From 6be3e78b3e81175a4f4a4541df8bb0664cf9784c Mon Sep 17 00:00:00 2001
From: Bernhard Manfred Gruber <bernhardmgruber@gmail.com>
Date: Fri, 31 Jan 2025 15:32:02 +0100
Subject: [PATCH] Backport to 2.8: Deprecate and replace
 `CUB_IS_INT128_ENABLED` (#3427) (#3629)

Co-authored-by: Federico Busato <50413820+fbusato@users.noreply.github.com>
---
 cub/cub/detail/fast_modulo_division.cuh       |  6 +++---
 .../device/dispatch/dispatch_histogram.cuh    | 12 +++++------
 .../tuning/tuning_run_length_encode.cuh       |  8 ++++----
 .../device/dispatch/tuning/tuning_scan.cuh    |  4 ++--
 .../dispatch/tuning/tuning_scan_by_key.cuh    | 20 +++++++++----------
 .../dispatch/tuning/tuning_select_if.cuh      | 16 +++++++--------
 cub/cub/util_ptx.cuh                          |  2 +-
 cub/cub/util_type.cuh                         | 13 ++----------
 .../catch2_test_device_for_each_in_extents.cu |  4 ++--
 cub/test/catch2_test_printing.cu              |  2 +-
 cub/test/internal/catch2_test_fast_div_mod.cu |  2 +-
 cub/test/test_util.h                          |  2 +-
 .../cuda/std/__cccl/extended_data_types.h     |  2 +-
 13 files changed, 42 insertions(+), 51 deletions(-)

diff --git a/cub/cub/detail/fast_modulo_division.cuh b/cub/cub/detail/fast_modulo_division.cuh
index 2ae65400e71..14d1fae32a1 100644
--- a/cub/cub/detail/fast_modulo_division.cuh
+++ b/cub/cub/detail/fast_modulo_division.cuh
@@ -38,7 +38,7 @@
 #endif // no system header
 
 #include <cub/detail/type_traits.cuh> // implicit_prom_t
-#include <cub/util_type.cuh> // CUB_IS_INT128_ENABLED
+#include <cub/util_type.cuh> // _CCCL_HAS_INT128()
 
 #include <cuda/cmath> // cuda::std::ceil_div
 #include <cuda/std/bit> // std::has_single_bit
@@ -79,7 +79,7 @@ struct larger_unsigned_type<T, typename ::cuda::std::enable_if<(sizeof(T) == 4)>
   using type = ::cuda::std::uint64_t;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 
 template <typename T>
 struct larger_unsigned_type<T, typename ::cuda::std::enable_if<(sizeof(T) == 8)>::type>
@@ -87,7 +87,7 @@ struct larger_unsigned_type<T, typename ::cuda::std::enable_if<(sizeof(T) == 8)>
   using type = __uint128_t;
 };
 
-#endif // CUB_IS_INT128_ENABLED
+#endif // _CCCL_HAS_INT128()
 
 template <typename T>
 using larger_unsigned_type_t = typename larger_unsigned_type<T>::type;
diff --git a/cub/cub/device/dispatch/dispatch_histogram.cuh b/cub/cub/device/dispatch/dispatch_histogram.cuh
index 32ef0b7e5d9..d4bc8306eca 100644
--- a/cub/cub/device/dispatch/dispatch_histogram.cuh
+++ b/cub/cub/device/dispatch/dispatch_histogram.cuh
@@ -646,27 +646,27 @@ public:
     using IntArithmeticT = ::cuda::std::_If< //
       sizeof(SampleT) + sizeof(CommonT) <= sizeof(uint32_t), //
       uint32_t, //
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
       ::cuda::std::_If< //
         (::cuda::std::is_same<CommonT, __int128_t>::value || //
          ::cuda::std::is_same<CommonT, __uint128_t>::value), //
         CommonT, //
         uint64_t> //
-#else // ^^^ CUB_IS_INT128_ENABLED ^^^ / vvv !CUB_IS_INT128_ENABLED vvv
+#else // ^^^ _CCCL_HAS_INT128() ^^^ / vvv !_CCCL_HAS_INT128() vvv
       uint64_t
-#endif // !CUB_IS_INT128_ENABLED
+#endif // !_CCCL_HAS_INT128()
       >;
 
