Ensure include order of insert_nested_NVTX_range_guard via clang-format (#4741)

bernhardmgruber · web-flow · commit b6a639076995 · 2025-05-19T09:44:53.000-07:00
diff --git a/.clang-format b/.clang-format
@@ -89,6 +89,9 @@ IncludeCategories:
   - Regex:           '^<cuda/experimental/__execution/prologue.cuh>'
     Priority:            0x7FFFFFFF
     SortPriority:        0x7FFFFFFF
+  - Regex:           '^"insert_nested_NVTX_range_guard.h"'
+    Priority:            -1
+    SortPriority:        -1
   - Regex:           '^<(cuda/std/detail/__config|cub/config.cuh|thrust/detail/config.h|thrust/system/cuda/config.h)'
     Priority:            0
     SortPriority:        0
diff --git a/cub/test/catch2_test_device_adjacent_difference_substract_left.cu b/cub/test/catch2_test_device_adjacent_difference_substract_left.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_adjacent_difference.cuh>
 
diff --git a/cub/test/catch2_test_device_adjacent_difference_substract_right.cu b/cub/test/catch2_test_device_adjacent_difference_substract_right.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_adjacent_difference.cuh>
 
diff --git a/cub/test/catch2_test_device_bulk.cu b/cub/test/catch2_test_device_bulk.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_for.cuh>
 
diff --git a/cub/test/catch2_test_device_decoupled_look_back.cu b/cub/test/catch2_test_device_decoupled_look_back.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #undef NDEBUG
 #include <cub/device/device_scan.cuh>
diff --git a/cub/test/catch2_test_device_for.cu b/cub/test/catch2_test_device_for.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_for.cuh>
 
diff --git a/cub/test/catch2_test_device_for_api.cu b/cub/test/catch2_test_device_for_api.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_for.cuh>
 
diff --git a/cub/test/catch2_test_device_for_copy.cu b/cub/test/catch2_test_device_for_copy.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_for.cuh>
 
diff --git a/cub/test/catch2_test_device_for_each_in_extents_api.cu b/cub/test/catch2_test_device_for_each_in_extents_api.cu
@@ -31,7 +31,6 @@
 #if !_CCCL_COMPILER(MSVC)
 
 #  include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #  include <cub/device/device_for.cuh>
 
diff --git a/cub/test/catch2_test_device_for_utils.cu b/cub/test/catch2_test_device_for_utils.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_for.cuh>
 
diff --git a/cub/test/catch2_test_device_merge.cu b/cub/test/catch2_test_device_merge.cu
@@ -2,7 +2,6 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_merge.cuh>
 
diff --git a/cub/test/catch2_test_device_merge_api.cu b/cub/test/catch2_test_device_merge_api.cu
@@ -2,7 +2,6 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_merge.cuh>
 
diff --git a/cub/test/catch2_test_device_merge_no_unroll.cu b/cub/test/catch2_test_device_merge_no_unroll.cu
@@ -7,7 +7,6 @@
 #define CCCL_AVOID_SORT_UNROLL
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_merge.cuh>
 
diff --git a/cub/test/catch2_test_device_merge_sort.cu b/cub/test/catch2_test_device_merge_sort.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_merge_sort.cuh>
 
diff --git a/cub/test/catch2_test_device_merge_sort_iterators.cu b/cub/test/catch2_test_device_merge_sort_iterators.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_merge_sort.cuh>
 
diff --git a/cub/test/catch2_test_device_merge_sort_vsmem.cu b/cub/test/catch2_test_device_merge_sort_vsmem.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_merge_sort.cuh>
 
diff --git a/cub/test/catch2_test_device_partition_flagged.cu b/cub/test/catch2_test_device_partition_flagged.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_partition.cuh>
 
diff --git a/cub/test/catch2_test_device_partition_if.cu b/cub/test/catch2_test_device_partition_if.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_partition.cuh>
 
diff --git a/cub/test/catch2_test_device_radix_sort_custom.cu b/cub/test/catch2_test_device_radix_sort_custom.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_radix_sort.cuh>
 
diff --git a/cub/test/catch2_test_device_radix_sort_keys.cu b/cub/test/catch2_test_device_radix_sort_keys.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_radix_sort.cuh>
 #include <cub/util_type.cuh>
diff --git a/cub/test/catch2_test_device_radix_sort_pairs.cu b/cub/test/catch2_test_device_radix_sort_pairs.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_radix_sort.cuh>
 #include <cub/util_type.cuh>
diff --git a/cub/test/catch2_test_device_reduce.cu b/cub/test/catch2_test_device_reduce.cu
@@ -25,7 +25,6 @@
  *
  ******************************************************************************/
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_reduce_by_key.cu b/cub/test/catch2_test_device_reduce_by_key.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_reduce_by_key_iterators.cu b/cub/test/catch2_test_device_reduce_by_key_iterators.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_reduce_fp_inf.cu b/cub/test/catch2_test_device_reduce_fp_inf.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_reduce_iterators.cu b/cub/test/catch2_test_device_reduce_iterators.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_reduce_large_offsets.cu b/cub/test/catch2_test_device_reduce_large_offsets.cu
@@ -2,7 +2,6 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_reduce.cuh>
 #include <cub/thread/thread_operators.cuh>
diff --git a/cub/test/catch2_test_device_run_length_encode.cu b/cub/test/catch2_test_device_run_length_encode.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_run_length_encode.cuh>
 
diff --git a/cub/test/catch2_test_device_run_length_encode_non_trivial_runs.cu b/cub/test/catch2_test_device_run_length_encode_non_trivial_runs.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_run_length_encode.cuh>
 
