[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/__bit/bit_reverse.h

@@ -0,0 +1,200 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___BIT_BIT_REVERSE_H
+#define _CUDA___BIT_BIT_REVERSE_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/__type_traits/is_unsigned_integer.h>
+#include <cuda/std/cstdint>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(builtin_bitreverse8)
+#  define _CCCL_BUILTIN_BITREVERSE8(...) __builtin_bitreverse8(__VA_ARGS__)
+#endif
+
+#if _CCCL_CHECK_BUILTIN(builtin_bitreverse16)
+#  define _CCCL_BUILTIN_BITREVERSE16(...) __builtin_bitreverse16(__VA_ARGS__)
+#endif
+
+#if _CCCL_CHECK_BUILTIN(builtin_bitreverse32)
+#  define _CCCL_BUILTIN_BITREVERSE32(...) __builtin_bitreverse32(__VA_ARGS__)
+#endif
+
+#if _CCCL_CHECK_BUILTIN(builtin_bitreverse64)
+#  define _CCCL_BUILTIN_BITREVERSE64(...) __builtin_bitreverse64(__VA_ARGS__)
+#endif
+
+// nvcc doesn't allow clang's bitreverse builtin
+#if _CCCL_CUDA_COMPILER(NVCC)
+#  undef _CCCL_BUILTIN_BITREVERSE8
+#  undef _CCCL_BUILTIN_BITREVERSE16
+#  undef _CCCL_BUILTIN_BITREVERSE32
+#  undef _CCCL_BUILTIN_BITREVERSE64
+#endif // _CCCL_CUDA_COMPILER(NVCC)
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+#if defined(_CCCL_BUILTIN_BITREVERSE32)
+
+template <typename _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp __bit_reverse_builtin(_Tp __value) noexcept
+{
+#  if _CCCL_HAS_INT128()
+  if constexpr (sizeof(_Tp) == sizeof(__uint128_t))
+  {
+    auto __high = static_cast<__uint128_t>(_CCCL_BUILTIN_BITREVERSE64(static_cast<::cuda::std::uint64_t>(__value)))
+               << 64;
+    auto __low =
+      static_cast<__uint128_t>(_CCCL_BUILTIN_BITREVERSE64(static_cast<::cuda::std::uint64_t>(__value >> 64)));
+    return __high | __low;
+  }
+#  endif // _CCCL_HAS_INT128()
+  if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint64_t))
+  {
+    return _CCCL_BUILTIN_BITREVERSE64(__value);
+  }
+  else if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint32_t))
+  {
+    return _CCCL_BUILTIN_BITREVERSE32(__value);
+  }
+  else if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint16_t))
+  {
+    return _CCCL_BUILTIN_BITREVERSE16(__value);
+  }
+  else
+  {
+    return _CCCL_BUILTIN_BITREVERSE8(__value);
+  }
+}
+
+#endif // defined(_CCCL_BUILTIN_BITREVERSE32)
+
+#if _CCCL_CUDA_COMPILATION()
+
+template <typename _Tp>
+[[nodiscard]] _CCCL_DEVICE_API constexpr _Tp __bit_reverse_device(_Tp __value) noexcept
+{
+#  if _CCCL_HAS_INT128()
+  if constexpr (sizeof(_Tp) == sizeof(__uint128_t))
+  {
+    auto __high = static_cast<__uint128_t>(::cuda::__bit_reverse_device(static_cast<::cuda::std::uint64_t>(__value)))
+               << 64;
+    auto __low =
+      static_cast<__uint128_t>(::cuda::__bit_reverse_device(static_cast<::cuda::std::uint64_t>(__value >> 64)));
+    return __high | __low;
+  }
+#  endif // _CCCL_HAS_INT128()
+  if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint64_t))
+  {
+    NV_IF_TARGET(NV_IS_DEVICE, (return ::__brevll(__value);))
+  }
+  else if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint32_t))
+  {
+    NV_IF_TARGET(NV_IS_DEVICE, (return ::__brev(__value);))
+  }
+  else if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint16_t))
+  {
+    NV_IF_TARGET(NV_IS_DEVICE, (return ::__brev(static_cast<::cuda::std::uint32_t>(__value) << 16);))
+  }
+  else
+  {
+    NV_IF_TARGET(NV_IS_DEVICE, (return ::__brev(static_cast<::cuda::std::uint32_t>(__value) << 24);))
+  }
+  _CCCL_UNREACHABLE();
+}
+
+#endif // _CCCL_CUDA_COMPILATION()
+
+template <typename _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp __bit_reverse_generic(_Tp __value) noexcept
+{
+#if _CCCL_HAS_INT128()
+  if constexpr (sizeof(_Tp) == sizeof(__uint128_t))
+  {
+    constexpr auto __c1 = __uint128_t{0x5555555555555555} << 64 | ::cuda::std::uint64_t{0x5555555555555555};
+    constexpr auto __c2 = __uint128_t{0x3333333333333333} << 64 | ::cuda::std::uint64_t{0x3333333333333333};
+    constexpr auto __c3 = __uint128_t{0x0F0F0F0F0F0F0F0F} << 64 | ::cuda::std::uint64_t{0x0F0F0F0F0F0F0F0F};
+    constexpr auto __c4 = __uint128_t{0x00FF00FF00FF00FF} << 64 | ::cuda::std::uint64_t{0x00FF00FF00FF00FF};
+    constexpr auto __c5 = __uint128_t{0x0000FFFF0000FFFF} << 64 | ::cuda::std::uint64_t{0x0000FFFF0000FFFF};
+    constexpr auto __c6 = __uint128_t{0x00000000FFFFFFFF} << 64 | ::cuda::std::uint64_t{0x00000000FFFFFFFF};
+    __value             = ((__value >> 1) & __c1) | ((__value & __c1) << 1);
+    __value             = ((__value >> 2) & __c2) | ((__value & __c2) << 2);
+    __value             = ((__value >> 4) & __c3) | ((__value & __c3) << 4);
+    __value             = ((__value >> 8) & __c4) | ((__value & __c4) << 8);
+    __value             = ((__value >> 16) & __c5) | ((__value & __c5) << 16);
+    __value             = ((__value >> 32) & __c6) | ((__value & __c6) << 32);
+    return (__value >> 64) | (__value << 64);
+  }
+#endif // _CCCL_HAS_INT128()
+  if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint64_t))
+  {
+    __value = ((__value >> 1) & 0x5555555555555555) | ((__value & 0x5555555555555555) << 1);
+    __value = ((__value >> 2) & 0x3333333333333333) | ((__value & 0x3333333333333333) << 2);
+    __value = ((__value >> 4) & 0x0F0F0F0F0F0F0F0F) | ((__value & 0x0F0F0F0F0F0F0F0F) << 4);
+    __value = ((__value >> 8) & 0x00FF00FF00FF00FF) | ((__value & 0x00FF00FF00FF00FF) << 8);
+    __value = ((__value >> 16) & 0x0000FFFF0000FFFF) | ((__value & 0x0000FFFF0000FFFF) << 16);
+    return (__value >> 32) | (__value << 32);
+  }
+  else if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint32_t))
+  {
+    __value = ((__value >> 1) & 0x55555555) | ((__value & 0x55555555) << 1);
+    __value = ((__value >> 2) & 0x33333333) | ((__value & 0x33333333) << 2);
+    __value = ((__value >> 4) & 0x0F0F0F0F) | ((__value & 0x0F0F0F0F) << 4);
+    __value = ((__value >> 8) & 0x00FF00FF) | ((__value & 0x00FF00FF) << 8);
+    return (__value >> 16) | (__value << 16);
+  }
+  else if constexpr (sizeof(_Tp) == sizeof(::cuda::std::uint16_t))
+  {
+    __value = ((__value >> 1) & 0x5555) | ((__value & 0x5555) << 1);
+    __value = ((__value >> 2) & 0x3333) | ((__value & 0x3333) << 2);
+    __value = ((__value >> 4) & 0x0F0F) | ((__value & 0x0F0F) << 4);
+    return (__value >> 8) | (__value << 8);
+  }
+  else
+  {
+    __value = ((__value >> 1) & 0x55) | ((__value & 0x55) << 1);
+    __value = ((__value >> 2) & 0x33) | ((__value & 0x33) << 2);
+    return (__value >> 4) | (__value << 4);
+  }
+}
+
+template <typename _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp bit_reverse(_Tp __value) noexcept
+{
+  static_assert(::cuda::std::__cccl_is_cv_unsigned_integer_v<_Tp>, "bit_reverse() requires unsigned integer types");
+#if !_CCCL_TILE_COMPILATION() // nvbug6085411: error: "call to non-tile function not supported!"
+  _CCCL_IF_NOT_CONSTEVAL_DEFAULT
+  {
+    NV_IF_TARGET(NV_IS_DEVICE, (return ::cuda::__bit_reverse_device(__value);))
+  }
+#endif // !_CCCL_TILE_COMPILATION()
+#if defined(_CCCL_BUILTIN_BITREVERSE32)
+  return ::cuda::__bit_reverse_builtin(__value);
+#else
+  return ::cuda::__bit_reverse_generic(__value);
+#endif
+}
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___BIT_BIT_REVERSE_H

