Fix the HLL type deduction bug

Author: PointKernel

include/cuco/detail/hyperloglog/hyperloglog_impl.cuh

@@ -54,9 +54,9 @@ namespace cuco::detail {
 template <class T, cuda::thread_scope Scope, class Hash>
 class hyperloglog_impl {
   // We use `int` here since this is the smallest type that supports native `atomicMax` on GPUs
-  using fp_type = double;  ///< Floating point type used for reduction
-  using hash_value_type =
-    decltype(cuda::std::declval<Hash>()(cuda::std::declval<T>()));  ///< Hash value type
+  using fp_type         = double;  ///< Floating point type used for reduction
+  using hash_value_type = cuda::std::remove_cvref_t<decltype(cuda::std::declval<Hash>()(
+    cuda::std::declval<T>()))>;  ///< Hash value type
  public:
   static constexpr auto thread_scope = Scope;  ///< CUDA thread scope
 

tests/CMakeLists.txt

@@ -146,7 +146,8 @@ ConfigureTest(DYNAMIC_BITSET_TEST
 ConfigureTest(HYPERLOGLOG_TEST
     hyperloglog/unique_sequence_test.cu
     hyperloglog/spark_parity_test.cu
-    hyperloglog/device_ref_test.cu)
+    hyperloglog/device_ref_test.cu
+    hyperloglog/type_deduction_test.cu)
 
 ###################################################################################################
 # - bloom_filter ----------------------------------------------------------------------------------