Merge branch 'develop' into woptim/rsc-update

RELEASE_NOTES.md

@@ -15,13 +15,14 @@ The format of this file is based on [Keep a Changelog](http://keepachangelog.com
 ## [Unreleased] - Release date YYYY-MM-DD
 
 ### Added
+- Added CARE\_DEEP\_COPY\_RAW\_PTR configuration option.
 - Added ATOMIC\_SUB, ATOMIC\_LOAD, ATOMIC\_STORE, ATOMIC\_EXCHANGE, and ATOMIC\_CAS macros.
-- Added TSAN\_ONLY\_ATOMIC\_\* macros to suppress tsan data race reports. Controlled by CARE\_ENABLE\_TSAN\_ONLY\_ATOMICS configuration option.
 
 ### Removed
 - Removed Accessor template parameter from host\_device\_ptr.
 - Removed NoOpAccessor and RaceConditionAccessor. It is recommended to use ThreadSanitizer (TSAN) instead to locate race conditions.
 - Removed CARE\_ENABLE\_RACE\_DETECTION configuration option.
+- Removed implicit conversions between raw pointers and host\_device\_ptrs/host\_ptrs and the corresponding CARE\_ENABLE\_IMPLICIT\_CONVERSIONS configuration option.
 
 ### Changed
 - Renamed host\_device\_ptr::getPointer to host\_device\_ptr::data.

cmake/SetupOptions.cmake

@@ -14,11 +14,8 @@ option(ENABLE_PICK "Enable pick and set methods on ManagedArrays" ON)
 option(ENABLE_PINNED "Enable pinned memory space" ON)
 option(CARE_ENABLE_PINNED_MEMORY_FOR_SCANS "Use pinned memory for scan lengths" ON)
 option(CARE_GPU_MEMORY_IS_ACCESSIBLE_ON_CPU "Allows default memory spaces for ZERO_COPY and PAGEABLE to be the GPU memory space" OFF)
-# Option to disable implicit conversion between host_device_ptr and raw arrays in CARE.
-option(CARE_ENABLE_IMPLICIT_CONVERSIONS "Enable implicit conversions to-from raw pointers" ON)
-# CHAI must also be configured with the same settings for implicit conversions.
-set(CHAI_ENABLE_IMPLICIT_CONVERSIONS ${CARE_ENABLE_IMPLICIT_CONVERSIONS} CACHE BOOL "Enable implicit conversions to-from raw pointers")
 option(CARE_LEGACY_COMPATIBILITY_MODE "Enable legacy compatibility mode" OFF)
+option(CARE_DEEP_COPY_RAW_PTR "Use deep copy for managed array initialization from raw pointer" OFF)
 option(CARE_ENABLE_MANAGED_PTR "Enable managed_ptr aliases, tests, and reproducer" ON)
 option(CARE_DISABLE_RAJAPLUGIN "Disable use of the RAJA plugin. WILL ALSO DISABLE MEMORY MOTION." OFF)
 option(CARE_ENABLE_EXTERN_INSTANTIATE "Enable extern instantiation of template functions" OFF)
@@ -32,8 +29,6 @@ option(CARE_NEVER_USE_RAJA_PARALLEL_SCAN "Disable RAJA parallel scans in SCAN lo
 option(CARE_ENABLE_FUSER_BIN_32 "Enable the 32 register fusible loop bin." OFF)
 option(CARE_ENABLE_PARALLEL_LOOP_BACKWARDS "Reverse the start and end for parallel loops." OFF)
 option(CARE_ENABLE_STALE_DATA_CHECK "Enable checking for stale host data. Only applicable for GPU (or GPU simulation) builds." OFF)
-# TODO: Investigate correctness and performance impact of this option
-option(CARE_ENABLE_TSAN_ONLY_ATOMICS "Enable atomics for ThreadSanitizer (TSAN) build." OFF)
 
 # Extra components
 cmake_dependent_option(CARE_ENABLE_TESTS "Build CARE tests"

configs/lc/toss_4_x86_64_ib_cray/amdclang.cmake

@@ -13,5 +13,3 @@ set(ENABLE_HIP ON CACHE BOOL "Enable Hip")
 set(ROCM_PATH "/usr/tce/packages/rocmcc/rocmcc-6.1.0-magic" CACHE PATH "")
 set(CMAKE_HIP_ARCHITECTURES "gfx942:xnack+" CACHE STRING "")
 set(AMDGPU_TARGETS "gfx942:xnack+" CACHE STRING "")
-
-set(CARE_ENABLE_IMPLICIT_CONVERSIONS OFF CACHE BOOL "Enable implicit conversions")

src/care/DefaultMacros.h

@@ -736,7 +736,7 @@ OMP_FOR_BEGIN for (auto INDEX = _care_openmp_for_loop_chunk_begin_ndx; INDEX < _
 #define CARE_CHECKED_PARALLEL_LOOP_END(CHECK) CARE_CHECKED_POLICY_LOOP_END(CHECK)
 
