[INFRA] Import NVIDIA/CCCL upstream as optimization reference library

CCCL (CUDA C++ Core Libraries) provides:
- CUB: device/block/warp-level GPU primitives (reduce, scan, sort, topk)
- Thrust: high-level parallel algorithms (transform_reduce, sort, scan)
- libcudacxx: CUDA C++ standard library (atomics, barriers, memory)
- cudax: experimental features (memory resources, allocators)
- Tuning policies: per-SM hardware-specific algorithm parameters

Competition optimization vectors mapped to CCCL:
- Output TPS (83% weight): warp_reduce, block_reduce, device_topk
- Input TPS (14% weight): device_scan, block_load, prefetch
- Cache TPS (3% weight): prefix caching strategy patterns
- Memory (0.9 util): pooled/cached/buddy allocators

Source: https://github.com/NVIDIA/cccl (shallow clone, HEAD only)
License: Apache-2.0

cccl_upstream/libcudacxx/include/cuda/__execution/determinism.h

@@ -0,0 +1,89 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___EXECUTION_DETERMINISM_H
+#define __CUDA___EXECUTION_DETERMINISM_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/__execution/require.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__execution/env.h>
+#include <cuda/std/__type_traits/is_one_of.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+namespace determinism
+{
+struct __get_determinism_t;
+
+enum class __determinism_t
+{
+  __not_guaranteed,
+  __run_to_run,
+  __gpu_to_gpu
+};
+
+template <__determinism_t _Guarantee>
+struct __determinism_holder_t : __requirement
+{
+  static constexpr __determinism_t value = _Guarantee;
+
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto query(const __get_determinism_t&) const noexcept
+    -> __determinism_holder_t<_Guarantee>
+  {
+    return *this;
+  }
+};
+
+using gpu_to_gpu_t     = __determinism_holder_t<__determinism_t::__gpu_to_gpu>;
+using run_to_run_t     = __determinism_holder_t<__determinism_t::__run_to_run>;
+using not_guaranteed_t = __determinism_holder_t<__determinism_t::__not_guaranteed>;
+
+_CCCL_GLOBAL_CONSTANT gpu_to_gpu_t gpu_to_gpu{};
+_CCCL_GLOBAL_CONSTANT run_to_run_t run_to_run{};
+_CCCL_GLOBAL_CONSTANT not_guaranteed_t not_guaranteed{};
+
+struct __get_determinism_t
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Env)
+  _CCCL_REQUIRES(::cuda::std::execution::__queryable_with<_Env, __get_determinism_t>)
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto operator()(const _Env& __env) const noexcept
+  {
+    static_assert(noexcept(__env.query(*this)));
+    return __env.query(*this);
+  }
+
+  [[nodiscard]]
+  _CCCL_NODEBUG_API static constexpr auto query(::cuda::std::execution::forwarding_query_t) noexcept -> bool
+  {
+    return true;
+  }
+};
+
+_CCCL_GLOBAL_CONSTANT auto __get_determinism = __get_determinism_t{};
+} // namespace determinism
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA___EXECUTION_DETERMINISM_H