     // Alias template that excludes __[u]int128 from the integral types
     template <typename T>
     using is_integral_excl_int128 =
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
       ::cuda::std::_If<::cuda::std::is_same<T, __int128_t>::value&& ::cuda::std::is_same<T, __uint128_t>::value,
                        ::cuda::std::false_type,
                        ::cuda::std::is_integral<T>>;
-#else // ^^^ CUB_IS_INT128_ENABLED ^^^ / vvv !CUB_IS_INT128_ENABLED vvv
+#else // ^^^ _CCCL_HAS_INT128() ^^^ / vvv !_CCCL_HAS_INT128() vvv
       ::cuda::std::is_integral<T>;
-#endif // !CUB_IS_INT128_ENABLED
+#endif // !_CCCL_HAS_INT128()
 
     union ScaleT
     {
diff --git a/cub/cub/device/dispatch/tuning/tuning_run_length_encode.cuh b/cub/cub/device/dispatch/tuning/tuning_run_length_encode.cuh
index 87631d1199e..a2d62e6ab2b 100644
--- a/cub/cub/device/dispatch/tuning/tuning_run_length_encode.cuh
+++ b/cub/cub/device/dispatch/tuning/tuning_run_length_encode.cuh
@@ -156,7 +156,7 @@ struct sm80_tuning<LengthT, KeyT, primitive_length::yes, primitive_key::yes, len
   using delay_constructor                            = detail::no_delay_constructor_t<1075>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class LengthT>
 struct sm80_tuning<LengthT, __int128_t, primitive_length::yes, primitive_key::no, length_size::_4, key_size::_16>
 {
@@ -216,7 +216,7 @@ struct sm90_tuning<LengthT, KeyT, primitive_length::yes, primitive_key::yes, len
   using delay_constructor                            = detail::no_delay_constructor_t<515>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class LengthT>
 struct sm90_tuning<LengthT, __int128_t, primitive_length::yes, primitive_key::no, length_size::_4, key_size::_16>
 {
@@ -349,7 +349,7 @@ struct sm80_tuning<LengthT, KeyT, primitive_length::yes, primitive_key::yes, len
   using delay_constructor                            = detail::no_delay_constructor_t<1065>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class LengthT>
 struct sm80_tuning<LengthT, __int128_t, primitive_length::yes, primitive_key::no, length_size::_4, key_size::_16>
 {
@@ -414,7 +414,7 @@ struct sm90_tuning<LengthT, KeyT, primitive_length::yes, primitive_key::yes, len
   using delay_constructor                            = detail::no_delay_constructor_t<840>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class LengthT>
 struct sm90_tuning<LengthT, __int128_t, primitive_length::yes, primitive_key::no, length_size::_4, key_size::_16>
 {
diff --git a/cub/cub/device/dispatch/tuning/tuning_scan.cuh b/cub/cub/device/dispatch/tuning/tuning_scan.cuh
index 1c76064da1d..1a0fac5db26 100644
--- a/cub/cub/device/dispatch/tuning/tuning_scan.cuh
+++ b/cub/cub/device/dispatch/tuning/tuning_scan.cuh
@@ -186,7 +186,7 @@ struct sm80_tuning<double, primitive_op::yes, primitive_accum::yes, accum_size::
   static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm80_tuning<__int128_t, primitive_op::yes, primitive_accum::no, accum_size::_16>
 {
@@ -218,7 +218,7 @@ template <class T> struct sm90_tuning<T, primitive_op::yes, primitive_accum::yes
 template <> struct sm90_tuning<float,  primitive_op::yes, primitive_accum::yes, accum_size::_4> : tuning<float,  128, 24, 688, 1140> {};
 template <> struct sm90_tuning<double, primitive_op::yes, primitive_accum::yes, accum_size::_8> : tuning<double, 224, 24, 576, 1215> {};
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <> struct sm90_tuning<__int128_t, primitive_op::yes, primitive_accum::no, accum_size::_16> : tuning<__int128_t, 576, 21, 860, 630> {};
 template <>
 struct sm90_tuning<__uint128_t, primitive_op::yes, primitive_accum::no, accum_size::_16>
diff --git a/cub/cub/device/dispatch/tuning/tuning_scan_by_key.cuh b/cub/cub/device/dispatch/tuning/tuning_scan_by_key.cuh
index cdd2468dc38..419aa059c08 100644
--- a/cub/cub/device/dispatch/tuning/tuning_scan_by_key.cuh
+++ b/cub/cub/device/dispatch/tuning/tuning_scan_by_key.cuh
@@ -172,7 +172,7 @@ struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<124, 1040>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
 {
@@ -229,7 +229,7 @@ struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<160, 695>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
 {
@@ -286,7 +286,7 @@ struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<888, 635>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
 {
@@ -343,7 +343,7 @@ struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_8,
   using delay_constructor                              = no_delay_constructor_t<1160>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