 #define CARE_CHECKED_REDUCE_LOOP_START(INDEX, START_INDEX, END_INDEX, CHECK) \
-   CARE_CHECKED_POLICY_LOOP_START(care::parallel_reduce,INDEX, START_INDEX, END_INDEX, CHECK)
+   CARE_CHECKED_POLICY_LOOP_START(care::gpu_reduce,INDEX, START_INDEX, END_INDEX, CHECK)
 
 #define CARE_CHECKED_REDUCE_LOOP_END(CHECK) CARE_CHECKED_POLICY_LOOP_END(CHECK)
 
@@ -771,7 +771,7 @@ OMP_FOR_BEGIN for (auto INDEX = _care_openmp_for_loop_chunk_begin_ndx; INDEX < _
 #define CARE_CHECKED_CHUNKED_PARALLEL_LOOP_END(CHECK) CARE_CHECKED_CHUNKED_POLICY_LOOP_END(CHECK)
 
 #define CARE_CHECKED_CHUNKED_REDUCE_LOOP_START(INDEX, START_INDEX, END_INDEX, CHUNK_SIZE, CHECK) \
-   CARE_CHECKED_CHUNKED_POLICY_LOOP_START(care::parallel_reduce,INDEX, START_INDEX, END_INDEX, CHUNK_SIZE, CHECK)
+   CARE_CHECKED_CHUNKED_POLICY_LOOP_START(care::gpu_reduce,INDEX, START_INDEX, END_INDEX, CHUNK_SIZE, CHECK)
 
 #define CARE_CHECKED_CHUNKED_REDUCE_LOOP_END(CHECK) CARE_CHECKED_CHUNKED_POLICY_LOOP_END(CHECK)
 
@@ -1278,6 +1278,9 @@ OMP_FOR_BEGIN for (auto INDEX = _care_openmp_for_loop_chunk_begin_ndx; INDEX < _
    launch_2D_jagged(care::gpu{}, XSTART, XEND, XLENGTHS.data(chai::DEFAULT, true), YSTART, YLENGTH, __FILE__, __LINE__, [=] CARE_DEVICE (int XINDEX, int YINDEX)->void  {
 #define CARE_LOOP_2D_STREAM_JAGGED_END });
 
+#define CARE_LOOP_2D_REDUCE_JAGGED(XINDEX, XSTART, XEND, XLENGTHS, YINDEX, YSTART, YLENGTH, FLAT_INDEX)  \
+   launch_2D_jagged(care::gpu_reduce{}, XSTART, XEND, XLENGTHS.data(chai::DEFAULT, true), YSTART, YLENGTH, __FILE__, __LINE__, [=] CARE_DEVICE (int XINDEX, int YINDEX)->void  {
+#define CARE_LOOP_2D_REDUCE_JAGGED_END });
 
 #endif // !defined(_CARE_DEFAULT_MACROS_H_)
 

src/care/KeyValueSorter_decl.h

@@ -174,31 +174,6 @@ class CARE_DLL_API KeyValueSorter<KeyType, ValueType, RAJADeviceExec> {
          setKeyValueArraysFromManagedArray(m_keys, m_values, len, arr);
       }
 
-#if defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
-      ///////////////////////////////////////////////////////////////////////////
-      /// @author Alan Dayton
-      ///
-      /// @brief Constructor
-      ///
-      /// Allocates space and initializes the KeyValueSorter by copying
-      ///    elements and ordering from the given managed array
-      ///
-      /// @note This overload is needed to prevent ambiguity when implicit
-      ///       casts are enabled
-      ///
-      /// @param[in] len - The number of elements to allocate space for
-      /// @param[in] arr - The managed array to copy elements from
-      ///
-      /// @return a KeyValueSorter instance
-      ///////////////////////////////////////////////////////////////////////////
-      KeyValueSorter<KeyType, ValueType, RAJADeviceExec>(const size_t len, const host_device_ptr<ValueType> & arr)
-      : KeyValueSorter<KeyType, ValueType, RAJADeviceExec>(len, host_device_ptr<const ValueType>(arr))
-      {
-      }
-
-#endif // defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
       ///////////////////////////////////////////////////////////////////////////
       /// @author Alan Dayton
       /// @brief (Shallow) Copy constructor
@@ -758,32 +733,6 @@ class CARE_DLL_API KeyValueSorter<KeyType, ValueType, RAJA::seq_exec> {
          setKeyValueArraysFromManagedArray(m_keyValues, len, arr);
       }
 
