Use isequal for generic atomic updates with compare-and-swap (#1300)

Prevents potential deadlocks with `NaN`.

Co-authored-by: Tim Besard <[email protected]>

Author: guyvdbroeck

Diff for: src/device/intrinsics/atomics.jl

@@ -225,7 +225,7 @@ end
         cmp = old
         new = convert(T, op(old, val))
         old = atomic_cas!(ptr, cmp, new)
-        (old == cmp) && return new
+        isequal(old, cmp) && return new
     end
 end
 

Diff for: test/device/intrinsics/atomics.jl

@@ -398,6 +398,24 @@ end
 end
 
 @testset "macro" begin
+    
+    @testset "NaN" begin
+        f(x,y) = 3x + 2y
+        
+        function kernel(x)
+            CUDA.@atomic x[1] = f(x[1],42f0) 
+            nothing
+        end
+
+        a = CuArray([0f0])
+        @cuda kernel(a)
+        @test Array(a)[1] ≈ 84
+            
+        a = CuArray([NaN32])
+        @cuda kernel(a)
+        @test isnan(Array(a)[1])
+    end
+            
     using CUDA: AtomicError
 
     @test_throws_macro AtomicError("right-hand side of an @atomic assignment should be a call") @macroexpand begin