Merge remote-tracking branch 'origin/develop' into grouped_gemm_args_const_buff

include/ck/utility/generic_memory_space_atomic.hpp

@@ -28,6 +28,12 @@ __device__ float atomic_add<float>(float* p_dst, const float& x)
     return atomicAdd(p_dst, x);
 }
 
+template <>
+__device__ double atomic_add<double>(double* p_dst, const double& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
 template <>
 __device__ float2_t atomic_add<float2_t>(float2_t* p_dst, const float2_t& x)
 {
@@ -45,6 +51,23 @@ __device__ float2_t atomic_add<float2_t>(float2_t* p_dst, const float2_t& x)
     return vy.template AsType<float2_t>()[I0];
 }
 
+template <>
+__device__ double2_t atomic_add<double2_t>(double2_t* p_dst, const double2_t& x)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    const vector_type<double, 2> vx{x};
+    vector_type<double, 2> vy{0};
+
+    vy.template AsType<double>()(I0) =
+        atomicAdd(c_style_pointer_cast<double*>(p_dst), vx.template AsType<double>()[I0]);
+    vy.template AsType<double>()(I1) =
+        atomicAdd(c_style_pointer_cast<double*>(p_dst) + 1, vx.template AsType<double>()[I1]);
+
+    return vy.template AsType<double2_t>()[I0];
+}
+
 // Caution: DO NOT REMOVE
 // intentionally have only declaration but no definition to cause compilation failure when trying to
 // instantiate this template. The purpose is to make the implementation of atomic_max explicit for