cccl_upstream/libcudacxx/include/cuda/__execution/guarantee.h

@@ -0,0 +1,94 @@
+//===----------------------------------------------------------------------===//
+//
+// 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) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CUDA___EXECUTION_GUARANTEE_H
+#define __CUDA___EXECUTION_GUARANTEE_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/__concepts/concept_macros.h>
+#include <cuda/std/__execution/env.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_nothrow_copy_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+// Base class of all guarantees that can be passed to cuda::execution::guarantee.
+//
+// A guarantee is a promise that the caller makes to an algorithm about its input or the problem being solved (e.g. an
+// upper bound on the total number of items). Algorithms may exploit guarantees to select faster code paths or smaller
+// intermediate types, but they are never required to. This is the dual of cuda::execution::__requirement, which
+// describes a property that the caller demands from the algorithm. Unlike requirements, guarantees may be stateful,
+// i.e. they may carry a runtime value.
+class __guarantee
+{};
+
+// Query used by algorithms to extract the environment of bundled guarantees from an environment assembled by
+// cuda::execution::guarantee.
+struct __get_guarantees_t
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Env)
+  _CCCL_REQUIRES(::cuda::std::execution::__queryable_with<_Env, __get_guarantees_t>)
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto operator()(const _Env& __env) const noexcept
+  {
+    static_assert(noexcept(__env.query(*this)), "Guarantees must be queryable without throwing");
+    return __env.query(*this);
+  }
+
+  // Answering the forwarding_query query with true makes __get_guarantees_t a forwarding query: environment adaptors
+  // pass it through to the environments they wrap, so guarantees stay visible to the algorithm no matter how deeply
+  // the environment carrying them is nested.
+  [[nodiscard]]
+  _CCCL_NODEBUG_API static constexpr auto query(::cuda::std::execution::forwarding_query_t) noexcept -> bool
+  {
+    return true;
+  }
+};
+
+_CCCL_GLOBAL_CONSTANT auto __get_guarantees = __get_guarantees_t{};
+
+//! @brief Bundles a pack of guarantees into an environment that can be passed to device-wide parallel algorithms.
+//!
+//! @param __guarantees The guarantees to bundle. Each guarantee must be nothrow copy constructible. It is stored by
+//! value, preserving any runtime state it carries.
+//! @return An environment carrying the bundled guarantees.
+// The returned property is keyed by __get_guarantees_t so that individual guarantees are only visible to algorithms
+// through the guarantees environment, mirroring how cuda::execution::require exposes requirements. The key is spelled
+// __get_guarantees_t{} rather than reusing the __get_guarantees global: odr-using the global inside this inline
+// host-device function would reference a per-TU internal-linkage object in device passes (see _CCCL_GLOBAL_CONSTANT).
+template <class... _Guarantees>
+[[nodiscard]] _CCCL_NODEBUG_API constexpr auto guarantee(_Guarantees... __guarantees) noexcept
+{
+  static_assert((::cuda::std::is_base_of_v<__guarantee, _Guarantees> && ...),
+                "Only guarantees can be passed to guarantee");
+  static_assert((::cuda::std::is_nothrow_copy_constructible_v<_Guarantees> && ...),
+                "Guarantees must be nothrow copy constructible");
+
+  ::cuda::std::execution::env<_Guarantees...> __env{__guarantees...};
+
+  return ::cuda::std::execution::prop{__get_guarantees_t{}, __env};
+}
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA___EXECUTION_GUARANTEE_H

cccl_upstream/libcudacxx/include/cuda/__execution/output_ordering.h

@@ -0,0 +1,99 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___EXECUTION_OUTPUT_ORDERING_H
+#define __CUDA___EXECUTION_OUTPUT_ORDERING_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/__execution/require.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__execution/env.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_one_of.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+//! @brief Requirement describing the order in which an algorithm writes its results to the output (e.g. the top-k
+//! elements of cub::DeviceBatchedTopK).
+//!
+//! The available options are:
+//! - output_ordering::unsorted: the results may be written in any order.
+//! - output_ordering::sorted: the results are written sorted by key. Among elements that compare
+//!   equal, their relative order is unspecified.
+//! - output_ordering::stable_sorted: the results are written sorted by key and, among elements that
+//!   compare equal, their relative order matches the order of their source indices (smaller source index first).
+namespace output_ordering
+{
+struct __get_output_ordering_t;
+
+enum class __output_ordering_t
+{
+  __sorted,
+  __unsorted,
+  __stable_sorted
+};
+
+template <__output_ordering_t _Guarantee>
+struct _CCCL_DECLSPEC_EMPTY_BASES __output_ordering_holder_t
+    : __requirement
+    , ::cuda::std::integral_constant<__output_ordering_t, _Guarantee>
+{
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto query(const __get_output_ordering_t&) const noexcept
+    -> __output_ordering_holder_t<_Guarantee>
+  {
+    return *this;
+  }
+};
+
+using sorted_t        = __output_ordering_holder_t<__output_ordering_t::__sorted>;
+using stable_sorted_t = __output_ordering_holder_t<__output_ordering_t::__stable_sorted>;
+using unsorted_t      = __output_ordering_holder_t<__output_ordering_t::__unsorted>;
+
+_CCCL_GLOBAL_CONSTANT sorted_t sorted{};
+_CCCL_GLOBAL_CONSTANT stable_sorted_t stable_sorted{};
+_CCCL_GLOBAL_CONSTANT unsorted_t unsorted{};
+
+struct __get_output_ordering_t
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Env)
+  _CCCL_REQUIRES(::cuda::std::execution::__queryable_with<_Env, __get_output_ordering_t>)
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto operator()(const _Env& __env) const noexcept
+  {
+    static_assert(noexcept(__env.query(*this)));
+    return __env.query(*this);
+  }
+
+  [[nodiscard]]
+  _CCCL_NODEBUG_API static constexpr auto query(::cuda::std::execution::forwarding_query_t) noexcept -> bool
+  {
+    return true;
+  }
+};
+
+_CCCL_GLOBAL_CONSTANT auto __get_output_ordering = __get_output_ordering_t{};
+} // namespace output_ordering
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif

