From 2c9afd2e5b798005d58009522e393e0365f0bc60 Mon Sep 17 00:00:00 2001
From: codingwithmagga <codingwithmagga@gmail.com>
Date: Fri, 19 Jun 2026 16:02:45 +0200
Subject: [PATCH 01/11] Rework constant_sequence taking a single auto NTTP to a
 variadic NTTP

---
 libcudacxx/include/cuda/__argument/argument.h | 163 +++++++++++++-----
 .../cuda/argument/argument_traits.pass.cpp    |  54 +++---
 .../cuda/argument/static_argument.pass.cpp    |  60 ++++---
 .../cuda/argument/usage_example.pass.cpp      |  28 ++-
 4 files changed, 204 insertions(+), 101 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index 0877a8df6d1..a11f8a91179 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -40,6 +40,7 @@
 #include <cuda/std/__utility/declval.h>
 #include <cuda/std/__utility/forward.h>
 #include <cuda/std/__utility/move.h>
+#include <cuda/std/array>
 #include <cuda/std/cstddef>
 #include <cuda/std/limits>
 
@@ -111,16 +112,79 @@ class constant
 };
 
 //! @brief Wraps a compile-time constant argument sequence.
-template <auto _Value>
+template <class T, T... Vs>
 class __constant_sequence
 {
 public:
-  using value_type     = ::cuda::std::remove_cvref_t<decltype(_Value)>;
-  using __element_type = __element_type_of_t<value_type>;
-
-  static_assert(__is_sequence_v<value_type>, "The value type of __constant_sequence must be a sequence");
+  using __element_type                      = ::cuda::std::remove_cvref_t<T>;
+  using value_type                          = ::cuda::std::array<__element_type, sizeof...(Vs)>;
+  static constexpr ::cuda::std::size_t size = sizeof...(Vs);
 };
 
+template <class T>
+struct __is_builtin_array : ::cuda::std::false_type
+{};
+
+template <class T, ::cuda::std::size_t N>
+struct __is_builtin_array<T[N]> : ::cuda::std::true_type
+{};
+
+template <class T>
+struct __is_cuda_array : ::cuda::std::false_type
+{};
+
+template <class T, ::cuda::std::size_t N>
+struct __is_cuda_array<::cuda::std::array<T, N>> : ::cuda::std::true_type
+{};
+
+template <class T>
+struct __array_extent;
+
+template <class T, ::cuda::std::size_t N>
+struct __array_extent<T[N]> : ::cuda::std::integral_constant<::cuda::std::size_t, N>{};
+
+template <class T, ::cuda::std::size_t N>
+struct __array_extent<::cuda::std::array<T, N>> : ::cuda::std::integral_constant<::cuda::std::size_t, N>{};
+
+template <class>
+inline constexpr bool __always_false_v = false;
+
+template <const auto& Arr, ::cuda::std::size_t... Is>
+constexpr auto make_constant_sequence_impl(::cuda::std::index_sequence<Is...>)
+{
+  using raw_array = ::cuda::std::remove_const_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
+
+  if constexpr (__is_builtin_array<raw_array>::value)
+  {
+    using T = ::cuda::std::remove_cv_t<::cuda::std::remove_extent_t<raw_array>>;
+    return __constant_sequence<T, Arr[Is]...>{};
+  }
+  else if constexpr (__is_cuda_array<raw_array>::value)
+  {
+    using T = typename raw_array::value_type;
+    return __constant_sequence<T, Arr[Is]...>{};
+  }
+  else
+  {
+    static_assert(__always_false_v<raw_array>, "unsupported array type");
+  }
+}
+
+//! @brief Makes a compile-time constant argument sequence.
+//! In C++17, Arr must have static storage duration.
+template <const auto& Arr>
+constexpr auto make_constant_sequence()
+{
+  using raw_array = ::cuda::std::remove_cv_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
+
+  static_assert(__is_builtin_array<raw_array>::value || __is_cuda_array<raw_array>::value,
+                "make_constant_sequence requires a cuda::std::array or non-empty C-style array");
+
+  constexpr ::cuda::std::size_t N = __array_extent<raw_array>::value;
+
+  return make_constant_sequence_impl<Arr>(::cuda::std::make_index_sequence<N>{});
+}
+
 // __assert_in_range
 // =====================================================================
 
@@ -621,8 +685,8 @@ template <class _Arg, class _StaticBounds>
 inline constexpr bool __is_wrapper_v<immediate<_Arg, _StaticBounds>> = true;
 template <auto _Value, class _Tp>
 inline constexpr bool __is_wrapper_v<constant<_Value, _Tp>> = true;
-template <auto _Value>
-inline constexpr bool __is_wrapper_v<__constant_sequence<_Value>> = true;
+template <class T, T... Vs>
+inline constexpr bool __is_wrapper_v<__constant_sequence<T, Vs...>> = true;
 template <class _Arg, class _StaticBounds>
 inline constexpr bool __is_wrapper_v<__immediate_sequence<_Arg, _StaticBounds>> = true;
 template <class _Arg, class _StaticBounds>
@@ -662,11 +726,11 @@ __unwrap(const constant<_Value, _Tp>&) noexcept
   return constant<_Value, _Tp>::__get_value();
 }
 
-template <auto _Value>
-[[nodiscard]] _CCCL_API constexpr ::cuda::std::remove_cvref_t<decltype(_Value)>
-__unwrap(const __constant_sequence<_Value>&) noexcept
+//! Unwraps a compile-time constant argument sequence into a canonical cuda::std::array value.
+template <class T, T... Vs>
+[[nodiscard]] _CCCL_API constexpr auto __unwrap(const __constant_sequence<T, Vs...>&) noexcept
 {
-  return _Value;
+  return ::cuda::std::array<T, sizeof...(Vs)>{Vs...};
 }
 
 template <class _Arg, class _StaticBounds>
@@ -735,32 +799,48 @@ _CCCL_API constexpr auto __constant_compute_highest() noexcept
   return constant<_Value, _Tp>::__get_value();
 }
 
-template <auto _Value>
-_CCCL_API constexpr auto __constant_sequence_compute_lowest() noexcept
+template <class T, T... Vs>
+_CCCL_API constexpr T __constant_sequence_compute_lowest() noexcept
 {
-  using _ElementType = __element_type_of_t<::cuda::std::remove_cvref_t<decltype(_Value)>>;
-  auto __first       = _Value.begin();
-  auto __last        = _Value.end();
-
-  if (__first == __last)
+  if constexpr (sizeof...(Vs) == 0)
   {
-    return ::cuda::std::numeric_limits<_ElementType>::lowest();
+    return ::cuda::std::numeric_limits<T>::lowest();
+  }
+  else
+  {
+    constexpr T values[] = {Vs...};
+    T __min              = values[0];
+    for (T __v : values)
+    {
+      if (__v < __min)
+      {
+        __min = __v;
+      }
+    }
+    return __min;
   }
-  return static_cast<_ElementType>(*::cuda::std::min_element(__first, __last));
 }
 
-template <auto _Value>
-_CCCL_API constexpr auto __constant_sequence_compute_highest() noexcept
+template <class T, T... Vs>
+_CCCL_API constexpr T __constant_sequence_compute_highest() noexcept
 {
-  using _ElementType = __element_type_of_t<::cuda::std::remove_cvref_t<decltype(_Value)>>;
-  auto __first       = _Value.begin();
-  auto __last        = _Value.end();
-
-  if (__first == __last)
+  if constexpr (sizeof...(Vs) == 0)
   {
-    return (::cuda::std::numeric_limits<_ElementType>::max)();
+    return ::cuda::std::numeric_limits<T>::max();
+  }
+  else
+  {
+    constexpr T values[] = {Vs...};
+    T __max              = values[0];
+    for (T __v : values)
+    {
+      if (__v > __max)
+      {
+        __max = __v;
+      }
+    }
+    return __max;
   }
-  return static_cast<_ElementType>(*::cuda::std::max_element(__first, __last));
 }
 