diff --git a/cub/test/catch2_test_device_scan.cu b/cub/test/catch2_test_device_scan.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_scan.cuh>
 
diff --git a/cub/test/catch2_test_device_scan_by_key.cu b/cub/test/catch2_test_device_scan_by_key.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_scan.cuh>
 
diff --git a/cub/test/catch2_test_device_scan_by_key_iterators.cu b/cub/test/catch2_test_device_scan_by_key_iterators.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_scan.cuh>
 
diff --git a/cub/test/catch2_test_device_scan_by_key_large_offsets.cu b/cub/test/catch2_test_device_scan_by_key_large_offsets.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_scan.cuh>
 
diff --git a/cub/test/catch2_test_device_scan_iterators.cu b/cub/test/catch2_test_device_scan_iterators.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_scan.cuh>
 
diff --git a/cub/test/catch2_test_device_scan_large_offsets.cu b/cub/test/catch2_test_device_scan_large_offsets.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_scan.cuh>
 
diff --git a/cub/test/catch2_test_device_segmented_radix_sort_keys.cu b/cub/test/catch2_test_device_segmented_radix_sort_keys.cu
@@ -25,7 +25,6 @@
  *
  ******************************************************************************/
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_segmented_radix_sort.cuh>
 #include <cub/util_type.cuh>
diff --git a/cub/test/catch2_test_device_segmented_radix_sort_pairs.cu b/cub/test/catch2_test_device_segmented_radix_sort_pairs.cu
@@ -25,7 +25,6 @@
  *
  ******************************************************************************/
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_segmented_radix_sort.cuh>
 #include <cub/util_type.cuh>
diff --git a/cub/test/catch2_test_device_segmented_reduce.cu b/cub/test/catch2_test_device_segmented_reduce.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_segmented_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_segmented_reduce_api.cu b/cub/test/catch2_test_device_segmented_reduce_api.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_segmented_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_segmented_reduce_iterators.cu b/cub/test/catch2_test_device_segmented_reduce_iterators.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_segmented_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_segmented_reduce_iterators_64bit.cu b/cub/test/catch2_test_device_segmented_reduce_iterators_64bit.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_segmented_reduce.cuh>
 
diff --git a/cub/test/catch2_test_device_segmented_reduce_large_offsets.cu b/cub/test/catch2_test_device_segmented_reduce_large_offsets.cu
@@ -2,7 +2,6 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_segmented_reduce.cuh>
 #include <cub/thread/thread_operators.cuh>
diff --git a/cub/test/catch2_test_device_segmented_sort_keys.cu b/cub/test/catch2_test_device_segmented_sort_keys.cu
@@ -26,7 +26,7 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
+
 #include <cub/device/device_segmented_sort.cuh>
 #include <cub/util_type.cuh>
 
diff --git a/cub/test/catch2_test_device_segmented_sort_pairs.cu b/cub/test/catch2_test_device_segmented_sort_pairs.cu
@@ -25,8 +25,6 @@
  *
  ******************************************************************************/
 #include "catch2_radix_sort_helper.cuh"
-// above header needs to be included first
-
 #include "catch2_segmented_sort_helper.cuh"
 #include <c2h/catch2_test_helper.h>
 
diff --git a/cub/test/catch2_test_device_select_api.cu b/cub/test/catch2_test_device_select_api.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_select.cuh>
 
diff --git a/cub/test/catch2_test_device_select_flagged.cu b/cub/test/catch2_test_device_select_flagged.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_select.cuh>
 
diff --git a/cub/test/catch2_test_device_select_flagged_if.cu b/cub/test/catch2_test_device_select_flagged_if.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_select.cuh>
 
diff --git a/cub/test/catch2_test_device_select_if.cu b/cub/test/catch2_test_device_select_if.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_select.cuh>
 #include <cub/device/dispatch/dispatch_select_if.cuh>
diff --git a/cub/test/catch2_test_device_select_if_vsmem.cu b/cub/test/catch2_test_device_select_if_vsmem.cu
@@ -26,7 +26,6 @@
  ******************************************************************************/
 
 #include "insert_nested_NVTX_range_guard.h"
-// above header needs to be included first
 
 #include <cub/device/device_select.cuh>
 
diff --git a/cub/test/catch2_test_device_select_unique.cu b/cub/test/catch2_test_device_select_unique.cu
diff --git a/cub/test/catch2_test_device_select_unique_by_key.cu b/cub/test/catch2_test_device_select_unique_by_key.cu
diff --git a/cub/test/catch2_test_device_three_way_partition.cu b/cub/test/catch2_test_device_three_way_partition.cu
diff --git a/cub/test/catch2_test_device_transform.cu b/cub/test/catch2_test_device_transform.cu
diff --git a/cub/test/catch2_test_device_transform_reduce.cu b/cub/test/catch2_test_device_transform_reduce.cu

Original file line number	Diff line number	Diff line change
`@@ -25,7 +25,6 @@`
`25`	`25`	`*`
`26`	`26`	`******************************************************************************/`
`27`	`27`	`#include "insert_nested_NVTX_range_guard.h"`
`28`		`-// above header needs to be included first`
`29`	`28`
`30`	`29`	`#include <cub/device/device_reduce.cuh>`
`31`	`30`
Original file line number	Diff line number	Diff line change
`@@ -25,8 +25,6 @@`
`25`	`25`	`*`
`26`	`26`	`******************************************************************************/`
`27`	`27`	`#include "catch2_radix_sort_helper.cuh"`
`28`		`-// above header needs to be included first`
`29`		`-`
`30`	`28`	`#include "catch2_segmented_sort_helper.cuh"`
`31`	`29`	`#include <c2h/catch2_test_helper.h>`
`32`	`30`