Merge branch 'develop' into reduceLoops

Author: liu15

RELEASE_NOTES.md

@@ -12,6 +12,26 @@ in this file.
 
 The format of this file is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
+## [Unreleased] - Release date YYYY-MM-DD
+
+### Added
+- Added CARE\_DEEP\_COPY\_RAW\_PTR configuration option.
+
+### Added
+- 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.
+
+### Changed
+- Renamed host\_device\_ptr::getPointer to host\_device\_ptr::data.
+
+### Fixed
+- Replaced calls to chai::ManagedArray::getPointer (previously deprecated and now removed) with calls to chai::ManagedArray::data.
+
 ## [Version 0.14.1] - Release date 2024-10-15
 
 ### Fixed

cmake/SetupOptions.cmake

@@ -19,6 +19,7 @@ option(CARE_ENABLE_IMPLICIT_CONVERSIONS "Enable implicit conversions to-from raw
 # 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)
@@ -32,6 +33,8 @@ 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"

src/care/Accessor.h

@@ -28,12 +28,12 @@ class CHAIDataGetter {
       T * getRawArrayData(chai::ManagedArray<T> data) {
          data.move(chai::CPU);
          data.registerTouch(chai::CPU);
-         return (T*)data.getPointer(chai::CPU);
+         return (T*)data.data(chai::CPU);
       }
 
       const T * getConstRawArrayData(chai::ManagedArray<T> data) {
          data.move(chai::CPU);
-         return (const T*)data.getPointer(chai::CPU);
+         return (const T*)data.data(chai::CPU);
       }
 
       static const auto ChaiPolicy = chai::CPU;
@@ -49,12 +49,12 @@ class CHAIDataGetter<T, RAJADeviceExec> {
       T * getRawArrayData(chai::ManagedArray<T> data) {
          data.move(chai::GPU);
          data.registerTouch(chai::GPU);
-         return (T*)data.getPointer(chai::GPU);
+         return (T*)data.data(chai::GPU);
       }
 
       const T * getConstRawArrayData(chai::ManagedArray<T> data) {
          data.move(chai::GPU);
-         return (const T*)data.getPointer(chai::GPU);
+         return (const T*)data.data(chai::GPU);
       }
 
       static const auto ChaiPolicy = chai::GPU;
@@ -70,12 +70,12 @@ class CHAIDataGetter<globalID, RAJADeviceExec> {
       GIDTYPE * getRawArrayData(chai::ManagedArray<globalID> data) {
          data.move(chai::GPU);
          data.registerTouch(chai::GPU);
-         return (GIDTYPE*)data.getPointer(chai::GPU);
+         return (GIDTYPE*)data.data(chai::GPU);
       }
 
       const GIDTYPE * getConstRawArrayData(chai::ManagedArray<globalID> data) {
          data.move(chai::GPU);
-         return (GIDTYPE*)data.getPointer(chai::GPU);
+         return (GIDTYPE*)data.data(chai::GPU);
       }
 
       static const auto ChaiPolicy = chai::GPU;
@@ -95,12 +95,12 @@ class CHAIDataGetter<globalID, RAJA::seq_exec> {
       GIDTYPE * getRawArrayData(chai::ManagedArray<globalID> data) {
          data.move(chai::CPU);
          data.registerTouch(chai::CPU);
-         return (GIDTYPE*)data.getPointer(chai::CPU);
+         return (GIDTYPE*)data.data(chai::CPU);
       }
 
       const GIDTYPE * getConstRawArrayData(chai::ManagedArray<globalID> data) {
          data.move(chai::CPU);
-         return (GIDTYPE*)data.getPointer(chai::CPU);
+         return (GIDTYPE*)data.data(chai::CPU);
       }
 
       static const auto ChaiPolicy = chai::CPU;

src/care/CHAIDataGetter.h

@@ -17,7 +17,6 @@ configure_file(
 
 set(care_headers
     ${PROJECT_BINARY_DIR}/include/care/config.h
-    Accessor.h
     algorithm.h
     algorithm_decl.h
     algorithm_impl.h

src/care/CMakeLists.txt

@@ -68,7 +68,6 @@ namespace care{
              it.second(space, PluginData::getFileName(), PluginData::getLineNumber());
           }
           PluginData::clear_post_parallel_forall_actions();
-          PluginData::s_threadID = -1;
       }
 #endif // !defined(CHAI_DISABLE_RM)
    }

src/care/DebugPlugin.cpp

@@ -34,13 +34,6 @@
 /// Used to capture variables by reference into a lambda (combine with FOR_EACH)
 #define CARE_REF_CAPTURE(X) , &X
 
-#ifdef CARE_ENABLE_RACE_DETECTION
-#define CARE_SET_THREAD_ID(INDEX) care::DebugPlugin::s_threadID = INDEX ;
-#else
-#define CARE_SET_THREAD_ID(INDEX)
-#endif
-
-
 
 
 
