Un-doc functions taking debug_synchronous (#2209)

* undoc functions taking debug_synchronous
NVIDIA · Aug 9, 2024 · 7473934 · 7473934
1 parent 8e20c9a
commit 7473934
Show file tree

Hide file tree

Showing 24 changed files with 244 additions and 34 deletions.
diff --git a/cub/cub/device/device_adjacent_difference.cuh b/cub/cub/device/device_adjacent_difference.cuh
@@ -267,6 +267,7 @@ public:
       d_temp_storage, temp_storage_bytes, d_input, d_output, num_items, difference_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename InputIteratorT, typename OutputIteratorT, typename DifferenceOpT, typename NumItemsT = std::uint32_t>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED static CUB_RUNTIME_FUNCTION cudaError_t SubtractLeftCopy(
     void* d_temp_storage,
@@ -282,6 +283,7 @@ public:
 
     return SubtractLeftCopy(d_temp_storage, temp_storage_bytes, d_input, d_output, num_items, difference_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Subtracts the left element of each adjacent pair of elements residing within device-accessible memory.
@@ -394,6 +396,7 @@ public:
       d_temp_storage, temp_storage_bytes, d_input, d_input, num_items, difference_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename RandomAccessIteratorT, typename DifferenceOpT, typename NumItemsT = std::uint32_t>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED static CUB_RUNTIME_FUNCTION cudaError_t SubtractLeft(
     void* d_temp_storage,
@@ -408,6 +411,7 @@ public:
 
     return SubtractLeft(d_temp_storage, temp_storage_bytes, d_input, num_items, difference_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Subtracts the right element of each adjacent pair of elements residing within device-accessible memory.
@@ -539,6 +543,7 @@ public:
       d_temp_storage, temp_storage_bytes, d_input, d_output, num_items, difference_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename InputIteratorT, typename OutputIteratorT, typename DifferenceOpT, typename NumItemsT = std::uint32_t>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED static CUB_RUNTIME_FUNCTION cudaError_t SubtractRightCopy(
     void* d_temp_storage,
@@ -554,6 +559,7 @@ public:
 
     return SubtractRightCopy(d_temp_storage, temp_storage_bytes, d_input, d_output, num_items, difference_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Subtracts the right element of each adjacent pair of elements residing within device-accessible memory.
@@ -655,6 +661,7 @@ public:
       d_temp_storage, temp_storage_bytes, d_input, d_input, num_items, difference_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename RandomAccessIteratorT, typename DifferenceOpT, typename NumItemsT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED static CUB_RUNTIME_FUNCTION cudaError_t SubtractRight(
     void* d_temp_storage,
@@ -669,6 +676,7 @@ public:
 
     return SubtractRight(d_temp_storage, temp_storage_bytes, d_input, num_items, difference_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 };
 
 CUB_NAMESPACE_END
diff --git a/cub/cub/device/device_histogram.cuh b/cub/cub/device/device_histogram.cuh
@@ -206,6 +206,7 @@ struct DeviceHistogram
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename SampleIteratorT, typename CounterT, typename LevelT, typename OffsetT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t HistogramEven(
     void* d_temp_storage,
@@ -232,6 +233,7 @@ struct DeviceHistogram
       num_samples,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Computes an intensity histogram from a sequence of data samples using equal-width bins.
@@ -384,6 +386,7 @@ struct DeviceHistogram
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename SampleIteratorT, typename CounterT, typename LevelT, typename OffsetT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t HistogramEven(
     void* d_temp_storage,
@@ -414,6 +417,7 @@ struct DeviceHistogram
       row_stride_bytes,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Computes per-channel intensity histograms from a sequence of multi-channel "pixel" data samples using
@@ -584,6 +588,7 @@ struct DeviceHistogram
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <int NUM_CHANNELS,
             int NUM_ACTIVE_CHANNELS,
             typename SampleIteratorT,
@@ -615,6 +620,7 @@ struct DeviceHistogram
       num_pixels,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Computes per-channel intensity histograms from a sequence of
@@ -830,6 +836,7 @@ struct DeviceHistogram
       is_byte_sample);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <int NUM_CHANNELS,
             int NUM_ACTIVE_CHANNELS,
             typename SampleIteratorT,
@@ -865,6 +872,7 @@ struct DeviceHistogram
       row_stride_bytes,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @}  end member group
   //! @name Custom bin ranges
@@ -991,6 +999,7 @@ struct DeviceHistogram
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename SampleIteratorT, typename CounterT, typename LevelT, typename OffsetT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t HistogramRange(
     void* d_temp_storage,
@@ -1008,6 +1017,7 @@ struct DeviceHistogram
     return HistogramRange(
       d_temp_storage, temp_storage_bytes, d_samples, d_histogram, num_levels, d_levels, num_samples, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Computes an intensity histogram from a sequence of data samples using the specified bin boundary levels.
@@ -1147,6 +1157,7 @@ struct DeviceHistogram
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename SampleIteratorT, typename CounterT, typename LevelT, typename OffsetT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t HistogramRange(
     void* d_temp_storage,
@@ -1175,6 +1186,7 @@ struct DeviceHistogram
       row_stride_bytes,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Computes per-channel intensity histograms from a sequence of multi-channel "pixel" data samples
@@ -1334,6 +1346,7 @@ struct DeviceHistogram
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <int NUM_CHANNELS,
             int NUM_ACTIVE_CHANNELS,
             typename SampleIteratorT,
@@ -1356,6 +1369,7 @@ struct DeviceHistogram
     return MultiHistogramRange(
       d_temp_storage, temp_storage_bytes, d_samples, d_histogram, num_levels, d_levels, num_pixels, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Computes per-channel intensity histograms from a sequence of multi-channel "pixel" data samples using
@@ -1560,6 +1574,7 @@ struct DeviceHistogram
       is_byte_sample);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <int NUM_CHANNELS,
             int NUM_ACTIVE_CHANNELS,
             typename SampleIteratorT,
@@ -1593,6 +1608,7 @@ struct DeviceHistogram
       row_stride_bytes,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //@}  end member group
 };

