cuda-cccl 0.3.0__cp313-cp313-manylinux_2_24_aarch64.whl → 0.3.2__cp313-cp313-manylinux_2_24_aarch64.whl

cuda/cccl/headers/include/cuda/__driver/driver_api.h

@@ -216,11 +216,10 @@ _CCCL_HOST_API inline void __deviceGetName(char* __name_out, int __len, int __or
   return __result;
 }
 
-_CCCL_HOST_API inline void __primaryCtxRelease(::CUdevice __dev)
+[[nodiscard]] _CCCL_HOST_API inline ::cudaError_t __primaryCtxReleaseNoThrow(::CUdevice __dev)
 {
   static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuDevicePrimaryCtxRelease);
-  // TODO we might need to ignore failure here
-  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to release context for a device", __dev);
+  return static_cast<::cudaError_t>(__driver_fn(__dev));
 }
 
 [[nodiscard]] _CCCL_HOST_API inline bool __isPrimaryCtxActive(::CUdevice __dev)
@@ -325,6 +324,109 @@ _CCCL_HOST_API void __memsetAsync(void* __dst, _Tp __value, size_t __count, ::CU
   }
 }
 
+_CCCL_HOST_API inline ::cudaError_t __mempoolCreateNoThrow(::CUmemoryPool* __pool, ::CUmemPoolProps* __props)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemPoolCreate);
+  return static_cast<::cudaError_t>(__driver_fn(__pool, __props));
+}
+
+_CCCL_HOST_API inline void __mempoolSetAttribute(::CUmemoryPool __pool, ::CUmemPool_attribute __attr, void* __value)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemPoolSetAttribute);
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to set attribute for a memory pool", __pool, __attr, __value);
+}
+
+_CCCL_HOST_API inline size_t __mempoolGetAttribute(::CUmemoryPool __pool, ::CUmemPool_attribute __attr)
+{
+  size_t __value          = 0;
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemPoolGetAttribute);
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to get attribute for a memory pool", __pool, __attr, &__value);
+  return __value;
+}
+
+_CCCL_HOST_API inline void __mempoolDestroy(::CUmemoryPool __pool)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemPoolDestroy);
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to destroy a memory pool", __pool);
+}
+
+_CCCL_HOST_API inline ::CUdeviceptr
+__mallocFromPoolAsync(::cuda::std::size_t __bytes, ::CUmemoryPool __pool, ::CUstream __stream)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemAllocFromPoolAsync);
+  ::CUdeviceptr __result  = 0;
+  ::cuda::__driver::__call_driver_fn(
+    __driver_fn, "Failed to allocate memory from a memory pool", &__result, __bytes, __pool, __stream);
+  return __result;
+}
+
+_CCCL_HOST_API inline void __mempoolTrimTo(::CUmemoryPool __pool, ::cuda::std::size_t __min_bytes_to_keep)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemPoolTrimTo);
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to trim a memory pool", __pool, __min_bytes_to_keep);
+}
+
+_CCCL_HOST_API inline ::cudaError_t __freeAsyncNoThrow(::CUdeviceptr __dptr, ::CUstream __stream)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemFreeAsync);
+  return static_cast<::cudaError_t>(__driver_fn(__dptr, __stream));
+}
+
+_CCCL_HOST_API inline void __mempoolSetAccess(::CUmemoryPool __pool, ::CUmemAccessDesc* __descs, ::size_t __count)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemPoolSetAccess);
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to set access of a memory pool", __pool, __descs, __count);
+}
+
+_CCCL_HOST_API inline ::CUmemAccess_flags __mempoolGetAccess(::CUmemoryPool __pool, ::CUmemLocation* __location)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemPoolGetAccess);
+  ::CUmemAccess_flags __flags;
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to get access of a memory pool", &__flags, __pool, __location);
+  return __flags;
+}
+
+#  if _CCCL_CTK_AT_LEAST(13, 0)
+_CCCL_HOST_API inline ::CUmemoryPool
+__getDefaultMemPool(CUmemLocation __location, CUmemAllocationType_enum __allocation_type)
+{
+  static auto __driver_fn =
+    _CCCLRT_GET_DRIVER_FUNCTION_VERSIONED(cuMemGetDefaultMemPool, cuMemGetDefaultMemPool, 13, 0);
+  ::CUmemoryPool __result = nullptr;
+  ::cuda::__driver::__call_driver_fn(
+    __driver_fn, "Failed to get default memory pool", &__result, &__location, __allocation_type);
+  return __result;
+}
+#  endif // _CCCL_CTK_AT_LEAST(13, 0)
+
+_CCCL_HOST_API inline ::CUdeviceptr __mallocManaged(::cuda::std::size_t __bytes, unsigned int __flags)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemAllocManaged);
+  ::CUdeviceptr __result  = 0;
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to allocate managed memory", &__result, __bytes, __flags);
+  return __result;
+}
+
+_CCCL_HOST_API inline void* __mallocHost(::cuda::std::size_t __bytes)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemAllocHost);
+  void* __result          = nullptr;
+  ::cuda::__driver::__call_driver_fn(__driver_fn, "Failed to allocate host memory", &__result, __bytes);
+  return __result;
+}
+
+_CCCL_HOST_API inline ::cudaError_t __freeNoThrow(::CUdeviceptr __dptr)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemFree);
+  return static_cast<::cudaError_t>(__driver_fn(__dptr));
+}
+
+_CCCL_HOST_API inline ::cudaError_t __freeHostNoThrow(void* __dptr)
+{
+  static auto __driver_fn = _CCCLRT_GET_DRIVER_FUNCTION(cuMemFreeHost);
+  return static_cast<::cudaError_t>(__driver_fn(__dptr));
+}
+
 // Unified Addressing
 
