Fix clang format conflicts un unchecked policy branch

include/RAJA/pattern/launch/launch_core.hpp

@@ -690,6 +690,21 @@ RAJA_HOST_DEVICE RAJA_INLINE void loop(CONTEXT const& ctx,
                                                        body);
 }
 
+RAJA_SUPPRESS_HD_WARN
+template<typename POLICY_LIST,
+         typename CONTEXT,
+         typename SEGMENT,
+         typename BODY>
+RAJA_HOST_DEVICE RAJA_INLINE void loop_icount(CONTEXT const& ctx,
+                                              SEGMENT const& segment0,
+                                              SEGMENT const& segment1,
+                                              BODY const& body)
+{
+
+  LoopICountExecute<loop_policy<POLICY_LIST>, SEGMENT>::exec(ctx, segment0,
+                                                             segment1, body);
+}
+
 RAJA_SUPPRESS_HD_WARN
 template<typename POLICY_LIST,
          typename CONTEXT,
@@ -796,6 +811,46 @@ RAJA_HOST_DEVICE RAJA_INLINE void tile_tcount(CONTEXT const& ctx,
       ctx, tile_size0, tile_size1, segment0, segment1, body);
 }
 
+template<typename POLICY_LIST,
+         typename CONTEXT,
+         typename TILE_T,
+         typename SEGMENT,
+         typename BODY>
+RAJA_HOST_DEVICE RAJA_INLINE void tile(CONTEXT const& ctx,
+                                       TILE_T tile_size0,
+                                       TILE_T tile_size1,
+                                       TILE_T tile_size2,
+                                       SEGMENT const& segment0,
+                                       SEGMENT const& segment1,
+                                       SEGMENT const& segment2,
+                                       BODY const& body)
+{
+
+  TileExecute<loop_policy<POLICY_LIST>, SEGMENT>::exec(
+      ctx, tile_size0, tile_size1, tile_size2, segment0, segment1, segment2,
+      body);
+}
+
+template<typename POLICY_LIST,
+         typename CONTEXT,
+         typename TILE_T,
+         typename SEGMENT,
+         typename BODY>
+RAJA_HOST_DEVICE RAJA_INLINE void tile_tcount(CONTEXT const& ctx,
+                                              TILE_T tile_size0,
+                                              TILE_T tile_size1,
+                                              TILE_T tile_size2,
+                                              SEGMENT const& segment0,
+                                              SEGMENT const& segment1,
+                                              SEGMENT const& segment2,
+                                              BODY const& body)
+{
+
+  TileTCountExecute<loop_policy<POLICY_LIST>, SEGMENT>::exec(
+      ctx, tile_size0, tile_size1, tile_size2, segment0, segment1, segment2,
+      body);
+}
+
 }  // namespace expt
 
 }  // namespace RAJA

include/RAJA/policy/cuda/kernel/For.hpp

@@ -29,6 +29,65 @@ namespace RAJA
 namespace internal
 {
 
+/*
+ * Executor for work sharing inside CudaKernel.
+ * Mapping without checking from IndexMapper to indices
+ * Assigns the loop index to offset ArgumentId
+ * Meets all sync requirements
+ */
+template<typename Data,
+         camp::idx_t ArgumentId,
+         typename IndexMapper,
+         kernel_sync_requirement sync,
+         typename... EnclosedStmts,
+         typename Types>
+struct CudaStatementExecutor<
+    Data,
+    statement::For<
+        ArgumentId,
+        RAJA::policy::cuda::
+            cuda_indexer<iteration_mapping::DirectUnchecked, sync, IndexMapper>,
+        EnclosedStmts...>,
+    Types>
+{
+
+  using stmt_list_t = StatementList<EnclosedStmts...>;
+
+  // Set the argument type for this loop
+  using NewTypes = setSegmentTypeFromData<Types, ArgumentId, Data>;
+
+  using enclosed_stmts_t =
+      CudaStatementListExecutor<Data, stmt_list_t, NewTypes>;
+
+  using diff_t = segment_diff_type<ArgumentId, Data>;
+
+  using DimensionCalculator = RAJA::internal::KernelDimensionCalculator<
+      RAJA::policy::cuda::
+          cuda_indexer<iteration_mapping::DirectUnchecked, sync, IndexMapper>>;
+
+  static inline RAJA_DEVICE void exec(Data& data, bool thread_active)
+  {
+    const diff_t i = IndexMapper::template index<diff_t>();
+
+    // Assign the index to the argument
+    data.template assign_offset<ArgumentId>(i);
+
+    // execute enclosed statements
+    enclosed_stmts_t::exec(data, thread_active);
+  }
+
+  static inline LaunchDims calculateDimensions(Data const& data)
+  {
+    const diff_t len = segment_length<ArgumentId>(data);
+
+    LaunchDims dims = DimensionCalculator::get_dimensions(len);
+
+    LaunchDims enclosed_dims = enclosed_stmts_t::calculateDimensions(data);
+
+    return combine(dims, enclosed_dims);
+  }
+};
+
 /*
  * Executor for work sharing inside CudaKernel.
  * Mapping directly from IndexMapper to indices