-#if defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
-      ///////////////////////////////////////////////////////////////////////////
-      /// @author Alan Dayton
-      ///
-      /// @brief Constructor
-      ///
-      /// Allocates space and initializes the KeyValueSorter by copying
-      ///    elements and ordering from the given managed array
-      ///
-      /// @note This overload is needed to prevent ambiguity when implicit
-      ///       casts are enabled
-      ///
-      /// @param[in] len - The number of elements to allocate space for
-      /// @param[in] arr - The managed array to copy elements from
-      ///
-      /// @return a KeyValueSorter instance
-      ///
-      ///////////////////////////////////////////////////////////////////////////
-      KeyValueSorter<KeyType, ValueType, RAJA::seq_exec>(const size_t len, const host_device_ptr<ValueType> & arr)
-      : KeyValueSorter<KeyType, ValueType, RAJA::seq_exec>(len, host_device_ptr<const ValueType>(arr))
-      {
-      }
-
-#endif // defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
       ///////////////////////////////////////////////////////////////////////////
       /// @author Alan Dayton
       /// @brief (Shallow) Copy constructor

src/care/LoopFuser.h

@@ -1273,9 +1273,14 @@ void LoopFuser<REGISTER_COUNT, XARGS...>::registerAction(const char * fileName,
 #define FUSIBLE_LOOP_STREAM_R_END \
                               } }); } FUSIBLE_FLUSH_IF_NEEDED
 
+#define FUSIBLE_REDUCE_LOOP_R FUSIBLE_LOOP_STREAM_R
+#define FUSIBLE_REDUCE_LOOP_R_END FUSIBLE_LOOP_STREAM_R_END
+
 #define FUSIBLE_LOOP_STREAM(INDEX, START, END) FUSIBLE_LOOP_STREAM_R(INDEX, START, END, CARE_DEFAULT_LOOP_FUSER_REGISTER_COUNT)
 #define FUSIBLE_LOOP_STREAM_END FUSIBLE_LOOP_STREAM_R_END
 
+#define FUSIBLE_REDUCE_LOOP FUSIBLE_LOOP_STREAM
+#define FUSIBLE_REDUCE_LOOP_END FUSIBLE_LOOP_STREAM_END
 
 #define FUSIBLE_KERNEL_R(REGISTER_COUNT) { \
    auto __fuser__ = LOOPFUSER(REGISTER_COUNT)::getInstance(); \
@@ -1304,9 +1309,14 @@ void LoopFuser<REGISTER_COUNT, XARGS...>::registerAction(const char * fileName,
                                     } \
                                     }); }}
 
+#define FUSIBLE_REDUCE_LOOP_PHASE_R FUSIBLE_LOOP_PHASE_R
+#define FUSIBLE_REDUCE_LOOP_PHASE_R_END FUSIBLE_LOOP_PHASE_R_END
+
 #define FUSIBLE_LOOP_PHASE(INDEX, START, END, PRIORITY) FUSIBLE_LOOP_PHASE_R(INDEX, START, END, PRIORITY, CARE_DEFAULT_LOOP_FUSER_REGISTER_COUNT)
 #define FUSIBLE_LOOP_PHASE_END FUSIBLE_LOOP_PHASE_R_END
 
+#define FUSIBLE_REDUCE_LOOP_PHASE FUSIBLE_LOOP_PHASE
+#define FUSIBLE_REDUCE_LOOP_PHASE_END FUSIBLE_LOOP_PHASE_END
 
 #define FUSIBLE_KERNEL_PHASE_R(PRIORITY, REGISTER_COUNT) { \
    LOOPFUSER(REGISTER_COUNT) * __fuser__ = FusedActionsObserver::getActiveObserver()->getFusedActions<LOOPFUSER(REGISTER_COUNT)>(PRIORITY); \
@@ -1382,21 +1392,32 @@ void LoopFuser<REGISTER_COUNT, XARGS...>::registerAction(const char * fileName,
 #define FUSIBLE_LOOP_SCAN(INDEX, START, END, POS, INIT_POS, BOOL_EXPR) \
    FUSIBLE_LOOP_SCAN_R(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, CARE_DEFAULT_LOOP_FUSER_REGISTER_COUNT)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_R FUSIBLE_LOOP_SCAN_R
+#define FUSIBLE_REDUCE_LOOP_SCAN FUSIBLE_LOOP_SCAN
+
 #define _FUSIBLE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION) } return 0; }, 1, POS_STORE_DESTINATION); }
 #define FUSIBLE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION) _FUSIBLE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION) FUSIBLE_FLUSH_IF_NEEDED
 