 // TODO: we don't want to have these functions here, refactoring expected

cuda/cccl/headers/include/cuda/__event/event.h

@@ -23,12 +23,13 @@
 
 #if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
 
+#  include <cuda/__device/device_ref.h>
 #  include <cuda/__driver/driver_api.h>
 #  include <cuda/__event/event_ref.h>
 #  include <cuda/__runtime/ensure_current_context.h>
 #  include <cuda/__utility/no_init.h>
+#  include <cuda/std/__utility/to_underlying.h>
 #  include <cuda/std/cstddef>
-#  include <cuda/std/utility>
 
 #  include <cuda/std/__cccl/prologue.h>
 
@@ -36,38 +37,43 @@ _CCCL_BEGIN_NAMESPACE_CUDA
 
 class timed_event;
 
+//! @brief Flags to use when creating the event.
+enum class event_flags : unsigned
+{
+  none          = cudaEventDefault,
+  blocking_sync = cudaEventBlockingSync,
+  interprocess  = cudaEventInterprocess,
+};
+
+[[nodiscard]] _CCCL_HOST_API constexpr event_flags operator|(event_flags __lhs, event_flags __rhs) noexcept
+{
+  return static_cast<event_flags>(::cuda::std::to_underlying(__lhs) | ::cuda::std::to_underlying(__rhs));
+}
+
 //! @brief An owning wrapper for an untimed `cudaEvent_t`.
 class event : public event_ref
 {
   friend class timed_event;
 
 public:
-  //! @brief Flags to use when creating the event.
-  enum class flags : unsigned
-  {
-    none          = cudaEventDefault,
-    blocking_sync = cudaEventBlockingSync,
-    interprocess  = cudaEventInterprocess,
-  };
-
   //! @brief Construct a new `event` object with timing disabled, and record
   //!        the event in the specified stream.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit event(stream_ref __stream, flags __flags = flags::none);
+  _CCCL_HOST_API explicit event(stream_ref __stream, event_flags __flags = event_flags::none);
 
   //! @brief Construct a new `event` object with timing disabled. The event can only be recorded on streams from the
   //! specified device.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit event(device_ref __device, flags __flags = flags::none)
-      : event(__device, static_cast<unsigned int>(__flags) | cudaEventDisableTiming)
+  _CCCL_HOST_API explicit event(device_ref __device, event_flags __flags = event_flags::none)
+      : event(__device, ::cuda::std::to_underlying(__flags) | cudaEventDisableTiming)
   {}
 
   //! @brief Construct a new `event` object into the moved-from state.
   //!
   //! @post `get()` returns `cudaEvent_t()`.
-  explicit constexpr event(no_init_t) noexcept
+  _CCCL_HOST_API explicit constexpr event(no_init_t) noexcept
       : event_ref(::cudaEvent_t{})
   {}
 
@@ -76,7 +82,7 @@ public:
   //! @param __other
   //!
   //! @post `__other` is in a moved-from state.
-  constexpr event(event&& __other) noexcept
+  _CCCL_HOST_API constexpr event(event&& __other) noexcept
       : event_ref(::cuda::std::exchange(__other.__event_, {}))
   {}
 