 // =====================================================================
@@ -811,17 +891,16 @@ struct __traits_impl<immediate<_Arg, _StaticBounds>>
   static constexpr element_type highest = __wrapper_static_highest<element_type, _StaticBounds>();
 };
 
-template <auto _Value>
-struct __traits_impl<__constant_sequence<_Value>>
+template <class T, T... Vs>
+struct __traits_impl<__constant_sequence<T, Vs...>>
 {
-  using value_type   = ::cuda::std::remove_cvref_t<decltype(_Value)>;
-  using element_type = __element_type_of_t<value_type>;
-  static_assert(__is_sequence_v<value_type>, "The value type of __constant_sequence must be a sequence");
+  using element_type                    = ::cuda::std::remove_cvref_t<T>;
+  using value_type                      = ::cuda::std::array<element_type, sizeof...(Vs)>;
   static constexpr bool is_constant     = true;
   static constexpr bool is_deferred     = false;
   static constexpr bool is_single_value = false;
-  static constexpr element_type lowest  = __constant_sequence_compute_lowest<_Value>();
-  static constexpr element_type highest = __constant_sequence_compute_highest<_Value>();
+  static constexpr element_type lowest  = __constant_sequence_compute_lowest<T, Vs...>();
+  static constexpr element_type highest = __constant_sequence_compute_highest<T, Vs...>();
 };
 
 template <class _Arg, class _StaticBounds>
@@ -896,10 +975,10 @@ template <auto _Value, class _Tp>
   return __constant_compute_lowest<_Value, _Tp>();
 }
 
-template <auto _Value>
-[[nodiscard]] _CCCL_API constexpr auto __lowest_(__constant_sequence<_Value>) noexcept
+template <class T, T... Vs>
+[[nodiscard]] _CCCL_API constexpr auto __lowest_(__constant_sequence<T, Vs...>) noexcept
 {
-  return __constant_sequence_compute_lowest<_Value>();
+  return __constant_sequence_compute_lowest<T, Vs...>();
 }
 
 template <class _Arg, class _StaticBounds>
@@ -949,10 +1028,10 @@ template <auto _Value, class _Tp>
   return __constant_compute_highest<_Value, _Tp>();
 }
 
-template <auto _Value>
-[[nodiscard]] _CCCL_API constexpr auto __highest_(__constant_sequence<_Value>) noexcept
+template <class T, T... Vs>
+[[nodiscard]] _CCCL_API constexpr auto __highest_(__constant_sequence<T, Vs...>) noexcept
 {
-  return __constant_sequence_compute_highest<_Value>();
+  return __constant_sequence_compute_highest<T, Vs...>();
 }
 
 template <class _Arg, class _StaticBounds>
diff --git a/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp b/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp
index aaf57291e23..89b4494a8df 100644
--- a/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp
@@ -103,9 +103,13 @@ TEST_FUNC void test()
   static_assert(!cuda::args::__traits<cuda::args::immediate<int>>::is_deferred);
   static_assert(!cuda::args::__traits<cuda::args::__immediate_sequence<cuda::std::span<int>>>::is_deferred);
   static_assert(!cuda::args::__traits<cuda::args::constant<42>>::is_deferred);
-#if TEST_HAS_CLASS_NTTP
-  static_assert(!cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{1, 2, 3}>>::is_deferred);
-#endif // TEST_HAS_CLASS_NTTP
+
+  static_assert(!cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::is_deferred);
+  static constexpr int carr[]                         = {1, 2, 3};
+  static constexpr ::cuda::std::array<int, 3> cudaarr = {1, 2, 3};
+  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::is_deferred);
+  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::is_deferred);
+
   static_assert(cuda::args::__traits<cuda::args::deferred<cuda::std::span<int, 1>>>::is_deferred);
   static_assert(cuda::args::__traits<cuda::args::deferred_sequence<cuda::std::span<int>>>::is_deferred);
 
@@ -118,10 +122,11 @@ TEST_FUNC void test()
   static_assert(cuda::args::__traits<cuda::args::immediate<cuda::counting_iterator<int>>>::is_single_value);
   static_assert(!cuda::args::__traits<cuda::args::__immediate_sequence<cuda::std::span<int>>>::is_single_value);
   static_assert(cuda::args::__traits<cuda::args::constant<42>>::is_single_value);
-#if TEST_HAS_CLASS_NTTP
-  static_assert(
-    !cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{1, 2, 3}>>::is_single_value);
-#endif // TEST_HAS_CLASS_NTTP
+
+  static_assert(!cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::is_single_value);
+  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::is_single_value);
+  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::is_single_value);
+
   static_assert(cuda::args::__traits<cuda::args::deferred<int*>>::is_single_value);
   static_assert(!cuda::args::__traits<cuda::args::deferred_sequence<cuda::std::span<int>>>::is_single_value);
 
@@ -134,11 +139,15 @@ TEST_FUNC void test()
                          cuda::std::span<int>>);
   static_assert(cuda::std::is_same_v<cuda::args::__traits<cuda::args::constant<42>>::value_type, int>);
   static_assert(cuda::std::is_same_v<cuda::args::__traits<cuda::args::constant<10, float>>::value_type, float>);
-#if TEST_HAS_CLASS_NTTP
-  static_assert(cuda::std::is_same_v<
-                cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{1, 2, 3}>>::value_type,
-                cuda::std::array<int, 3>>);
-#endif // TEST_HAS_CLASS_NTTP
+
+  static_assert(cuda::std::is_same_v<cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::value_type,
+                                     cuda::std::array<int, 3>>);
+  static_assert(
+    cuda::std::is_same_v<cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::value_type,
+                         cuda::std::array<int, 3>>);
+  static_assert(
+    cuda::std::is_same_v<cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::value_type,
+                         cuda::std::array<int, 3>>);
 
   // --- argument_traits: lowest / highest ---
 
@@ -155,10 +164,13 @@ TEST_FUNC void test()
     == 8);
   static_assert(cuda::args::__traits<cuda::args::constant<10, float>>::lowest == 10.0f);
   static_assert(cuda::args::__traits<cuda::args::constant<10, float>>::highest == 10.0f);
-#if TEST_HAS_CLASS_NTTP
-  static_assert(cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{3, 1, 2}>>::lowest == 1);
-  static_assert(cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{3, 1, 2}>>::highest == 3);
-#endif // TEST_HAS_CLASS_NTTP
+
+  static_assert(cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::lowest == 1);
+  static_assert(cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::highest == 3);
+  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::lowest == 1);
+  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::highest == 3);
+  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::lowest == 1);
+  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::highest == 3);
 
   // --- Free function bounds on plain values ---
 
@@ -166,16 +178,6 @@ TEST_FUNC void test()
   static_assert(cuda::args::__highest_(42) == (cuda::std::numeric_limits<int>::max)());
   static_assert(cuda::args::__lowest_(1.0f) == cuda::std::numeric_limits<float>::lowest());
   static_assert(cuda::args::__highest_(1.0f) == (cuda::std::numeric_limits<float>::max)());
-
-  // --- Scalar and sequence wrappers expose distinct single-value traits ---
-
-  static_assert(cuda::args::__traits<cuda::args::constant<42>>::is_single_value);
-  static_assert(cuda::args::__traits<cuda::args::immediate<int>>::is_single_value);
-  static_assert(!cuda::args::__traits<cuda::args::__immediate_sequence<cuda::std::span<int>>>::is_single_value);
-#if TEST_HAS_CLASS_NTTP
-  static_assert(
-    !cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{1, 2, 3}>>::is_single_value);
-#endif // TEST_HAS_CLASS_NTTP
 }
 
 int main(int, char**)
diff --git a/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp b/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp
index f3cc7a2a993..3b85a8db9a8 100644
--- a/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp
@@ -55,16 +55,24 @@ TEST_FUNC void test()
   }
 #endif // TEST_HAS_CLASS_NTTP
 