@@ -732,8 +725,7 @@ OMP_FOR_BEGIN for (auto INDEX = _care_openmp_for_loop_chunk_begin_ndx; INDEX < _
    decltype(_care_checked_loop_end) _care_checked_loop_begin = START_INDEX; \
    if (_care_checked_loop_end > _care_checked_loop_begin) { \
       CARE_NEST_BEGIN(CHECK) \
-      care::forall(POLICY{}, __FILE__, __LINE__, _care_checked_loop_begin, _care_checked_loop_end, 0, [=] CARE_DEVICE (decltype(_care_checked_loop_end) INDEX) { \
-         CARE_SET_THREAD_ID(INDEX)
+      care::forall(POLICY{}, __FILE__, __LINE__, _care_checked_loop_begin, _care_checked_loop_end, 0, [=] CARE_DEVICE (decltype(_care_checked_loop_end) INDEX) {
 
 #define CARE_CHECKED_POLICY_LOOP_END(CHECK) }); \
    CARE_NEST_END(CHECK) }}
@@ -768,8 +760,7 @@ OMP_FOR_BEGIN for (auto INDEX = _care_openmp_for_loop_chunk_begin_ndx; INDEX < _
    decltype(_care_checked_loop_end) _care_checked_loop_begin = START_INDEX; \
    if (_care_checked_loop_end > _care_checked_loop_begin) { \
       CARE_NEST_BEGIN(CHECK) \
-      care::forall(POLICY{}, __FILE__, __LINE__, _care_checked_loop_begin, _care_checked_loop_end, CHUNK_SIZE, [=] CARE_DEVICE (decltype(_care_checked_loop_end) INDEX) { \
-         CARE_SET_THREAD_ID(INDEX)
+      care::forall(POLICY{}, __FILE__, __LINE__, _care_checked_loop_begin, _care_checked_loop_end, CHUNK_SIZE, [=] CARE_DEVICE (decltype(_care_checked_loop_end) INDEX) {
 
 #define CARE_CHECKED_CHUNKED_POLICY_LOOP_END(CHECK) }); \
    CARE_NEST_END(CHECK) }}

src/care/DefaultMacros.h

@@ -13,7 +13,6 @@ namespace care{
    CARE_DLL_API bool PluginData::s_parallel_context = false;
    CARE_DLL_API ActionMap PluginData::s_post_parallel_forall_actions = ActionMap{};
    CARE_DLL_API std::vector<const chai::PointerRecord*> PluginData::s_active_pointers_in_loop = std::vector<const chai::PointerRecord*>{};
-   CARE_DLL_API int PluginData::s_threadID = -1;
 
    void PluginData::setFileName(const char * name) {PluginData::s_file_name = name;}
 

src/care/PluginData.cpp

@@ -22,7 +22,7 @@ namespace care{
    using ActionMap = std::unordered_map<void *, std::function<void(chai::ExecutionSpace, const char *, int)>>;
 
    //class for shared plugin functions and variables
-   CARE_DLL_API class PluginData	{
+   class CARE_DLL_API PluginData	{
       public:
          PluginData() = default;
 
@@ -54,8 +54,6 @@ namespace care{
 
          static void clearActivePointers();
 
-         static int s_threadID;
-
       private:
          static const char * s_file_name;
 

src/care/PluginData.h

@@ -299,8 +299,8 @@ namespace care {
       // set up a 2D kernel, put per-array meta-data in pinned memory to eliminate cudaMemcpy's of the smaller dimension of data
       host_device_ptr<int> lengths(chai::ManagedArray<int>(m_num_arrays, chai::ZERO_COPY));
       host_device_ptr<host_device_ptr<int> > out_arrays(chai::ManagedArray<host_device_ptr<int>>(m_num_arrays, chai::ZERO_COPY));
-      host_ptr<int> pinned_lengths = lengths.getPointer(care::ZERO_COPY, false);
-      host_ptr<host_device_ptr<int>>  pinned_out_arrays = out_arrays.getPointer(care::ZERO_COPY, false);
+      host_ptr<int> pinned_lengths = lengths.data(care::ZERO_COPY, false);
+      host_ptr<host_device_ptr<int>>  pinned_out_arrays = out_arrays.data(care::ZERO_COPY, false);
       // initialized lengths, maxLength, and array of arrays for the 2D kernel
       int maxLength = 0;
       for (int a = 0; a < m_num_arrays; ++a ) {
@@ -316,7 +316,7 @@ namespace care {
       }
       // subtract out the offset, copy the result into individual arrays
       // (use of device pointer is to avoid clang-query rules that prevent capture of raw pointer)
-      device_ptr<int> dev_pinned_lengths = lengths.getPointer(ZERO_COPY, false);
+      device_ptr<int> dev_pinned_lengths = lengths.data(ZERO_COPY, false);
       CARE_LOOP_2D_STREAM_JAGGED(i, 0, maxLength, lengths, a, 0, m_num_arrays, iFlattened)  {
          result[i+out_offsets[a]] -= max_range*a;
          out_arrays[a][i] = result[i+out_offsets[a]];