@@ -86,7 +92,7 @@ public:
   //! @brief Destroy the `event` object
   //!
   //! @note If the event fails to be destroyed, the error is silently ignored.
-  ~event()
+  _CCCL_HOST_API ~event()
   {
     if (__event_ != nullptr)
     {
@@ -101,7 +107,7 @@ public:
   //! @param __other
   //!
   //! @post `__other` is in a moved-from state.
-  event& operator=(event&& __other) noexcept
+  _CCCL_HOST_API event& operator=(event&& __other) noexcept
   {
     event __tmp(::cuda::std::move(__other));
     ::cuda::std::swap(__event_, __tmp.__event_);
@@ -118,7 +124,7 @@ public:
   //! @return event The constructed `event` object
   //!
   //! @note The constructed `event` object takes ownership of the native handle.
-  [[nodiscard]] static event from_native_handle(::cudaEvent_t __evnt) noexcept
+  [[nodiscard]] static _CCCL_HOST_API event from_native_handle(::cudaEvent_t __evnt) noexcept
   {
     return event(__evnt);
   }
@@ -134,26 +140,21 @@ public:
   //! @return cudaEvent_t The native handle being held by the `event` object.
   //!
   //! @post The event object is in a moved-from state.
-  [[nodiscard]] constexpr ::cudaEvent_t release() noexcept
+  [[nodiscard]] _CCCL_HOST_API constexpr ::cudaEvent_t release() noexcept
   {
     return ::cuda::std::exchange(__event_, {});
   }
 
-  [[nodiscard]] friend constexpr flags operator|(flags __lhs, flags __rhs) noexcept
-  {
-    return static_cast<flags>(static_cast<unsigned>(__lhs) | static_cast<unsigned>(__rhs));
-  }
-
 private:
   // Use `event::from_native_handle(e)` to construct an owning `event`
   // object from a `cudaEvent_t` handle.
-  explicit constexpr event(::cudaEvent_t __evnt) noexcept
+  _CCCL_HOST_API explicit constexpr event(::cudaEvent_t __evnt) noexcept
       : event_ref(__evnt)
   {}
 
-  explicit event(stream_ref __stream, unsigned __flags);
+  _CCCL_HOST_API explicit event(stream_ref __stream, unsigned __flags);
 
-  explicit event(device_ref __device, unsigned __flags)
+  _CCCL_HOST_API explicit event(device_ref __device, unsigned __flags)
       : event_ref(::cudaEvent_t{})
   {
     [[maybe_unused]] __ensure_current_context __ctx_setter(__device);

cuda/cccl/headers/include/cuda/__event/event_ref.h

@@ -56,7 +56,7 @@ public:
   //!
   //! @note: It is the callers responsibility to ensure the `event_ref` does not
   //! outlive the event denoted by the `cudaEvent_t` handle.
-  constexpr event_ref(::cudaEvent_t __evnt) noexcept
+  _CCCL_HOST_API constexpr event_ref(::cudaEvent_t __evnt) noexcept
       : __event_(__evnt)
   {}
 
@@ -108,7 +108,7 @@ public:
   //! @brief Retrieve the native `cudaEvent_t` handle.
   //!
   //! @return cudaEvent_t The native handle being held by the event_ref object.
-  [[nodiscard]] constexpr ::cudaEvent_t get() const noexcept
+  [[nodiscard]] _CCCL_HOST_API constexpr ::cudaEvent_t get() const noexcept
   {
     return __event_;
   }
@@ -116,7 +116,7 @@ public:
   //! @brief Checks if the `event_ref` is valid
   //!
   //! @return true if the `event_ref` is valid, false otherwise.
-  [[nodiscard]] explicit constexpr operator bool() const noexcept
+  [[nodiscard]] _CCCL_HOST_API explicit constexpr operator bool() const noexcept
   {
     return __event_ != nullptr;
   }
@@ -129,7 +129,7 @@ public:
   //! @param __lhs The first `event_ref` to compare
   //! @param __rhs The second `event_ref` to compare
   //! @return true if `lhs` and `rhs` refer to the same `cudaEvent_t` object.
-  [[nodiscard]] friend constexpr bool operator==(event_ref __lhs, event_ref __rhs) noexcept
+  [[nodiscard]] friend _CCCL_HOST_API constexpr bool operator==(event_ref __lhs, event_ref __rhs) noexcept
   {
     return __lhs.__event_ == __rhs.__event_;
   }
@@ -142,7 +142,7 @@ public:
   //! @param __lhs The first `event_ref` to compare
   //! @param __rhs The second `event_ref` to compare
   //! @return true if `lhs` and `rhs` refer to different `cudaEvent_t` objects.
-  [[nodiscard]] friend constexpr bool operator!=(event_ref __lhs, event_ref __rhs) noexcept
+  [[nodiscard]] friend _CCCL_HOST_API constexpr bool operator!=(event_ref __lhs, event_ref __rhs) noexcept
   {
     return __lhs.__event_ != __rhs.__event_;
   }