 {
@@ -400,7 +400,7 @@ struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_8,
   using delay_constructor                              = no_delay_constructor_t<1030>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
 {
@@ -465,7 +465,7 @@ struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<488, 1070>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
 {
@@ -522,7 +522,7 @@ struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<352, 1170>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
 {
@@ -579,7 +579,7 @@ struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<556, 1195>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
 {
@@ -636,7 +636,7 @@ struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<600, 930>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
 {
@@ -693,7 +693,7 @@ struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_8,
   using delay_constructor                              = fixed_delay_constructor_t<320, 1200>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <class KeyT>
 struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
 {
diff --git a/cub/cub/device/dispatch/tuning/tuning_select_if.cuh b/cub/cub/device/dispatch/tuning/tuning_select_if.cuh
index 792b1669fa1..6093d187732 100644
--- a/cub/cub/device/dispatch/tuning/tuning_select_if.cuh
+++ b/cub/cub/device/dispatch/tuning/tuning_select_if.cuh
@@ -121,7 +121,7 @@ struct sm80_tuning<Input, flagged::no, keep_rejects::no, offset_size::_4, primit
   using delay_constructor                            = detail::fixed_delay_constructor_t<832, 1165>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm80_tuning<__int128_t, flagged::no, keep_rejects::no, offset_size::_4, primitive::no, input_size::_16>
 {
@@ -174,7 +174,7 @@ struct sm80_tuning<Input, flagged::yes, keep_rejects::no, offset_size::_4, primi
   using delay_constructor                            = detail::no_delay_constructor_t<1130>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm80_tuning<__int128_t, flagged::yes, keep_rejects::no, offset_size::_4, primitive::no, input_size::_16>
 {
@@ -227,7 +227,7 @@ struct sm80_tuning<Input, flagged::no, keep_rejects::yes, offset_size::_4, primi
   using delay_constructor                            = detail::fixed_delay_constructor_t<68, 1160>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm80_tuning<__int128_t, flagged::no, keep_rejects::yes, offset_size::_4, primitive::no, input_size::_16>
 {
@@ -280,7 +280,7 @@ struct sm80_tuning<Input, flagged::yes, keep_rejects::yes, offset_size::_4, prim
   using delay_constructor                            = detail::fixed_delay_constructor_t<884, 1130>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm80_tuning<__int128_t, flagged::yes, keep_rejects::yes, offset_size::_4, primitive::no, input_size::_16>
 {
@@ -336,7 +336,7 @@ struct sm90_tuning<Input, flagged::no, keep_rejects::no, offset_size::_4, primit
   using delay_constructor                            = detail::fixed_delay_constructor_t<380, 1140>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm90_tuning<__int128_t, flagged::no, keep_rejects::no, offset_size::_4, primitive::no, input_size::_16>
 {
@@ -389,7 +389,7 @@ struct sm90_tuning<Input, flagged::yes, keep_rejects::no, offset_size::_4, primi
   using delay_constructor                            = detail::fixed_delay_constructor_t<360, 1170>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm90_tuning<__int128_t, flagged::yes, keep_rejects::no, offset_size::_4, primitive::no, input_size::_16>
 {
@@ -442,7 +442,7 @@ struct sm90_tuning<Input, flagged::no, keep_rejects::yes, offset_size::_4, primi
   using delay_constructor                            = detail::fixed_delay_constructor_t<512, 1075>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm90_tuning<__int128_t, flagged::no, keep_rejects::yes, offset_size::_4, primitive::no, input_size::_16>
 {
@@ -495,7 +495,7 @@ struct sm90_tuning<Input, flagged::yes, keep_rejects::yes, offset_size::_4, prim
   using delay_constructor                            = detail::fixed_delay_constructor_t<532, 1180>;
 };
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 template <>
 struct sm90_tuning<__int128_t, flagged::yes, keep_rejects::yes, offset_size::_4, primitive::no, input_size::_16>
 {
diff --git a/cub/cub/util_ptx.cuh b/cub/cub/util_ptx.cuh
index 99beeed313e..e6bb45c4a31 100644
--- a/cub/cub/util_ptx.cuh
+++ b/cub/cub/util_ptx.cuh
@@ -99,7 +99,7 @@ BFE(UnsignedBits source, unsigned int bit_start, unsigned int num_bits, Int2Type
   return (source >> bit_start) & MASK;
 }
 