-#if TEST_HAS_CLASS_NTTP
   // Array sequence
   {
-    constexpr auto sa_arr = cuda::args::__constant_sequence<cuda::std::array<int, 3>{128, 256, 512}>{};
+    static constexpr int carr[] = {128, 256, 512};
+    constexpr auto sa_arr       = cuda::args::make_constant_sequence<carr>();
+    static_assert(cuda::args::__unwrap(sa_arr)[0] == 128);
+    static_assert(cuda::args::__unwrap(sa_arr)[1] == 256);
+    static_assert(cuda::args::__unwrap(sa_arr)[2] == 512);
+    static_assert(cuda::std::is_same_v<decltype(sa_arr)::value_type, cuda::std::array<int, 3>>);
+  }
+
+  {
+    static constexpr ::cuda::std::array<int, 3> cudaarr = {128, 256, 512};
+    constexpr auto sa_arr                               = cuda::args::make_constant_sequence<cudaarr>();
     static_assert(cuda::args::__unwrap(sa_arr)[0] == 128);
     static_assert(cuda::args::__unwrap(sa_arr)[1] == 256);
     static_assert(cuda::args::__unwrap(sa_arr)[2] == 512);
     static_assert(cuda::std::is_same_v<decltype(sa_arr)::value_type, cuda::std::array<int, 3>>);
   }
-#endif // TEST_HAS_CLASS_NTTP
 
   // Bounds: scalar
   {
@@ -73,23 +81,27 @@ TEST_FUNC void test()
     static_assert(cuda::args::__highest_(sa) == 42);
   }
 
-#if TEST_HAS_CLASS_NTTP
   // Bounds: array sequence computes lowest/highest of elements
   {
-    constexpr auto sa = cuda::args::__constant_sequence<cuda::std::array<int, 3>{128, 256, 512}>{};
+    static constexpr int carr[] = {128, 256, 512};
+    constexpr auto sa           = cuda::args::make_constant_sequence<carr>();
+    static_assert(cuda::args::__lowest_(sa) == 128);
+    static_assert(cuda::args::__highest_(sa) == 512);
+  }
+  {
+    static constexpr ::cuda::std::array<int, 3> cudaarr = {128, 256, 512};
+    constexpr auto sa                                   = cuda::args::make_constant_sequence<cudaarr>();
     static_assert(cuda::args::__lowest_(sa) == 128);
     static_assert(cuda::args::__highest_(sa) == 512);
   }
-#endif // TEST_HAS_CLASS_NTTP
 
-#if TEST_HAS_CLASS_NTTP
   // Bounds: empty array sequence has unconstrained element bounds
   {
-    constexpr auto sa = cuda::args::__constant_sequence<cuda::std::array<int, 0>{}>{};
+    static constexpr ::cuda::std::array<int, 0> cudaarr = {};
+    constexpr auto sa                                   = cuda::args::make_constant_sequence<cudaarr>();
     static_assert(cuda::args::__lowest_(sa) == cuda::std::numeric_limits<int>::lowest());
     static_assert(cuda::args::__highest_(sa) == (cuda::std::numeric_limits<int>::max)());
   }
-#endif // TEST_HAS_CLASS_NTTP
 
   // Traits
   {
@@ -114,25 +126,29 @@ TEST_FUNC void test()
     static_assert(traits::highest == 10.0f);
   }
 
-#if TEST_HAS_CLASS_NTTP
   // Sequence traits
   {
-    using traits = cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{1, 2, 3}>>;
+    static constexpr int carr[] = {1, 2, 3};
+    using traits                = cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>;
+    static_assert(traits::is_constant);
+    static_assert(!traits::is_deferred);
+    static_assert(!traits::is_single_value);
+    static_assert(cuda::std::is_same_v<traits::value_type, cuda::std::array<int, 3>>);
+    static_assert(cuda::std::is_same_v<traits::element_type, int>);
+  }
+  {
+    static constexpr ::cuda::std::array<int, 3> cudaarr = {1, 2, 3};
+    using traits = cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>;
     static_assert(traits::is_constant);
     static_assert(!traits::is_deferred);
     static_assert(!traits::is_single_value);
     static_assert(cuda::std::is_same_v<traits::value_type, cuda::std::array<int, 3>>);
     static_assert(cuda::std::is_same_v<traits::element_type, int>);
   }
-#endif // TEST_HAS_CLASS_NTTP
 
-  // Single value: scalar is single, sequence is not
+  // Single value: scalar is single
   {
     static_assert(!cuda::args::__is_sequence_v<cuda::args::__traits<cuda::args::constant<42>>::value_type>);
-#if TEST_HAS_CLASS_NTTP
-    static_assert(
-      !cuda::args::__traits<cuda::args::__constant_sequence<cuda::std::array<int, 3>{1, 2, 3}>>::is_single_value);
-#endif // TEST_HAS_CLASS_NTTP
   }
 
   // Unwrap: scalar
@@ -149,16 +165,6 @@ TEST_FUNC void test()
     static_assert(val == 10.0f);
     static_assert(cuda::std::is_same_v<decltype(val), const float>);
   }
-
-#if TEST_HAS_CLASS_NTTP
-  // Unwrap: sequence
-  {
-    constexpr auto sa  = cuda::args::__constant_sequence<cuda::std::array<int, 3>{10, 20, 30}>{};
-    constexpr auto val = cuda::args::__unwrap(sa);
-    static_assert(val[0] == 10);
-    static_assert(val[2] == 30);
-  }
-#endif // TEST_HAS_CLASS_NTTP
 }
 
 int main(int, char**)
diff --git a/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp b/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp
index eada29e23de..ad54a8c745e 100644
--- a/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp
@@ -73,6 +73,11 @@ TEST_FUNC constexpr int process_segments(_SegSizeArg __seg_size)
   }
 }
 
+constexpr ::cuda::std::array<int, 3> cudaarr_shared = {64, 128, 256};
+constexpr int carr_shared[]                         = {64, 128, 256};
+constexpr ::cuda::std::array<int, 3> cudaarr_global = {64, 128, 512};
+constexpr int carr_global[]                         = {64, 128, 512};
+
 TEST_FUNC constexpr bool test()
 {
   // Plain scalar: no bounds, global memory, buffer clamped to default
@@ -101,10 +106,15 @@ TEST_FUNC constexpr bool test()
     assert(process_segments(seg_size) == 128);
   }
 
-#if TEST_HAS_CLASS_NTTP
   // __constant_sequence: array sequence, highest fits in shared memory
   {
-    constexpr auto seg_sizes = cuda::args::__constant_sequence<cuda::std::array{64, 128, 256}>{};
+    constexpr auto seg_sizes = cuda::args::make_constant_sequence<carr_shared>();
+    static_assert(select_variant(seg_sizes) == algorithm_variant::shared_memory);
+    assert(compute_buffer_size(seg_sizes, 3) == 256 * 3);
+    assert(process_segments(seg_sizes) == 64 + 128 + 256);
+  }
+  {
+    constexpr auto seg_sizes = cuda::args::make_constant_sequence<cudaarr_shared>();
     static_assert(select_variant(seg_sizes) == algorithm_variant::shared_memory);
     assert(compute_buffer_size(seg_sizes, 3) == 256 * 3);
     assert(process_segments(seg_sizes) == 64 + 128 + 256);
@@ -112,12 +122,17 @@ TEST_FUNC constexpr bool test()
 