cuda/cccl/headers/include/cuda/__event/timed_event.h

@@ -26,10 +26,12 @@
 
 #if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
 
+#  include <cuda/__device/device_ref.h>
 #  include <cuda/__driver/driver_api.h>
 #  include <cuda/__event/event.h>
 #  include <cuda/__utility/no_init.h>
 #  include <cuda/std/__chrono/duration.h>
+#  include <cuda/std/__utility/to_underlying.h>
 #  include <cuda/std/cstddef>
 
 #  include <cuda/std/__cccl/prologue.h>
@@ -44,20 +46,20 @@ public:
   //!        and record the event on the specified stream.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit timed_event(stream_ref __stream, flags __flags = flags::none);
+  _CCCL_HOST_API explicit timed_event(stream_ref __stream, event_flags __flags = event_flags::none);
 
   //! @brief Construct a new `timed_event` object with the specified flags. The event can only be recorded on streams
   //! from the specified device.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit timed_event(device_ref __device, flags __flags = flags::none)
-      : event(__device, static_cast<unsigned>(__flags))
+  _CCCL_HOST_API explicit timed_event(device_ref __device, event_flags __flags = event_flags::none)
+      : event(__device, ::cuda::std::to_underlying(__flags))
   {}
 
   //! @brief Construct a new `timed_event` object into the moved-from state.
   //!
   //! @post `get()` returns `cudaEvent_t()`.
-  explicit constexpr timed_event(no_init_t) noexcept
+  _CCCL_HOST_API explicit constexpr timed_event(no_init_t) noexcept
       : event(no_init)
   {}
 
@@ -73,7 +75,7 @@ public:
   //! @return timed_event The constructed `timed_event` object
   //!
   //! @note The constructed `timed_event` object takes ownership of the native handle.
-  [[nodiscard]] static timed_event from_native_handle(::cudaEvent_t __evnt) noexcept
+  [[nodiscard]] static _CCCL_HOST_API timed_event from_native_handle(::cudaEvent_t __evnt) noexcept
   {
     return timed_event(__evnt);
   }
@@ -94,7 +96,8 @@ public:
   //! @return cuda::std::chrono::nanoseconds The elapsed time in nanoseconds.
   //!
   //! @note The elapsed time has a resolution of approximately 0.5 microseconds.
-  [[nodiscard]] friend ::cuda::std::chrono::nanoseconds operator-(const timed_event& __end, const timed_event& __start)
+  [[nodiscard]] friend _CCCL_HOST_API ::cuda::std::chrono::nanoseconds
+  operator-(const timed_event& __end, const timed_event& __start)
   {
     const auto __ms = ::cuda::__driver::__eventElapsedTime(__start.get(), __end.get());
     return ::cuda::std::chrono::nanoseconds(static_cast<::cuda::std::chrono::nanoseconds::rep>(__ms * 1'000'000.0));
@@ -103,7 +106,7 @@ public:
 private:
   // Use `timed_event::from_native_handle(e)` to construct an owning `timed_event`
   // object from a `cudaEvent_t` handle.
-  explicit constexpr timed_event(::cudaEvent_t __evnt) noexcept
+  _CCCL_HOST_API explicit constexpr timed_event(::cudaEvent_t __evnt) noexcept
       : event(__evnt)
   {}
 };

cuda/cccl/headers/include/cuda/__fwd/devices.h

@@ -0,0 +1,44 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___FWD_DEVICES_H
+#define _CUDA___FWD_DEVICES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/span.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+class __physical_device;
+class device_ref;
+template <::cudaDeviceAttr _Attr>
+struct __dev_attr;
+struct arch_traits_t;
+class compute_capability;
+enum class arch_id : int;
+
+inline constexpr int __arch_specific_id_multiplier = 100000;
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___FWD_DEVICES_H

cuda/cccl/headers/include/cuda/__fwd/zip_iterator.h

@@ -42,6 +42,15 @@ inline constexpr bool __is_zip_function = false;
 template <class _Fn>
 inline constexpr bool __is_zip_function<zip_function<_Fn>> = true;
 