 #define FUSIBLE_LOOP_SCAN_END(LENGTH, POS, POS_STORE_DESTINATION) FUSIBLE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_R_END FUSIBLE_LOOP_SCAN_R_END
+#define FUSIBLE_REDUCE_LOOP_SCAN_END FUSIBLE_LOOP_SCAN_END
+
 #define FUSIBLE_LOOP_SCAN_PHASE_R(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, PRIORITY, REGISTER_COUNT) \
    _FUSIBLE_LOOP_SCAN_R(FusedActionsObserver::getActiveObserver()->getFusedActions<LOOPFUSER(REGISTER_COUNT)>(PRIORITY), \
                         INDEX, START, END, POS, INIT_POS, BOOL_EXPR, REGISTER_COUNT)
 
 #define FUSIBLE_LOOP_SCAN_PHASE(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, PRIORITY) \
    FUSIBLE_LOOP_SCAN_PHASE_R(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, PRIORITY, CARE_DEFAULT_LOOP_FUSER_REGISTER_COUNT)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE_R FUSIBLE_LOOP_SCAN_PHASE_R
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE FUSIBLE_LOOP_SCAN_PHASE
+
 #define FUSIBLE_LOOP_SCAN_PHASE_END(LENGTH, POS, POS_STORE_DESTINATION) _FUSIBLE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION)
 #define FUSIBLE_LOOP_SCAN_PHASE_R_END(LENGTH, POS, POS_STORE_DESTINATION) _FUSIBLE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE_R_END FUSIBLE_LOOP_SCAN_PHASE_R_END
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE_END FUSIBLE_LOOP_SCAN_PHASE_END
 
 // note - FUSED_SCANVAR will be nullptr if we are not recording, as there will be no need for an intermediate
 // FUSED_SCANVAR, so we won't need to write to it in the action or store into it in the conditional
@@ -1466,13 +1487,21 @@ void LoopFuser<REGISTER_COUNT, XARGS...>::registerAction(const char * fileName,
 
 #define FUSIBLE_LOOP_STREAM_R(INDEX, START, END, REGISTER_COUNT) CARE_STREAM_LOOP(INDEX, START, END)
 #define FUSIBLE_LOOP_STREAM(INDEX, START, END) CARE_STREAM_LOOP(INDEX, START, END)
+#define FUSIBLE_REDUCE_LOOP_R(INDEX, START, END, REGISTER_COUNT) CARE_REDUCE_LOOP(INDEX, START, END)
+#define FUSIBLE_REDUCE_LOOP(INDEX, START, END) CARE_REDUCE_LOOP(INDEX, START, END)
 
 #define FUSIBLE_LOOP_PHASE_R(INDEX, START, END, PRIORITY, REGISTER_COUNT) CARE_STREAM_LOOP(INDEX, START, END)
 #define FUSIBLE_LOOP_PHASE(INDEX, START, END, PRIORITY) CARE_STREAM_LOOP(INDEX, START, END)
 
 #define FUSIBLE_LOOP_PHASE_END CARE_STREAM_LOOP_END
 #define FUSIBLE_LOOP_PHASE_R_END CARE_STREAM_LOOP_END
 
+#define FUSIBLE_REDUCE_LOOP_PHASE_R(INDEX, START, END, PRIORITY, REGISTER_COUNT) CARE_STREAM_LOOP(INDEX, START, END)
+#define FUSIBLE_REDUCE_LOOP_PHASE(INDEX, START, END, PRIORITY) CARE_STREAM_LOOP(INDEX, START, END)
+
+#define FUSIBLE_REDUCE_LOOP_PHASE_END CARE_REDUCE_LOOP_END
+#define FUSIBLE_REDUCE_LOOP_PHASE_R_END CARE_REDUCE_LOOP_END
+
 #define FUSIBLE_FLUSH_IF_NEEDED
 #define FUSIBLE_PHASE_RESET
 