diff --git a/cub/cub/device/device_merge_sort.cuh b/cub/cub/device/device_merge_sort.cuh
@@ -246,6 +246,7 @@ public:
     return SortPairsNoNVTX(d_temp_storage, temp_storage_bytes, d_keys, d_items, num_items, compare_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename KeyIteratorT, typename ValueIteratorT, typename OffsetT, typename CompareOpT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t SortPairs(
     void* d_temp_storage,
@@ -262,6 +263,7 @@ public:
     return SortPairs<KeyIteratorT, ValueIteratorT, OffsetT, CompareOpT>(
       d_temp_storage, temp_storage_bytes, d_keys, d_items, num_items, compare_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   /**
    * @brief Sorts items using a merge sorting method.
@@ -409,6 +411,7 @@ public:
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename KeyInputIteratorT,
             typename ValueInputIteratorT,
             typename KeyIteratorT,
@@ -440,6 +443,7 @@ public:
       compare_op,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
 private:
   // Internal version without NVTX range
@@ -566,6 +570,7 @@ public:
     return SortKeysNoNVTX(d_temp_storage, temp_storage_bytes, d_keys, num_items, compare_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename KeyIteratorT, typename OffsetT, typename CompareOpT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t SortKeys(
     void* d_temp_storage,
@@ -581,6 +586,7 @@ public:
     return SortKeys<KeyIteratorT, OffsetT, CompareOpT>(
       d_temp_storage, temp_storage_bytes, d_keys, num_items, compare_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
 private:
   // Internal version without NVTX range
@@ -723,6 +729,7 @@ public:
       d_temp_storage, temp_storage_bytes, d_input_keys, d_output_keys, num_items, compare_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename KeyInputIteratorT, typename KeyIteratorT, typename OffsetT, typename CompareOpT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t SortKeysCopy(
     void* d_temp_storage,
@@ -739,6 +746,7 @@ public:
     return SortKeysCopy<KeyInputIteratorT, KeyIteratorT, OffsetT, CompareOpT>(
       d_temp_storage, temp_storage_bytes, d_input_keys, d_output_keys, num_items, compare_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   /**
    * @brief Sorts items using a merge sorting method.
@@ -849,6 +857,7 @@ public:
       d_temp_storage, temp_storage_bytes, d_keys, d_items, num_items, compare_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename KeyIteratorT, typename ValueIteratorT, typename OffsetT, typename CompareOpT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t StableSortPairs(
     void* d_temp_storage,
@@ -865,6 +874,7 @@ public:
     return StableSortPairs<KeyIteratorT, ValueIteratorT, OffsetT, CompareOpT>(
       d_temp_storage, temp_storage_bytes, d_keys, d_items, num_items, compare_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   /**
    * @brief Sorts items using a merge sorting method.
@@ -966,6 +976,7 @@ public:
       d_temp_storage, temp_storage_bytes, d_keys, num_items, compare_op, stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename KeyIteratorT, typename OffsetT, typename CompareOpT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION static cudaError_t StableSortKeys(
     void* d_temp_storage,
@@ -981,6 +992,7 @@ public:
     return StableSortKeys<KeyIteratorT, OffsetT, CompareOpT>(
       d_temp_storage, temp_storage_bytes, d_keys, num_items, compare_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   /**
    * @brief Sorts items using a merge sorting method.

diff --git a/cub/cub/device/device_partition.cuh b/cub/cub/device/device_partition.cuh
@@ -206,6 +206,7 @@ struct DevicePartition
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename InputIteratorT, typename FlagIterator, typename OutputIteratorT, typename NumSelectedIteratorT>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION _CCCL_FORCEINLINE static cudaError_t Flagged(
     void* d_temp_storage,
@@ -223,6 +224,7 @@ struct DevicePartition
     return Flagged<InputIteratorT, FlagIterator, OutputIteratorT, NumSelectedIteratorT>(
       d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
   //! @rst
   //! Uses the ``select_op`` functor to split the corresponding items from ``d_in`` into
@@ -367,6 +369,7 @@ struct DevicePartition
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename InputIteratorT, typename OutputIteratorT, typename NumSelectedIteratorT, typename SelectOp>
   CUB_DETAIL_RUNTIME_DEBUG_SYNC_IS_NOT_SUPPORTED CUB_RUNTIME_FUNCTION _CCCL_FORCEINLINE static cudaError_t
   If(void* d_temp_storage,
@@ -384,6 +387,7 @@ struct DevicePartition
     return If<InputIteratorT, OutputIteratorT, NumSelectedIteratorT, SelectOp>(
       d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op, stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 
 private:
   template <bool IS_DESCENDING,
@@ -656,6 +660,7 @@ public:
       stream);
   }
 
+#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document
   template <typename InputIteratorT,
             typename FirstOutputIteratorT,
             typename SecondOutputIteratorT,
@@ -698,6 +703,7 @@ public:
       select_second_part_op,
       stream);
   }
+#endif // DOXYGEN_SHOULD_SKIP_THIS
 };
 
 CUB_NAMESPACE_END