+template <class _Fn, class... _Iterators>
+class zip_transform_iterator;
+
+template <class>
+inline constexpr bool __is_zip_transform_iterator = false;
+
+template <class _Fn, class... _Iterators>
+inline constexpr bool __is_zip_transform_iterator<zip_transform_iterator<_Fn, _Iterators...>> = true;
+
 _CCCL_END_NAMESPACE_CUDA
 
 #include <cuda/std/__cccl/epilogue.h>

cuda/cccl/headers/include/cuda/__iterator/constant_iterator.h

@@ -23,6 +23,7 @@
 
 #include <cuda/std/__iterator/concepts.h>
 #include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__ranges/compressed_movable_box.h>
 #include <cuda/std/__ranges/movable_box.h>
 #include <cuda/std/__type_traits/is_nothrow_copy_constructible.h>
 #include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
@@ -64,8 +65,28 @@ class constant_iterator
 private:
   static_assert(::cuda::std::__integer_like<_Index>, "The index type of cuda::constant_iterator must be integer-like!");
 
-  ::cuda::std::ranges::__movable_box<_Tp> __value_{::cuda::std::in_place};
-  _Index __index_ = 0;
+  // Not a base because then the friend operators would be ambiguous
+  ::cuda::std::__compressed_movable_box<_Index, _Tp> __store_;
+
+  [[nodiscard]] _CCCL_API constexpr _Index& __index() noexcept
+  {
+    return __store_.template __get<0>();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const _Index& __index() const noexcept
+  {
+    return __store_.template __get<0>();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr _Tp& __value() noexcept
+  {
+    return __store_.template __get<1>();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const _Tp& __value() const noexcept
+  {
+    return __store_.template __get<1>();
+  }
 
 public:
   using iterator_concept  = ::cuda::std::random_access_iterator_tag;
@@ -78,22 +99,17 @@ public:
   using reference = _Tp;
   using pointer   = void;
 
-#if _CCCL_HAS_CONCEPTS()
-  _CCCL_HIDE_FROM_ABI constant_iterator()
-    requires ::cuda::std::default_initializable<_Tp>
-  = default;
-#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
   _CCCL_TEMPLATE(class _Tp2 = _Tp)
   _CCCL_REQUIRES(::cuda::std::default_initializable<_Tp2>)
-  _CCCL_API constexpr constant_iterator() noexcept(::cuda::std::is_nothrow_default_constructible_v<_Tp2>) {}
-#endif // !_CCCL_HAS_CONCEPTS()
+  _CCCL_API constexpr constant_iterator() noexcept(::cuda::std::is_nothrow_default_constructible_v<_Tp2>)
+      : __store_()
+  {}
 
   //! @brief Creates a @c constant_iterator from a value. The index is set to zero
   //! @param __value The value to store in the @c constant_iterator
   _CCCL_EXEC_CHECK_DISABLE
   _CCCL_API constexpr constant_iterator(_Tp __value) noexcept(::cuda::std::is_nothrow_move_constructible_v<_Tp>)
-      : __value_(::cuda::std::in_place, ::cuda::std::move(__value))
-      , __index_()
+      : __store_(0, ::cuda::std::move(__value))
   {}
 
   //! @brief Creates @c constant_iterator from a value and an index
@@ -104,32 +120,31 @@ public:
   _CCCL_REQUIRES(::cuda::std::__integer_like<_Index2>)
   _CCCL_API constexpr explicit constant_iterator(_Tp __value, _Index2 __index) noexcept(
     ::cuda::std::is_nothrow_move_constructible_v<_Tp>)
-      : __value_(::cuda::std::in_place, ::cuda::std::move(__value))
-      , __index_(static_cast<_Index>(__index))
+      : __store_(static_cast<_Index>(__index), ::cuda::std::move(__value))
   {}
 
   //! @brief Returns a the current index
   [[nodiscard]] _CCCL_API constexpr difference_type index() const noexcept
   {
-    return static_cast<difference_type>(__index_);
+    return static_cast<difference_type>(__index());
   }
 
   //! @brief Returns a const reference to the stored value
   [[nodiscard]] _CCCL_API constexpr const _Tp& operator*() const noexcept
   {
-    return *__value_;
+    return __value();
   }
 