   // __constant_sequence: array sequence, highest exceeds shared memory, buffer clamped
   {
-    constexpr auto seg_sizes = cuda::args::__constant_sequence<cuda::std::array{64, 128, 512}>{};
+    constexpr auto seg_sizes = cuda::args::make_constant_sequence<carr_global>();
+    static_assert(select_variant(seg_sizes) == algorithm_variant::global_memory);
+    assert(compute_buffer_size(seg_sizes, 3) == 512 * 3);
+    assert(process_segments(seg_sizes) == 64 + 128 + 512);
+  }
+  {
+    constexpr auto seg_sizes = cuda::args::make_constant_sequence<cudaarr_global>();
     static_assert(select_variant(seg_sizes) == algorithm_variant::global_memory);
     assert(compute_buffer_size(seg_sizes, 3) == 512 * 3);
     assert(process_segments(seg_sizes) == 64 + 128 + 512);
   }
-#endif // TEST_HAS_CLASS_NTTP
 
   // immediate: tight static bounds, shared memory, buffer = value
   {
@@ -145,8 +160,9 @@ TEST_FUNC constexpr bool test()
 
   // __immediate_sequence: per-segment span with runtime bounds only
   {
-    int sizes[3]   = {64, 128, 96};
-    auto seg_sizes = cuda::args::__immediate_sequence{cuda::std::span<int>{sizes, 3}, cuda::args::bounds(1, 200)};
+    int sizes[3] = {64, 128, 96};
+    [[maybe_unused]] auto seg_sizes =
+      cuda::args::__immediate_sequence{cuda::std::span<int>{sizes, 3}, cuda::args::bounds(1, 200)};
     assert(select_variant(seg_sizes) == algorithm_variant::global_memory);
     assert(compute_buffer_size(seg_sizes, 3) == 200 * 3);
     assert(process_segments(seg_sizes) == 64 + 128 + 96);

From 7ed4d84926a9846183262ed796bf1972b3985d87 Mon Sep 17 00:00:00 2001
From: "Schoos, Marco" <marco.schoos@spark-radiance.eu>
Date: Mon, 22 Jun 2026 14:49:01 +0200
Subject: [PATCH 02/11] Fix template parameter naming

---
 libcudacxx/include/cuda/__argument/argument.h | 100 +++++++++---------
 1 file changed, 50 insertions(+), 50 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index a11f8a91179..3d5d6d4f957 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -112,39 +112,39 @@ class constant
 };
 
 //! @brief Wraps a compile-time constant argument sequence.
-template <class T, T... Vs>
+template <class _Tp, _Tp... _Vs>
 class __constant_sequence
 {
 public:
-  using __element_type                      = ::cuda::std::remove_cvref_t<T>;
-  using value_type                          = ::cuda::std::array<__element_type, sizeof...(Vs)>;
-  static constexpr ::cuda::std::size_t size = sizeof...(Vs);
+  using __element_type                      = ::cuda::std::remove_cvref_t<_Tp>;
+  using value_type                          = ::cuda::std::array<__element_type, sizeof...(_Vs)>;
+  static constexpr ::cuda::std::size_t size = sizeof...(_Vs);
 };
 
-template <class T>
+template <class _Tp>
 struct __is_builtin_array : ::cuda::std::false_type
 {};
 
-template <class T, ::cuda::std::size_t N>
-struct __is_builtin_array<T[N]> : ::cuda::std::true_type
+template <class _Tp, ::cuda::std::size_t _Size>
+struct __is_builtin_array<_Tp[_Size]> : ::cuda::std::true_type
 {};
 
-template <class T>
+template <class _Tp>
 struct __is_cuda_array : ::cuda::std::false_type
 {};
 
-template <class T, ::cuda::std::size_t N>
-struct __is_cuda_array<::cuda::std::array<T, N>> : ::cuda::std::true_type
+template <class _Tp, ::cuda::std::size_t _Size>
+struct __is_cuda_array<::cuda::std::array<_Tp, _Size>> : ::cuda::std::true_type
 {};
 
-template <class T>
+template <class _Tp>
 struct __array_extent;
 
-template <class T, ::cuda::std::size_t N>
-struct __array_extent<T[N]> : ::cuda::std::integral_constant<::cuda::std::size_t, N>{};
+template <class _Tp, ::cuda::std::size_t _Size>
+struct __array_extent<_Tp[_Size]> : ::cuda::std::integral_constant<::cuda::std::size_t, _Size>{};
 
-template <class T, ::cuda::std::size_t N>
-struct __array_extent<::cuda::std::array<T, N>> : ::cuda::std::integral_constant<::cuda::std::size_t, N>{};
+template <class _Tp, ::cuda::std::size_t _Size>
+struct __array_extent<::cuda::std::array<_Tp, _Size>> : ::cuda::std::integral_constant<::cuda::std::size_t, _Size>{};
 
 template <class>
 inline constexpr bool __always_false_v = false;
@@ -156,13 +156,13 @@ constexpr auto make_constant_sequence_impl(::cuda::std::index_sequence<Is...>)
 
   if constexpr (__is_builtin_array<raw_array>::value)
   {
-    using T = ::cuda::std::remove_cv_t<::cuda::std::remove_extent_t<raw_array>>;
-    return __constant_sequence<T, Arr[Is]...>{};
+    using _Tp = ::cuda::std::remove_cv_t<::cuda::std::remove_extent_t<raw_array>>;
+    return __constant_sequence<_Tp, Arr[Is]...>{};
   }
   else if constexpr (__is_cuda_array<raw_array>::value)
   {
-    using T = typename raw_array::value_type;
-    return __constant_sequence<T, Arr[Is]...>{};
+    using _Tp = typename raw_array::value_type;
+    return __constant_sequence<_Tp, Arr[Is]...>{};
   }
   else
   {
@@ -685,8 +685,8 @@ template <class _Arg, class _StaticBounds>
 inline constexpr bool __is_wrapper_v<immediate<_Arg, _StaticBounds>> = true;
 template <auto _Value, class _Tp>
 inline constexpr bool __is_wrapper_v<constant<_Value, _Tp>> = true;
-template <class T, T... Vs>
-inline constexpr bool __is_wrapper_v<__constant_sequence<T, Vs...>> = true;
+template <class _Tp, _Tp... _Vs>
+inline constexpr bool __is_wrapper_v<__constant_sequence<_Tp, _Vs...>> = true;
 template <class _Arg, class _StaticBounds>
 inline constexpr bool __is_wrapper_v<__immediate_sequence<_Arg, _StaticBounds>> = true;
 template <class _Arg, class _StaticBounds>
@@ -727,10 +727,10 @@ __unwrap(const constant<_Value, _Tp>&) noexcept
 }
 
 //! Unwraps a compile-time constant argument sequence into a canonical cuda::std::array value.
-template <class T, T... Vs>
-[[nodiscard]] _CCCL_API constexpr auto __unwrap(const __constant_sequence<T, Vs...>&) noexcept
+template <class _Tp, _Tp... _Vs>
+[[nodiscard]] _CCCL_API constexpr auto __unwrap(const __constant_sequence<_Tp, _Vs...>&) noexcept
 {
-  return ::cuda::std::array<T, sizeof...(Vs)>{Vs...};
+  return ::cuda::std::array<_Tp, sizeof...(_Vs)>{_Vs...};
 }
 
 template <class _Arg, class _StaticBounds>
@@ -799,18 +799,18 @@ _CCCL_API constexpr auto __constant_compute_highest() noexcept
   return constant<_Value, _Tp>::__get_value();
 }
 