cccl_upstream/libcudacxx/include/cuda/__execution/policy.h

@@ -0,0 +1,43 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, 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___EXECUTION_POLICY_H
+#define _CUDA___EXECUTION_POLICY_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
+
+#if _CCCL_HAS_BACKEND_CUDA()
+
+#  include <cuda/std/__execution/policy.h>
+
+#  include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+using __cub_parallel_unsequenced_policy =
+  ::cuda::std::execution::__execution_policy_base<::cuda::std::execution::__with_backend<
+    static_cast<uint32_t>(::cuda::std::execution::__execution_policy::__parallel_unsequenced),
+    ::cuda::std::execution::__execution_backend::__cuda>()>;
+_CCCL_GLOBAL_CONSTANT __cub_parallel_unsequenced_policy gpu{};
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#  include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CCCL_HAS_BACKEND_CUDA()
+
+#endif // _CUDA___EXECUTION_POLICY_H

cccl_upstream/libcudacxx/include/cuda/__execution/require.h

@@ -0,0 +1,75 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___EXECUTION_REQUIRE_H
+#define __CUDA___EXECUTION_REQUIRE_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/__concepts/concept_macros.h>
+#include <cuda/std/__execution/env.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_empty.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+class __requirement
+{};
+
+struct __get_requirements_t
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Env)
+  _CCCL_REQUIRES(::cuda::std::execution::__queryable_with<_Env, __get_requirements_t>)
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto operator()(const _Env& __env) const noexcept
+  {
+    static_assert(noexcept(__env.query(*this)));
+    return __env.query(*this);
+  }
+
+  [[nodiscard]]
+  _CCCL_NODEBUG_API static constexpr auto query(::cuda::std::execution::forwarding_query_t) noexcept -> bool
+  {
+    return true;
+  }
+};
+
+_CCCL_GLOBAL_CONSTANT auto __get_requirements = __get_requirements_t{};
+
+template <class... _Requirements>
+[[nodiscard]] _CCCL_NODEBUG_API auto require(_Requirements...)
+{
+  static_assert((::cuda::std::is_base_of_v<__requirement, _Requirements> && ...),
+                "Only requirements can be passed to require");
+  static_assert((::cuda::std::is_empty_v<_Requirements> && ...), "Stateful requirements are not implemented");
+
+  // clang < 19 doesn't like this code
+  // since the only requirements we currently allow are in determinism.h and
+  // all of them are stateless, let's ignore incoming parameters
+  ::cuda::std::execution::env<_Requirements...> __env{};
+
+  return ::cuda::std::execution::prop{__get_requirements_t{}, __env};
+}
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA___EXECUTION_REQUIRE_H