   //! @brief Returns a const reference to the stored value
   [[nodiscard]] _CCCL_API constexpr const _Tp& operator[](difference_type) const noexcept
   {
-    return *__value_;
+    return __value();
   }
 
   //! @brief Increments the stored index
   _CCCL_API constexpr constant_iterator& operator++() noexcept
   {
-    ++__index_;
+    ++__index();
     return *this;
   }
 
@@ -147,9 +162,9 @@ public:
   {
     if constexpr (::cuda::std::is_signed_v<_Index>)
     {
-      _CCCL_ASSERT(__index_ > 0, "The index must be greater than or equal to 0");
+      _CCCL_ASSERT(__index() > 0, "The index must be greater than or equal to 0");
     }
-    --__index_;
+    --__index();
     return *this;
   }
 
@@ -168,9 +183,9 @@ public:
   {
     if constexpr (::cuda::std::is_signed_v<_Index>)
     {
-      _CCCL_ASSERT(__index_ + __n >= 0, "The index must be greater than or equal to 0");
+      _CCCL_ASSERT(__index() + __n >= 0, "The index must be greater than or equal to 0");
     }
-    __index_ += static_cast<_Index>(__n);
+    __index() += static_cast<_Index>(__n);
     return *this;
   }
 
@@ -200,9 +215,9 @@ public:
   {
     if constexpr (::cuda::std::is_signed_v<_Index>)
     {
-      _CCCL_ASSERT(__index_ - __n >= 0, "The index must be greater than or equal to 0");
+      _CCCL_ASSERT(__index() - __n >= 0, "The index must be greater than or equal to 0");
     }
-    __index_ -= static_cast<_Index>(__n);
+    __index() -= static_cast<_Index>(__n);
     return *this;
   }
 
@@ -220,14 +235,14 @@ public:
   [[nodiscard]] _CCCL_API friend constexpr difference_type
   operator-(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return static_cast<difference_type>(__lhs.__index_) - static_cast<difference_type>(__rhs.__index_);
+    return static_cast<difference_type>(__lhs.__index()) - static_cast<difference_type>(__rhs.__index());
   }
 
   //! @brief Compares two @c constant_iterator for equality by comparing the index in the sequence
   [[nodiscard]] _CCCL_API friend constexpr bool
   operator==(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return __lhs.__index_ == __rhs.__index_;
+    return __lhs.__index() == __rhs.__index();
   }
 
 #if _CCCL_STD_VER <= 2017
@@ -235,7 +250,7 @@ public:
   [[nodiscard]] _CCCL_API friend constexpr bool
   operator!=(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return __lhs.__index_ != __rhs.__index_;
+    return __lhs.__index() != __rhs.__index();
   }
 #endif // _CCCL_STD_VER <= 2017
 
@@ -244,32 +259,32 @@ public:
   [[nodiscard]] _CCCL_API friend constexpr auto
   operator<=>(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return __lhs.__index_ <=> __rhs.__index_;
+    return __lhs.__index() <=> __rhs.__index();
   }
 #else // ^^^ _LIBCUDACXX_HAS_SPACESHIP_OPERATOR() ^^^ / vvv !_LIBCUDACXX_HAS_SPACESHIP_OPERATOR() vvv
   //! @brief Compares two @c constant_iterator for less than by comparing the index in the sequence
   [[nodiscard]] _CCCL_API friend constexpr bool
   operator<(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return __lhs.__index_ < __rhs.__index_;
+    return __lhs.__index() < __rhs.__index();
   }
   //! @brief Compares two @c constant_iterator for less equal by comparing the index in the sequence
   [[nodiscard]] _CCCL_API friend constexpr bool
   operator<=(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return __lhs.__index_ <= __rhs.__index_;
+    return __lhs.__index() <= __rhs.__index();
   }
   //! @brief Compares two @c constant_iterator for greater than by comparing the index in the sequence
   [[nodiscard]] _CCCL_API friend constexpr bool
   operator>(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return __lhs.__index_ > __rhs.__index_;
+    return __lhs.__index() > __rhs.__index();
   }
   //! @brief Compares two @c constant_iterator for greater equal by comparing the index in the sequence
   [[nodiscard]] _CCCL_API friend constexpr bool
   operator>=(const constant_iterator& __lhs, const constant_iterator& __rhs) noexcept
   {
-    return __lhs.__index_ >= __rhs.__index_;
+    return __lhs.__index() >= __rhs.__index();
   }
 #endif // !_LIBCUDACXX_HAS_NO_SPACESHIP_OPERATOR()
 };