-template <class T, T... Vs>
-_CCCL_API constexpr T __constant_sequence_compute_lowest() noexcept
+template <class _Tp, _Tp... _Vs>
+_CCCL_API constexpr _Tp __constant_sequence_compute_lowest() noexcept
 {
-  if constexpr (sizeof...(Vs) == 0)
+  if constexpr (sizeof...(_Vs) == 0)
   {
-    return ::cuda::std::numeric_limits<T>::lowest();
+    return ::cuda::std::numeric_limits<_Tp>::lowest();
   }
   else
   {
-    constexpr T values[] = {Vs...};
-    T __min              = values[0];
-    for (T __v : values)
+    constexpr _Tp values[] = {_Vs...};
+    _Tp __min              = values[0];
+    for (_Tp __v : values)
     {
       if (__v < __min)
       {
@@ -821,18 +821,18 @@ _CCCL_API constexpr T __constant_sequence_compute_lowest() noexcept
   }
 }
 
-template <class T, T... Vs>
-_CCCL_API constexpr T __constant_sequence_compute_highest() noexcept
+template <class _Tp, _Tp... _Vs>
+_CCCL_API constexpr _Tp __constant_sequence_compute_highest() noexcept
 {
-  if constexpr (sizeof...(Vs) == 0)
+  if constexpr (sizeof...(_Vs) == 0)
   {
-    return ::cuda::std::numeric_limits<T>::max();
+    return ::cuda::std::numeric_limits<_Tp>::max();
   }
   else
   {
-    constexpr T values[] = {Vs...};
-    T __max              = values[0];
-    for (T __v : values)
+    constexpr _Tp values[] = {_Vs...};
+    _Tp __max              = values[0];
+    for (_Tp __v : values)
     {
       if (__v > __max)
       {
@@ -891,16 +891,16 @@ struct __traits_impl<immediate<_Arg, _StaticBounds>>
   static constexpr element_type highest = __wrapper_static_highest<element_type, _StaticBounds>();
 };
 
-template <class T, T... Vs>
-struct __traits_impl<__constant_sequence<T, Vs...>>
+template <class _Tp, _Tp... _Vs>
+struct __traits_impl<__constant_sequence<_Tp, _Vs...>>
 {
-  using element_type                    = ::cuda::std::remove_cvref_t<T>;
-  using value_type                      = ::cuda::std::array<element_type, sizeof...(Vs)>;
+  using element_type                    = ::cuda::std::remove_cvref_t<_Tp>;
+  using value_type                      = ::cuda::std::array<element_type, sizeof...(_Vs)>;
   static constexpr bool is_constant     = true;
   static constexpr bool is_deferred     = false;
   static constexpr bool is_single_value = false;
-  static constexpr element_type lowest  = __constant_sequence_compute_lowest<T, Vs...>();
-  static constexpr element_type highest = __constant_sequence_compute_highest<T, Vs...>();
+  static constexpr element_type lowest  = __constant_sequence_compute_lowest<_Tp, _Vs...>();
+  static constexpr element_type highest = __constant_sequence_compute_highest<_Tp, _Vs...>();
 };
 
 template <class _Arg, class _StaticBounds>
@@ -975,10 +975,10 @@ template <auto _Value, class _Tp>
   return __constant_compute_lowest<_Value, _Tp>();
 }
 
-template <class T, T... Vs>
-[[nodiscard]] _CCCL_API constexpr auto __lowest_(__constant_sequence<T, Vs...>) noexcept
+template <class _Tp, _Tp... _Vs>
+[[nodiscard]] _CCCL_API constexpr auto __lowest_(__constant_sequence<_Tp, _Vs...>) noexcept
 {
-  return __constant_sequence_compute_lowest<T, Vs...>();
+  return __constant_sequence_compute_lowest<_Tp, _Vs...>();
 }
 
 template <class _Arg, class _StaticBounds>
@@ -1028,10 +1028,10 @@ template <auto _Value, class _Tp>
   return __constant_compute_highest<_Value, _Tp>();
 }
 
-template <class T, T... Vs>
-[[nodiscard]] _CCCL_API constexpr auto __highest_(__constant_sequence<T, Vs...>) noexcept
+template <class _Tp, _Tp... _Vs>
+[[nodiscard]] _CCCL_API constexpr auto __highest_(__constant_sequence<_Tp, _Vs...>) noexcept
 {
-  return __constant_sequence_compute_highest<T, Vs...>();
+  return __constant_sequence_compute_highest<_Tp, _Vs...>();
 }
 
 template <class _Arg, class _StaticBounds>

From 23f15cf1a926ba09f48b2537b356cdfe36304d76 Mon Sep 17 00:00:00 2001
From: "Schoos, Marco" <marco.schoos@spark-radiance.eu>
Date: Mon, 22 Jun 2026 14:55:21 +0200
Subject: [PATCH 03/11] Add _CCCL_API for make_constant_sequence. Add "__"
 prefix to funtions

---
 libcudacxx/include/cuda/__argument/argument.h |  6 +++---
 .../cuda/argument/argument_traits.pass.cpp    | 20 +++++++++----------
 .../cuda/argument/static_argument.pass.cpp    | 14 ++++++-------
 .../cuda/argument/usage_example.pass.cpp      |  8 ++++----
 4 files changed, 24 insertions(+), 24 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index 3d5d6d4f957..b989383ffe4 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -150,7 +150,7 @@ template <class>
 inline constexpr bool __always_false_v = false;
 
 template <const auto& Arr, ::cuda::std::size_t... Is>
-constexpr auto make_constant_sequence_impl(::cuda::std::index_sequence<Is...>)
+_CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequence<Is...>)
 {
   using raw_array = ::cuda::std::remove_const_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
 
@@ -173,7 +173,7 @@ constexpr auto make_constant_sequence_impl(::cuda::std::index_sequence<Is...>)
 //! @brief Makes a compile-time constant argument sequence.
 //! In C++17, Arr must have static storage duration.
 template <const auto& Arr>
-constexpr auto make_constant_sequence()
+_CCCL_API constexpr auto __make_constant_sequence()
 {
   using raw_array = ::cuda::std::remove_cv_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
 
@@ -182,7 +182,7 @@ constexpr auto make_constant_sequence()
 
   constexpr ::cuda::std::size_t N = __array_extent<raw_array>::value;
 
-  return make_constant_sequence_impl<Arr>(::cuda::std::make_index_sequence<N>{});
+  return __make_constant_sequence_impl<Arr>(::cuda::std::make_index_sequence<N>{});
 }
 