-#  if CUB_IS_INT128_ENABLED
+#  if _CCCL_HAS_INT128()
 /**
  * Bitfield-extract for 128-bit types.
  */
diff --git a/cub/cub/util_type.cuh b/cub/cub/util_type.cuh
index c6be2a5209f..a89cd159309 100644
--- a/cub/cub/util_type.cuh
+++ b/cub/cub/util_type.cuh
@@ -76,17 +76,8 @@ _CCCL_DIAG_POP
 CUB_NAMESPACE_BEGIN
 
 #ifndef CUB_IS_INT128_ENABLED
-#  if defined(__CUDACC_RTC__)
-#    if defined(__CUDACC_RTC_INT128__)
-#      define CUB_IS_INT128_ENABLED 1
-#    endif // !defined(__CUDACC_RTC_INT128__)
-#  else // !defined(__CUDACC_RTC__)
-#    if _CCCL_CUDACC_AT_LEAST(11, 5)
-#      if _CCCL_COMPILER(GCC) || _CCCL_COMPILER(CLANG) || _CCCL_COMPILER(ICC) || _CCCL_COMPILER(NVHPC)
-#        define CUB_IS_INT128_ENABLED 1
-#      endif // GCC || CLANG || ICC || NVHPC
-#    endif // _CCCL_CUDACC_AT_LEAST(11, 5)
-#  endif // !defined(__CUDACC_RTC__)
+// Deprecated [Since 2.8]
+#  define CUB_IS_INT128_ENABLED _CCCL_HAS_INT128()
 #endif // !defined(CUB_IS_INT128_ENABLED)
 
 /******************************************************************************
diff --git a/cub/test/catch2_test_device_for_each_in_extents.cu b/cub/test/catch2_test_device_for_each_in_extents.cu
index 6f11810101c..2d87fdfb193 100644
--- a/cub/test/catch2_test_device_for_each_in_extents.cu
+++ b/cub/test/catch2_test_device_for_each_in_extents.cu
@@ -107,7 +107,7 @@ using index_types =
                  uint16_t,
                  int32_t,
                  uint32_t
-#  if CUB_IS_INT128_ENABLED
+#  if _CCCL_HAS_INT128()
                  ,
                  int64_t,
                  uint64_t
@@ -120,7 +120,7 @@ using index_types_dynamic =
                  uint16_t,
                  int32_t,
                  uint32_t
-#  if CUB_IS_INT128_ENABLED
+#  if _CCCL_HAS_INT128()
                  ,
                  int64_t,
                  uint64_t
diff --git a/cub/test/catch2_test_printing.cu b/cub/test/catch2_test_printing.cu
index 6f93515114a..63b622f3554 100644
--- a/cub/test/catch2_test_printing.cu
+++ b/cub/test/catch2_test_printing.cu
@@ -11,7 +11,7 @@ std::string print(T val)
   return ss.str();
 }
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 TEST_CASE("Test utils can print __int128", "[test][utils]")
 {
   REQUIRE(print(__int128_t{0}) == "0");
diff --git a/cub/test/internal/catch2_test_fast_div_mod.cu b/cub/test/internal/catch2_test_fast_div_mod.cu
index 8a1a3e96a27..ec3b5e20d68 100644
--- a/cub/test/internal/catch2_test_fast_div_mod.cu
+++ b/cub/test/internal/catch2_test_fast_div_mod.cu
@@ -42,7 +42,7 @@ using index_types =
                  uint16_t,
                  int32_t,
                  uint32_t
-#  if CUB_IS_INT128_ENABLED
+#  if _CCCL_HAS_INT128()
                  ,
                  int64_t,
                  uint64_t
diff --git a/cub/test/test_util.h b/cub/test/test_util.h
index c06d803ecb1..ac472cba693 100644
--- a/cub/test/test_util.h
+++ b/cub/test/test_util.h
@@ -716,7 +716,7 @@ std::ostream& operator<<(std::ostream& os, const CUB_NS_QUALIFIER::KeyValuePair<
   return os;
 }
 
-#if CUB_IS_INT128_ENABLED
+#if _CCCL_HAS_INT128()
 inline std::ostream& operator<<(std::ostream& os, __uint128_t val)
 {
   constexpr int max_digits      = 40;
diff --git a/libcudacxx/include/cuda/std/__cccl/extended_data_types.h b/libcudacxx/include/cuda/std/__cccl/extended_data_types.h
index 034ebc51b5f..7f160656e5b 100644
--- a/libcudacxx/include/cuda/std/__cccl/extended_data_types.h
+++ b/libcudacxx/include/cuda/std/__cccl/extended_data_types.h
@@ -29,7 +29,7 @@
 #if !defined(_CCCL_DISABLE_INT128)
 #  if _CCCL_COMPILER(NVRTC) && defined(__CUDACC_RTC_INT128__) && _CCCL_OS(LINUX)
 #    define _CCCL_HAS_INT128() 1
-#  elif defined(__SIZEOF_INT128__) && _CCCL_OS(LINUX)
+#  elif defined(__SIZEOF_INT128__) && _CCCL_OS(LINUX) && _CCCL_CUDACC_AT_LEAST(11, 5)
 #    define _CCCL_HAS_INT128() 1
 #  else
 #    define _CCCL_HAS_INT128() 0