include/RAJA/policy/cuda/kernel/ForICount.hpp

@@ -29,6 +29,67 @@ namespace RAJA
 namespace internal
 {
 
+/*
+ * Executor for work sharing inside CudaKernel.
+ * Provides a direct unchecked mapping.
+ * Assigns the loop index to offset ArgumentId
+ * Assigns the loop index to param ParamId
+ * Meets all sync requirements
+ */
+template<typename Data,
+         camp::idx_t ArgumentId,
+         typename ParamId,
+         typename IndexMapper,
+         kernel_sync_requirement sync,
+         typename... EnclosedStmts,
+         typename Types>
+struct CudaStatementExecutor<
+    Data,
+    statement::ForICount<
+        ArgumentId,
+        ParamId,
+        RAJA::policy::cuda::
+            cuda_indexer<iteration_mapping::DirectUnchecked, sync, IndexMapper>,
+        EnclosedStmts...>,
+    Types>
+    : CudaStatementExecutor<
+          Data,
+          statement::For<ArgumentId,
+                         RAJA::policy::cuda::cuda_indexer<
+                             iteration_mapping::DirectUnchecked,
+                             sync,
+                             IndexMapper>,
+                         EnclosedStmts...>,
+          Types>
+{
+
+  using Base = CudaStatementExecutor<
+      Data,
+      statement::For<
+          ArgumentId,
+          RAJA::policy::cuda::cuda_indexer<iteration_mapping::DirectUnchecked,
+                                           sync,
+                                           IndexMapper>,
+          EnclosedStmts...>,
+      Types>;
+
+  using typename Base::diff_t;
+  using typename Base::enclosed_stmts_t;
+
+  static inline RAJA_DEVICE void exec(Data& data, bool thread_active)
+  {
+    // grid stride loop
+    const diff_t i = IndexMapper::template index<diff_t>();
+
+    // Assign the index to the argument and param
+    data.template assign_offset<ArgumentId>(i);
+    data.template assign_param<ParamId>(i);
+
+    // execute enclosed statements
+    enclosed_stmts_t::exec(data, thread_active);
+  }
+};
+
 /*
  * Executor for work sharing inside CudaKernel.
  * Provides a direct mapping.

include/RAJA/policy/cuda/kernel/Tile.hpp

@@ -42,6 +42,92 @@ namespace RAJA
 namespace internal
 {
 
+/*!
+ * A specialized RAJA::kernel cuda_impl executor for statement::Tile
+ * Assigns the tile segment to segment ArgumentId
+ * Meets all sync requirements
+ */
+template<typename Data,
+         camp::idx_t ArgumentId,
+         camp::idx_t chunk_size,
+         typename IndexMapper,
+         kernel_sync_requirement sync,
+         typename... EnclosedStmts,
+         typename Types>
+struct CudaStatementExecutor<
+    Data,
+    statement::Tile<
+        ArgumentId,
+        RAJA::tile_fixed<chunk_size>,
+        RAJA::policy::cuda::
+            cuda_indexer<iteration_mapping::DirectUnchecked, sync, IndexMapper>,
+        EnclosedStmts...>,
+    Types>
+{
+
+  using stmt_list_t = StatementList<EnclosedStmts...>;
+
+  using enclosed_stmts_t = CudaStatementListExecutor<Data, stmt_list_t, Types>;
+
+  using diff_t = segment_diff_type<ArgumentId, Data>;
+
+  using DimensionCalculator = KernelDimensionCalculator<
+      RAJA::policy::cuda::
+          cuda_indexer<iteration_mapping::DirectUnchecked, sync, IndexMapper>>;
+
+  static inline RAJA_DEVICE void exec(Data& data, bool thread_active)
+  {
+    // Get the segment referenced by this Tile statement
+    auto& segment = camp::get<ArgumentId>(data.segment_tuple);
+
+    using segment_t = camp::decay<decltype(segment)>;
+
+    // compute trip count
+    const diff_t i =
+        IndexMapper::template index<diff_t>() * static_cast<diff_t>(chunk_size);
+
+    // Keep copy of original segment, so we can restore it
+    segment_t orig_segment = segment;
+
+    // Assign our new tiled segment
+    segment = orig_segment.slice(i, static_cast<diff_t>(chunk_size));
+
+    // execute enclosed statements
+    enclosed_stmts_t::exec(data, thread_active);
+
+    // Set range back to original values
+    segment = orig_segment;
+  }
+
+  static inline LaunchDims calculateDimensions(Data const& data)
+  {
+    // Compute how many chunks
+    const diff_t full_len = segment_length<ArgumentId>(data);
+    const diff_t len =
+        RAJA_DIVIDE_CEILING_INT(full_len, static_cast<diff_t>(chunk_size));
+
+    LaunchDims dims = DimensionCalculator::get_dimensions(len);
+
+    // privatize data, so we can mess with the segments
+    using data_t        = camp::decay<Data>;
+    data_t private_data = data;
+
+    // Get original segment
+    auto& segment = camp::get<ArgumentId>(private_data.segment_tuple);
+
+    // restrict to first tile
+    segment = segment.slice(0, static_cast<diff_t>(chunk_size));
+
+    // NOTE: We do not detect improper uses of direct_unchecked policies under
+    // tiling. This happens when using a direct unchecked policy on a tiled
+    // range that is not evenly divisible by chunk_size.
+    LaunchDims enclosed_dims =
+        enclosed_stmts_t::calculateDimensions(private_data);
+
+    return combine(dims, enclosed_dims);
+  }
+};
+
 /*!
  * A specialized RAJA::kernel cuda_impl executor for statement::Tile
  * Assigns the tile segment to segment ArgumentId