 // __assert_in_range
diff --git a/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp b/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp
index 89b4494a8df..e3fef4affb8 100644
--- a/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/argument/argument_traits.pass.cpp
@@ -107,8 +107,8 @@ TEST_FUNC void test()
   static_assert(!cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::is_deferred);
   static constexpr int carr[]                         = {1, 2, 3};
   static constexpr ::cuda::std::array<int, 3> cudaarr = {1, 2, 3};
-  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::is_deferred);
-  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::is_deferred);
+  static_assert(!cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<carr>())>::is_deferred);
+  static_assert(!cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<cudaarr>())>::is_deferred);
 
   static_assert(cuda::args::__traits<cuda::args::deferred<cuda::std::span<int, 1>>>::is_deferred);
   static_assert(cuda::args::__traits<cuda::args::deferred_sequence<cuda::std::span<int>>>::is_deferred);
@@ -124,8 +124,8 @@ TEST_FUNC void test()
   static_assert(cuda::args::__traits<cuda::args::constant<42>>::is_single_value);
 
   static_assert(!cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::is_single_value);
-  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::is_single_value);
-  static_assert(!cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::is_single_value);
+  static_assert(!cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<carr>())>::is_single_value);
+  static_assert(!cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<cudaarr>())>::is_single_value);
 
   static_assert(cuda::args::__traits<cuda::args::deferred<int*>>::is_single_value);
   static_assert(!cuda::args::__traits<cuda::args::deferred_sequence<cuda::std::span<int>>>::is_single_value);
@@ -143,10 +143,10 @@ TEST_FUNC void test()
   static_assert(cuda::std::is_same_v<cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::value_type,
                                      cuda::std::array<int, 3>>);
   static_assert(
-    cuda::std::is_same_v<cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::value_type,
+    cuda::std::is_same_v<cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<carr>())>::value_type,
                          cuda::std::array<int, 3>>);
   static_assert(
-    cuda::std::is_same_v<cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::value_type,
+    cuda::std::is_same_v<cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<cudaarr>())>::value_type,
                          cuda::std::array<int, 3>>);
 
   // --- argument_traits: lowest / highest ---
@@ -167,10 +167,10 @@ TEST_FUNC void test()
 
   static_assert(cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::lowest == 1);
   static_assert(cuda::args::__traits<cuda::args::__constant_sequence<int, 1, 2, 3>>::highest == 3);
-  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::lowest == 1);
-  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>::highest == 3);
-  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::lowest == 1);
-  static_assert(cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>::highest == 3);
+  static_assert(cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<carr>())>::lowest == 1);
+  static_assert(cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<carr>())>::highest == 3);
+  static_assert(cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<cudaarr>())>::lowest == 1);
+  static_assert(cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<cudaarr>())>::highest == 3);
 
   // --- Free function bounds on plain values ---
 
diff --git a/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp b/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp
index 3b85a8db9a8..b7d98a70aae 100644
--- a/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/argument/static_argument.pass.cpp
@@ -58,7 +58,7 @@ TEST_FUNC void test()
   // Array sequence
   {
     static constexpr int carr[] = {128, 256, 512};
-    constexpr auto sa_arr       = cuda::args::make_constant_sequence<carr>();
+    constexpr auto sa_arr       = cuda::args::__make_constant_sequence<carr>();
     static_assert(cuda::args::__unwrap(sa_arr)[0] == 128);
     static_assert(cuda::args::__unwrap(sa_arr)[1] == 256);
     static_assert(cuda::args::__unwrap(sa_arr)[2] == 512);
@@ -67,7 +67,7 @@ TEST_FUNC void test()
 
   {
     static constexpr ::cuda::std::array<int, 3> cudaarr = {128, 256, 512};
-    constexpr auto sa_arr                               = cuda::args::make_constant_sequence<cudaarr>();
+    constexpr auto sa_arr                               = cuda::args::__make_constant_sequence<cudaarr>();
     static_assert(cuda::args::__unwrap(sa_arr)[0] == 128);
     static_assert(cuda::args::__unwrap(sa_arr)[1] == 256);
     static_assert(cuda::args::__unwrap(sa_arr)[2] == 512);
@@ -84,13 +84,13 @@ TEST_FUNC void test()
   // Bounds: array sequence computes lowest/highest of elements
   {
     static constexpr int carr[] = {128, 256, 512};
-    constexpr auto sa           = cuda::args::make_constant_sequence<carr>();
+    constexpr auto sa           = cuda::args::__make_constant_sequence<carr>();
     static_assert(cuda::args::__lowest_(sa) == 128);
     static_assert(cuda::args::__highest_(sa) == 512);
   }
   {
     static constexpr ::cuda::std::array<int, 3> cudaarr = {128, 256, 512};
-    constexpr auto sa                                   = cuda::args::make_constant_sequence<cudaarr>();
+    constexpr auto sa                                   = cuda::args::__make_constant_sequence<cudaarr>();
     static_assert(cuda::args::__lowest_(sa) == 128);
     static_assert(cuda::args::__highest_(sa) == 512);
   }
@@ -98,7 +98,7 @@ TEST_FUNC void test()
   // Bounds: empty array sequence has unconstrained element bounds
   {
     static constexpr ::cuda::std::array<int, 0> cudaarr = {};
-    constexpr auto sa                                   = cuda::args::make_constant_sequence<cudaarr>();
+    constexpr auto sa                                   = cuda::args::__make_constant_sequence<cudaarr>();
     static_assert(cuda::args::__lowest_(sa) == cuda::std::numeric_limits<int>::lowest());
     static_assert(cuda::args::__highest_(sa) == (cuda::std::numeric_limits<int>::max)());
   }
@@ -129,7 +129,7 @@ TEST_FUNC void test()
   // Sequence traits
   {
     static constexpr int carr[] = {1, 2, 3};
-    using traits                = cuda::args::__traits<decltype(cuda::args::make_constant_sequence<carr>())>;
+    using traits                = cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<carr>())>;
     static_assert(traits::is_constant);
     static_assert(!traits::is_deferred);
     static_assert(!traits::is_single_value);
@@ -138,7 +138,7 @@ TEST_FUNC void test()
   }
   {
     static constexpr ::cuda::std::array<int, 3> cudaarr = {1, 2, 3};
-    using traits = cuda::args::__traits<decltype(cuda::args::make_constant_sequence<cudaarr>())>;
+    using traits = cuda::args::__traits<decltype(cuda::args::__make_constant_sequence<cudaarr>())>;
     static_assert(traits::is_constant);
     static_assert(!traits::is_deferred);
     static_assert(!traits::is_single_value);
diff --git a/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp b/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp
index ad54a8c745e..c8801a56472 100644
--- a/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp
+++ b/libcudacxx/test/libcudacxx/cuda/argument/usage_example.pass.cpp
@@ -108,13 +108,13 @@ TEST_FUNC constexpr bool test()
 
   // __constant_sequence: array sequence, highest fits in shared memory
   {
-    constexpr auto seg_sizes = cuda::args::make_constant_sequence<carr_shared>();
+    constexpr auto seg_sizes = cuda::args::__make_constant_sequence<carr_shared>();
     static_assert(select_variant(seg_sizes) == algorithm_variant::shared_memory);
     assert(compute_buffer_size(seg_sizes, 3) == 256 * 3);
     assert(process_segments(seg_sizes) == 64 + 128 + 256);
   }
   {
-    constexpr auto seg_sizes = cuda::args::make_constant_sequence<cudaarr_shared>();
+    constexpr auto seg_sizes = cuda::args::__make_constant_sequence<cudaarr_shared>();
     static_assert(select_variant(seg_sizes) == algorithm_variant::shared_memory);
     assert(compute_buffer_size(seg_sizes, 3) == 256 * 3);
     assert(process_segments(seg_sizes) == 64 + 128 + 256);
@@ -122,13 +122,13 @@ TEST_FUNC constexpr bool test()
 
