_Cst()
+#             atomic_thread_fence(Seq_Cst(), scope)
+#         end
+#         val = __load_volatile(ptr)
+#         if order == Relaxed()
+#             return val
+#         end
+#         atomic_thread_fence(order, scope)
+#         return val
+#     end
+# end
+
 function Atomix.get(ref::CuIndexableRef, order)
     ptr = Atomix.pointer(ref)
-    return UnsafeAtomics.load(ptr, order)
+    return UnsafeAtomics.load(ptr, UnsafeAtomics.monotonic)
 end
 
+# function atomic_store!(ptr::LLVMPtr{T}, val::T, order, scope::System=System()) where T
+#     if order == Acq_Rel() || order == Consume() || order == Acquire()
+#         assert(false)
+#     end
+#     if compute_capability() >= sv"7.0"
+#         if order == Release()
+#             __store!(ptr, val, Release(), scope)
+#             return
+#         end
+#         if order == Seq_Cst()
+#             atomic_thread_fence(Seq_Cst(), scope)
+#         end
+#         __store!(ptr, val, Relaxed(), scope)
+#     else
+#         if order == Seq_Cst()
+#             atomic_thread_fence(Seq_Cst(), scope)
+#         end
+#         __store_volatile!(ptr, val)
+#     end
+# end
 function Atomix.set!(ref::CuIndexableRef, v, order)
     ptr = Atomix.pointer(ref)
-    return UnsafeAtomics.store!(ptr, v, order)
+    return UnsafeAtomics.store!(ptr, v, UnsafeAtomics.monotonic)
 end
 
 @inline function Atomix.replace!(
@@ -53,7 +101,7 @@ end
         elseif op === Atomix.right
             CUDA.atomic_xchg!(ptr, x)
         else
-            return UnsafeAtomics.modify(ptr, op, x)
+            return UnsafeAtomics.modify(ptr, op, x, UnsafeAtomics.monotonic)
         end
     end
     return old => op(old, x)

From 16bd0fec61bf36e21846a6f8e8b62162abc80ede Mon Sep 17 00:00:00 2001
From: Valentin Churavy <v.churavy@gmail.com>
Date: Tue, 13 May 2025 12:03:58 +0200
Subject: [PATCH 3/4] UnsafeAtomics defined in Internal

---
 ext/AtomixCUDAExt.jl | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ext/AtomixCUDAExt.jl b/ext/AtomixCUDAExt.jl
index 48d6d5e..1948430 100644
--- a/ext/AtomixCUDAExt.jl
+++ b/ext/AtomixCUDAExt.jl
@@ -1,7 +1,8 @@
 # TODO: respect ordering
 module AtomixCUDAExt
 
-using Atomix: Atomix, IndexableRef, UnsafeAtomics
+using Atomix: Atomix, IndexableRef
+using Atomix.Internal: UnsafeAtomics
 using CUDA: CUDA, CuDeviceArray
 
 const CuIndexableRef{Indexable<:CuDeviceArray} = IndexableRef{Indexable}

From a8ec317ee48e089e17e71a298286a36a52e53380 Mon Sep 17 00:00:00 2001
From: Valentin Churavy <v.churavy@gmail.com>
Date: Tue, 13 May 2025 12:05:35 +0200
Subject: [PATCH 4/4] enable test

---
 test/test_atomix_cuda.jl | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/test/test_atomix_cuda.jl b/test/test_atomix_cuda.jl
index 487eae2..dd153db 100644
--- a/test/test_atomix_cuda.jl
+++ b/test/test_atomix_cuda.jl
@@ -15,10 +15,7 @@ function cuda(f)
     CUDA.@cuda g()
 end
 
-
-# Not implemented:
-#=
-function test_get_set()
+@testset "AtomixCUDAExt:test_get_set" begin
     A = CUDA.ones(Int, 3)
     cuda() do
         GC.@preserve A begin
@@ -29,8 +26,6 @@ function test_get_set()
     end
     @test collect(A) == [-1, 1, 1]
 end
-=#
-
 
 @testset "AtomixCUDAExt:test_cas" begin
     idx = (