include/RAJA/policy/cuda/kernel/TileTCount.hpp

@@ -42,6 +42,82 @@ namespace RAJA
 namespace internal
 {
 
+/*!
+ * A specialized RAJA::kernel cuda_impl executor for statement::TileTCount
+ * Assigns the tile segment to segment ArgumentId
+ * Assigns the tile index to param ParamId
+ * Meets all sync requirements
+ */
+template<typename Data,
+         camp::idx_t ArgumentId,
+         typename ParamId,
+         camp::idx_t chunk_size,
+         typename IndexMapper,
+         kernel_sync_requirement sync,
+         typename... EnclosedStmts,
+         typename Types>
+struct CudaStatementExecutor<
+    Data,
+    statement::TileTCount<
+        ArgumentId,
+        ParamId,
+        RAJA::tile_fixed<chunk_size>,
+        RAJA::policy::cuda::
+            cuda_indexer<iteration_mapping::DirectUnchecked, sync, IndexMapper>,
+        EnclosedStmts...>,
+    Types>
+    : public CudaStatementExecutor<
+          Data,
+          statement::Tile<ArgumentId,
+                          RAJA::tile_fixed<chunk_size>,
+                          RAJA::policy::cuda::cuda_indexer<
+                              iteration_mapping::DirectUnchecked,
+                              sync,
+                              IndexMapper>,
+                          EnclosedStmts...>,
+          Types>
+{
+
+  using Base = CudaStatementExecutor<
+      Data,
+      statement::Tile<
+          ArgumentId,
+          RAJA::tile_fixed<chunk_size>,
+          RAJA::policy::cuda::cuda_indexer<iteration_mapping::DirectUnchecked,
+                                           sync,
+                                           IndexMapper>,
+          EnclosedStmts...>,
+      Types>;
+
+  using typename Base::diff_t;
+  using typename Base::enclosed_stmts_t;
+
+  static inline RAJA_DEVICE void exec(Data& data, bool thread_active)
+  {
+    // Get the segment referenced by this Tile statement
+    auto& segment = camp::get<ArgumentId>(data.segment_tuple);
+
+    using segment_t = camp::decay<decltype(segment)>;
+
+    // compute trip count
+    const diff_t t = IndexMapper::template index<diff_t>();
+    const diff_t i = t * static_cast<diff_t>(chunk_size);
+
+    // Keep copy of original segment, so we can restore it
+    segment_t orig_segment = segment;
+
+    // Assign our new tiled segment
+    segment = orig_segment.slice(i, static_cast<diff_t>(chunk_size));
+    data.template assign_param<ParamId>(t);
+
+    // execute enclosed statements
+    enclosed_stmts_t::exec(data, thread_active);
+
+    // Set range back to original values
+    segment = orig_segment;
+  }
+};
+
 /*!
  * A specialized RAJA::kernel cuda_impl executor for statement::TileTCount
  * Assigns the tile segment to segment ArgumentId