   // __constant_sequence: array sequence, highest exceeds shared memory, buffer clamped
   {
-    constexpr auto seg_sizes = cuda::args::make_constant_sequence<carr_global>();
+    constexpr auto seg_sizes = cuda::args::__make_constant_sequence<carr_global>();
     static_assert(select_variant(seg_sizes) == algorithm_variant::global_memory);
     assert(compute_buffer_size(seg_sizes, 3) == 512 * 3);
     assert(process_segments(seg_sizes) == 64 + 128 + 512);
   }
   {
-    constexpr auto seg_sizes = cuda::args::make_constant_sequence<cudaarr_global>();
+    constexpr auto seg_sizes = cuda::args::__make_constant_sequence<cudaarr_global>();
     static_assert(select_variant(seg_sizes) == algorithm_variant::global_memory);
     assert(compute_buffer_size(seg_sizes, 3) == 512 * 3);
     assert(process_segments(seg_sizes) == 64 + 128 + 512);

From c7c27f9ba57033d89f936c56adcecf91f5cd01c8 Mon Sep 17 00:00:00 2001
From: "Schoos, Marco" <marco.schoos@spark-radiance.eu>
Date: Mon, 22 Jun 2026 14:58:34 +0200
Subject: [PATCH 04/11] Remove cv of _Tp in __unwrap function

---
 libcudacxx/include/cuda/__argument/argument.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index b989383ffe4..4e8654bef6d 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -730,7 +730,7 @@ __unwrap(const constant<_Value, _Tp>&) noexcept
 template <class _Tp, _Tp... _Vs>
 [[nodiscard]] _CCCL_API constexpr auto __unwrap(const __constant_sequence<_Tp, _Vs...>&) noexcept
 {
-  return ::cuda::std::array<_Tp, sizeof...(_Vs)>{_Vs...};
+  return ::cuda::std::array<::cuda::std::remove_cvref_t<_Tp>, sizeof...(_Vs)>{_Vs...};
 }
 
 template <class _Arg, class _StaticBounds>

From a62cb39109f04e674313838b91a41856fc0af174 Mon Sep 17 00:00:00 2001
From: "Schoos, Marco" <marco.schoos@spark-radiance.eu>
Date: Mon, 22 Jun 2026 15:00:55 +0200
Subject: [PATCH 05/11] Use safe version of max() function call

---
 libcudacxx/include/cuda/__argument/argument.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index 4e8654bef6d..4459b9540b1 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -826,7 +826,7 @@ _CCCL_API constexpr _Tp __constant_sequence_compute_highest() noexcept
 {
   if constexpr (sizeof...(_Vs) == 0)
   {
-    return ::cuda::std::numeric_limits<_Tp>::max();
+    return (::cuda::std::numeric_limits<_Tp>::max)();
   }
   else
   {

From 6bf8684e703a5cd69b45207cd315a8d16dacd2f3 Mon Sep 17 00:00:00 2001
From: "Schoos, Marco" <marco.schoos@spark-radiance.eu>
Date: Mon, 22 Jun 2026 15:52:47 +0200
Subject: [PATCH 06/11] Replace custom traits with existing ones

---
 libcudacxx/include/cuda/__argument/argument.h | 44 +++++++------------
 1 file changed, 15 insertions(+), 29 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index 4459b9540b1..fc7f1ebaa9c 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -28,8 +28,10 @@
 #include <cuda/std/__iterator/iterator_traits.h>
 #include <cuda/std/__iterator/readable_traits.h>
 #include <cuda/std/__ranges/concepts.h>
+#include <cuda/std/__type_traits/extent.h>
 #include <cuda/std/__type_traits/is_arithmetic.h>
 #include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_bounded_array.h>
 #include <cuda/std/__type_traits/is_integer.h>
 #include <cuda/std/__type_traits/is_integral.h>
 #include <cuda/std/__type_traits/is_same.h>
@@ -121,31 +123,6 @@ class __constant_sequence
   static constexpr ::cuda::std::size_t size = sizeof...(_Vs);
 };
 
-template <class _Tp>
-struct __is_builtin_array : ::cuda::std::false_type
-{};
-
-template <class _Tp, ::cuda::std::size_t _Size>
-struct __is_builtin_array<_Tp[_Size]> : ::cuda::std::true_type
-{};
-
-template <class _Tp>
-struct __is_cuda_array : ::cuda::std::false_type
-{};
-
-template <class _Tp, ::cuda::std::size_t _Size>
-struct __is_cuda_array<::cuda::std::array<_Tp, _Size>> : ::cuda::std::true_type
-{};
-
-template <class _Tp>
-struct __array_extent;
-
-template <class _Tp, ::cuda::std::size_t _Size>
-struct __array_extent<_Tp[_Size]> : ::cuda::std::integral_constant<::cuda::std::size_t, _Size>{};
-
-template <class _Tp, ::cuda::std::size_t _Size>
-struct __array_extent<::cuda::std::array<_Tp, _Size>> : ::cuda::std::integral_constant<::cuda::std::size_t, _Size>{};
-
 template <class>
 inline constexpr bool __always_false_v = false;
 
@@ -154,12 +131,12 @@ _CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequen
 {
   using raw_array = ::cuda::std::remove_const_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
 
-  if constexpr (__is_builtin_array<raw_array>::value)
+  if constexpr (::cuda::std::is_bounded_array_v<raw_array>)
   {
     using _Tp = ::cuda::std::remove_cv_t<::cuda::std::remove_extent_t<raw_array>>;
     return __constant_sequence<_Tp, Arr[Is]...>{};
   }
-  else if constexpr (__is_cuda_array<raw_array>::value)
+  else if constexpr (::cuda::std::__is_cuda_std_array_v<raw_array>)
   {
     using _Tp = typename raw_array::value_type;
     return __constant_sequence<_Tp, Arr[Is]...>{};
@@ -177,10 +154,19 @@ _CCCL_API constexpr auto __make_constant_sequence()
 {
   using raw_array = ::cuda::std::remove_cv_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
 
-  static_assert(__is_builtin_array<raw_array>::value || __is_cuda_array<raw_array>::value,
+  static_assert(::cuda::std::is_bounded_array_v<raw_array> || ::cuda::std::__is_cuda_std_array_v<raw_array>,
                 "make_constant_sequence requires a cuda::std::array or non-empty C-style array");
 
-  constexpr ::cuda::std::size_t N = __array_extent<raw_array>::value;
+  constexpr ::cuda::std::size_t N = []() constexpr {
+    if constexpr (::cuda::std::is_bounded_array_v<raw_array>)
+    {
+      return ::cuda::std::extent_v<raw_array>;
+    }
+    else
+    {
+      return ::cuda::std::tuple_size_v<raw_array>;
+    }
+  }();
 
   return __make_constant_sequence_impl<Arr>(::cuda::std::make_index_sequence<N>{});
 }

From f1a042639ff0e905321f9ace45704ad3bc0545fd Mon Sep 17 00:00:00 2001
From: codingwithmagga <codingwithmagga@gmail.com>
Date: Mon, 22 Jun 2026 20:10:53 +0200
Subject: [PATCH 07/11] Replace __always_false_v with existing trait

---
 libcudacxx/include/cuda/__argument/argument.h | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index fc7f1ebaa9c..5b682ac585f 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -123,9 +123,6 @@ class __constant_sequence
   static constexpr ::cuda::std::size_t size = sizeof...(_Vs);
 };
 
-template <class>
-inline constexpr bool __always_false_v = false;
-
 template <const auto& Arr, ::cuda::std::size_t... Is>
 _CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequence<Is...>)
 {
@@ -143,7 +140,7 @@ _CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequen
   }
   else
   {
-    static_assert(__always_false_v<raw_array>, "unsupported array type");
+    static_assert(::cuda::std::__always_false_v<raw_array>, "unsupported array type");
   }
 }
 

From 1580330ff24137eef0f60a5fc939dab8ce363943 Mon Sep 17 00:00:00 2001
From: codingwithmagga <codingwithmagga@gmail.com>
Date: Mon, 22 Jun 2026 20:12:29 +0200
Subject: [PATCH 08/11] Rename template parameter