@@ -1485,6 +1514,9 @@ void LoopFuser<REGISTER_COUNT, XARGS...>::registerAction(const char * fileName,
 #define FUSIBLE_LOOP_STREAM_R_END  CARE_STREAM_LOOP_END
 #define FUSIBLE_LOOP_STREAM_END  CARE_STREAM_LOOP_END
 
+#define FUSIBLE_REDUCE_LOOP_R_END  CARE_REDUCE_LOOP_END
+#define FUSIBLE_REDUCE_LOOP_END  CARE_REDUCE_LOOP_END
+
 #define FUSIBLE_KERNEL_PHASE_R_END CARE_PARALLEL_KERNEL_END
 #define FUSIBLE_KERNEL_R_END CARE_PARALLEL_KERNEL_END
 #define FUSIBLE_KERNEL_PHASE_END CARE_PARALLEL_KERNEL_END
@@ -1501,16 +1533,29 @@ void LoopFuser<REGISTER_COUNT, XARGS...>::registerAction(const char * fileName,
 #define FUSIBLE_LOOP_SCAN_R(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, REGISTER_COUNT) SCAN_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
 #define FUSIBLE_LOOP_SCAN(INDEX, START, END, POS, INIT_POS, BOOL_EXPR) SCAN_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_R(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, REGISTER_COUNT) SCAN_REDUCE_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
+#define FUSIBLE_REDUCE_LOOP_SCAN(INDEX, START, END, POS, INIT_POS, BOOL_EXPR) SCAN_REDUCE_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
+
 #define FUSIBLE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
 #define FUSIBLE_LOOP_SCAN_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_R_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_REDUCE_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
+#define FUSIBLE_REDUCE_LOOP_SCAN_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_REDUCE_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
+
 #define FUSIBLE_LOOP_SCAN_PHASE_R(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, PRIORITY, REGISTER_COUNT) \
     SCAN_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
 #define FUSIBLE_LOOP_SCAN_PHASE(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, PRIORITY) SCAN_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE_R(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, PRIORITY, REGISTER_COUNT) \
+    SCAN_REDUCE_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE(INDEX, START, END, POS, INIT_POS, BOOL_EXPR, PRIORITY) SCAN_REDUCE_LOOP(INDEX, START, END, POS, INIT_POS, BOOL_EXPR)
+
 #define FUSIBLE_LOOP_SCAN_PHASE_R_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
 #define FUSIBLE_LOOP_SCAN_PHASE_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
 
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE_R_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_REDUCE_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
+#define FUSIBLE_REDUCE_LOOP_SCAN_PHASE_END(LENGTH, POS, POS_STORE_DESTINATION) SCAN_REDUCE_LOOP_END(LENGTH, POS, POS_STORE_DESTINATION)
+
 #define FUSIBLE_FREE(A) A.free()
 #define FUSIBLE_FREE_DEVICE(A,DEST,ELEM) care::wrappedFreeDeviceMemory(A, DEST, ELEM);
 

src/care/algorithm_decl.h

@@ -250,16 +250,6 @@ CARE_HOST_DEVICE bool checkSorted(const care::host_device_ptr<const T>& array, c
                                   const bool allowDuplicates = false,
                                   const bool warnOnFailure = true);
 
-#if defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
-template <typename T>
-CARE_HOST_DEVICE bool checkSorted(const care::host_device_ptr<T>& array, const int len,
-                                  const char* name, const char* argname,
-                                  const bool allowDuplicates = false,
-                                  const bool warnOnFailure = true);
-
-#endif // defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
 template<typename mapType>
 CARE_HOST_DEVICE CARE_DLL_API int BinarySearch(const mapType *map,
                                                const int start,

src/care/algorithm_impl.h

@@ -103,21 +103,6 @@ CARE_HOST_DEVICE CARE_INLINE bool checkSorted(const care::host_device_ptr<const
    return checkSorted<T>(array.data(), len, name, argname, allowDuplicates, warnOnFailure);
 }
 
-#if defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
-template <typename T>
-CARE_HOST_DEVICE CARE_INLINE bool checkSorted(const care::host_device_ptr<T>& array,
-                                              const int len,
-                                              const char* name,
-                                              const char* argname,
-                                              const bool allowDuplicates,
-                                              const bool warnOnFailure)
-{
-   return checkSorted(care::host_device_ptr<const T>(array), len, name, argname, allowDuplicates, warnOnFailure);
-}
-
-#endif // defined(CARE_ENABLE_IMPLICIT_CONVERSIONS)
-
 /************************************************************************
  * Function  : IntersectArrays<A,RAJAExec>
  * Author(s) : Peter Robinson, based on IntersectGlobalIDArrays by Al Nichols