cccl_upstream/libcudacxx/include/cuda/__execution/tie_break.h

@@ -0,0 +1,96 @@
+//===----------------------------------------------------------------------===//
+//
+// 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) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CUDA___EXECUTION_TIE_BREAK_H
+#define __CUDA___EXECUTION_TIE_BREAK_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/__execution/require.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__execution/env.h>
+#include <cuda/std/__type_traits/is_one_of.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+//! @brief Requirement describing how an algorithm breaks ties among elements that compare equal at its selection
+//! boundary (e.g. the K-th element of cub::DeviceBatchedTopK).
+//!
+//! A tie-break requirement only constrains *which* of the equal-comparing elements end up in the result set; it says
+//! nothing about the order in which the results are written (that is controlled independently by
+//! cuda::execution::output_ordering). A tie-break is only meaningful together with a deterministic execution
+//! requirement (cuda::execution::determinism::run_to_run or gpu_to_gpu).
+namespace tie_break
+{
+struct __get_tie_break_t;
+
+enum class __tie_break_t
+{
+  __unspecified, //!< Any (implementation-defined) deterministic tie-break is acceptable.
+  __prefer_smaller_index, //!< Among elements that compare equal, prefer the one(s) with the smaller source index.
+  __prefer_larger_index //!< Among elements that compare equal, prefer the one(s) with the larger source index.
+};
+
+template <__tie_break_t _Preference>
+struct __tie_break_holder_t : __requirement
+{
+  static constexpr __tie_break_t value = _Preference;
+
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto query(const __get_tie_break_t&) const noexcept
+    -> __tie_break_holder_t<_Preference>
+  {
+    return *this;
+  }
+};
+
+using unspecified_t          = __tie_break_holder_t<__tie_break_t::__unspecified>;
+using prefer_smaller_index_t = __tie_break_holder_t<__tie_break_t::__prefer_smaller_index>;
+using prefer_larger_index_t  = __tie_break_holder_t<__tie_break_t::__prefer_larger_index>;
+
+_CCCL_GLOBAL_CONSTANT unspecified_t unspecified{};
+_CCCL_GLOBAL_CONSTANT prefer_smaller_index_t prefer_smaller_index{};
+_CCCL_GLOBAL_CONSTANT prefer_larger_index_t prefer_larger_index{};
+
+struct __get_tie_break_t
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Env)
+  _CCCL_REQUIRES(::cuda::std::execution::__queryable_with<_Env, __get_tie_break_t>)
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto operator()(const _Env& __env) const noexcept
+  {
+    static_assert(noexcept(__env.query(*this)));
+    return __env.query(*this);
+  }
+
+  [[nodiscard]]
+  _CCCL_NODEBUG_API static constexpr auto query(::cuda::std::execution::forwarding_query_t) noexcept -> bool
+  {
+    return true;
+  }
+};
+
+_CCCL_GLOBAL_CONSTANT auto __get_tie_break = __get_tie_break_t{};
+} // namespace tie_break
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA___EXECUTION_TIE_BREAK_H

cccl_upstream/libcudacxx/include/cuda/__execution/tune.h

@@ -0,0 +1,81 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___EXECUTION_TUNE_H
+#define __CUDA___EXECUTION_TUNE_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/__device/compute_capability.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/semiregular.h>
+#include <cuda/std/__execution/env.h>
+#include <cuda/std/__functional/invoke.h>
+#include <cuda/std/__type_traits/is_empty.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION
+
+struct __get_tuning_t
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Env)
+  _CCCL_REQUIRES(::cuda::std::execution::__queryable_with<_Env, __get_tuning_t>)
+  [[nodiscard]] _CCCL_NODEBUG_API constexpr auto operator()(const _Env& __env) const noexcept
+  {
+    static_assert(noexcept(__env.query(*this)));
+    return __env.query(*this);
+  }
+
+  [[nodiscard]]
+  _CCCL_NODEBUG_API static constexpr auto query(::cuda::std::execution::forwarding_query_t) noexcept -> bool
+  {
+    return true;
+  }
+};
+
+_CCCL_GLOBAL_CONSTANT auto __get_tuning = __get_tuning_t{};
+
+//! @rst
+//! Creates an environment from a pack of policy selectors that can be passed to device-wide parallel algorithms to
+//! select tunings for different target architectures. See the :ref:`policy selector documentation
+//! <cub-policy-selectors>` for more information on how algorithms can be tuned.
+//! @endrst
+template <class... _PolicySelectors>
+[[nodiscard]] _CCCL_NODEBUG_API auto tune(_PolicySelectors...)
+{
+  static_assert((::cuda::std::is_empty_v<_PolicySelectors> && ...), "Policy selectors must be stateless");
+  static_assert((::cuda::std::semiregular<_PolicySelectors> && ...), "Policy selectors must be semiregular types");
+  static_assert((::cuda::std::is_invocable_v<_PolicySelectors, ::cuda::compute_capability> && ...),
+                "Policy selectors must be invocable with cuda::compute_capability");
+
+  // since all the tunings are stateless, let's ignore incoming parameters
+
+  // we use the return type of the policy_selector as tag
+  using tuning_env = ::cuda::std::execution::env<
+    ::cuda::std::execution::prop<decltype(_PolicySelectors{}(::cuda::compute_capability{})), _PolicySelectors>...>;
+
+  return ::cuda::std::execution::prop{__get_tuning_t{}, tuning_env{}};
+}
+
+_CCCL_END_NAMESPACE_CUDA_EXECUTION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA___EXECUTION_TUNE_H