---
 libcudacxx/include/cuda/__argument/argument.h | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index 5b682ac585f..974f7be0f2e 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -123,20 +123,20 @@ class __constant_sequence
   static constexpr ::cuda::std::size_t size = sizeof...(_Vs);
 };
 
-template <const auto& Arr, ::cuda::std::size_t... Is>
-_CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequence<Is...>)
+template <const auto& _Arr, ::cuda::std::size_t... _Is>
+_CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequence<_Is...>)
 {
-  using raw_array = ::cuda::std::remove_const_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
+  using raw_array = ::cuda::std::remove_const_t<::cuda::std::remove_reference_t<decltype(_Arr)>>;
 
   if constexpr (::cuda::std::is_bounded_array_v<raw_array>)
   {
     using _Tp = ::cuda::std::remove_cv_t<::cuda::std::remove_extent_t<raw_array>>;
-    return __constant_sequence<_Tp, Arr[Is]...>{};
+    return __constant_sequence<_Tp, _Arr[_Is]...>{};
   }
   else if constexpr (::cuda::std::__is_cuda_std_array_v<raw_array>)
   {
     using _Tp = typename raw_array::value_type;
-    return __constant_sequence<_Tp, Arr[Is]...>{};
+    return __constant_sequence<_Tp, _Arr[_Is]...>{};
   }
   else
   {

From 199c31253d8469c9f236e32fc75ba2cd33f45414 Mon Sep 17 00:00:00 2001
From: codingwithmagga <codingwithmagga@gmail.com>
Date: Mon, 22 Jun 2026 20:15:21 +0200
Subject: [PATCH 09/11] Apply suggestion for __raw_array

---
 libcudacxx/include/cuda/__argument/argument.h | 22 +++++++++----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index 974f7be0f2e..4a785a32e67 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -126,21 +126,21 @@ class __constant_sequence
 template <const auto& _Arr, ::cuda::std::size_t... _Is>
 _CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequence<_Is...>)
 {
-  using raw_array = ::cuda::std::remove_const_t<::cuda::std::remove_reference_t<decltype(_Arr)>>;
+  using __raw_array = ::cuda::std::remove_cvref_t<decltype(_Arr)>;
 
-  if constexpr (::cuda::std::is_bounded_array_v<raw_array>)
+  if constexpr (::cuda::std::is_bounded_array_v<__raw_array>)
   {
-    using _Tp = ::cuda::std::remove_cv_t<::cuda::std::remove_extent_t<raw_array>>;
+    using _Tp = ::cuda::std::remove_cv_t<::cuda::std::remove_extent_t<__raw_array>>;
     return __constant_sequence<_Tp, _Arr[_Is]...>{};
   }
-  else if constexpr (::cuda::std::__is_cuda_std_array_v<raw_array>)
+  else if constexpr (::cuda::std::__is_cuda_std_array_v<__raw_array>)
   {
-    using _Tp = typename raw_array::value_type;
+    using _Tp = typename __raw_array::value_type;
     return __constant_sequence<_Tp, _Arr[_Is]...>{};
   }
   else
   {
-    static_assert(::cuda::std::__always_false_v<raw_array>, "unsupported array type");
+    static_assert(::cuda::std::__always_false_v<__raw_array>, "unsupported array type");
   }
 }
 
@@ -149,19 +149,19 @@ _CCCL_API constexpr auto __make_constant_sequence_impl(::cuda::std::index_sequen
 template <const auto& Arr>
 _CCCL_API constexpr auto __make_constant_sequence()
 {
-  using raw_array = ::cuda::std::remove_cv_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
+  using __raw_array = ::cuda::std::remove_cv_t<::cuda::std::remove_reference_t<decltype(Arr)>>;
 
-  static_assert(::cuda::std::is_bounded_array_v<raw_array> || ::cuda::std::__is_cuda_std_array_v<raw_array>,
+  static_assert(::cuda::std::is_bounded_array_v<__raw_array> || ::cuda::std::__is_cuda_std_array_v<__raw_array>,
                 "make_constant_sequence requires a cuda::std::array or non-empty C-style array");
 
   constexpr ::cuda::std::size_t N = []() constexpr {
-    if constexpr (::cuda::std::is_bounded_array_v<raw_array>)
+    if constexpr (::cuda::std::is_bounded_array_v<__raw_array>)
     {
-      return ::cuda::std::extent_v<raw_array>;
+      return ::cuda::std::extent_v<__raw_array>;
     }
     else
     {
-      return ::cuda::std::tuple_size_v<raw_array>;
+      return ::cuda::std::tuple_size_v<__raw_array>;
     }
   }();
 

From bc667511ec4992baa780e284124967da2b877e40 Mon Sep 17 00:00:00 2001
From: codingwithmagga <codingwithmagga@gmail.com>
Date: Mon, 22 Jun 2026 20:22:06 +0200
Subject: [PATCH 10/11] Use min and max_element in lowest resp. highest
 computation of constant sequence

---
 libcudacxx/include/cuda/__argument/argument.h | 24 ++++---------------
 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index 4a785a32e67..9dc4b36f53d 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -791,16 +791,8 @@ _CCCL_API constexpr _Tp __constant_sequence_compute_lowest() noexcept
   }
   else
   {
-    constexpr _Tp values[] = {_Vs...};
-    _Tp __min              = values[0];
-    for (_Tp __v : values)
-    {
-      if (__v < __min)
-      {
-        __min = __v;
-      }
-    }
-    return __min;
+    constexpr _Tp __values[] = {_Vs...};
+    return static_cast<_Tp>(*::cuda::std::min_element(__values, __values + sizeof...(_Vs)));
   }
 }
 
@@ -813,16 +805,8 @@ _CCCL_API constexpr _Tp __constant_sequence_compute_highest() noexcept
   }
   else
   {
-    constexpr _Tp values[] = {_Vs...};
-    _Tp __max              = values[0];
-    for (_Tp __v : values)
-    {
-      if (__v > __max)
-      {
-        __max = __v;
-      }
-    }
-    return __max;
+    constexpr _Tp __values[] = {_Vs...};
+    return static_cast<_Tp>(*::cuda::std::max_element(__values, __values + sizeof...(_Vs)));
   }
 }
 

From 7e70275946a6a1825119218f6e2df03751b1f88b Mon Sep 17 00:00:00 2001
From: codingwithmagga <codingwithmagga@gmail.com>
Date: Fri, 26 Jun 2026 10:07:16 +0200
Subject: [PATCH 11/11] Fix merging conflict, use new __type_lowest and
 __type_highest function for constant_sequence

---
 libcudacxx/include/cuda/__argument/argument.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/libcudacxx/include/cuda/__argument/argument.h b/libcudacxx/include/cuda/__argument/argument.h
index b757020f4e8..81f1d180589 100644
--- a/libcudacxx/include/cuda/__argument/argument.h
+++ b/libcudacxx/include/cuda/__argument/argument.h
@@ -787,7 +787,7 @@ _CCCL_API constexpr _Tp __constant_sequence_compute_lowest() noexcept
 {
   if constexpr (sizeof...(_Vs) == 0)
   {
-    return ::cuda::std::numeric_limits<_Tp>::lowest();
+    return __type_lowest<_Tp>();
   }
   else
   {
@@ -801,7 +801,7 @@ _CCCL_API constexpr _Tp __constant_sequence_compute_highest() noexcept
 {
   if constexpr (sizeof...(_Vs) == 0)
   {
-    return (::cuda::std::numeric_limits<_Tp>::max)();
+    return __type_highest<_Tp>();
   }
   else
   {