cccl_upstream/libcudacxx/include/cuda/__bit/bitfield.h

@@ -0,0 +1,129 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___BIT_BITFILED_INSERT_EXTRACT_H
+#define _CUDA___BIT_BITFILED_INSERT_EXTRACT_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/__bit/bitmask.h>
+#include <cuda/std/__bit/shl.h>
+#include <cuda/std/__bit/shr.h>
+#include <cuda/std/__limits/numeric_limits.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/is_unsigned_integer.h>
+#include <cuda/std/cstdint>
+#include <cuda/std/limits>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+#if !_CCCL_TILE_COMPILATION() // error: asm statement is unsupported in tile code
+#  if __cccl_ptx_isa >= 200
+
+[[nodiscard]]
+_CCCL_DEVICE_API inline uint32_t __bfi(uint32_t __dest, uint32_t __source, int __start, int __width) noexcept
+{
+  asm("bfi.b32 %0, %1, %2, %3, %4;" : "=r"(__dest) : "r"(__source), "r"(__dest), "r"(__start), "r"(__width));
+  return __dest;
+}
+
+[[nodiscard]] _CCCL_DEVICE_API inline uint64_t
+__bfi(uint64_t __dest, uint64_t __source, int __start, int __width) noexcept
+{
+  asm("bfi.b64 %0, %1, %2, %3, %4;" : "=l"(__dest) : "l"(__source), "l"(__dest), "r"(__start), "r"(__width));
+  return __dest;
+}
+
+[[nodiscard]] _CCCL_DEVICE_API inline uint32_t __bfe(uint32_t __value, int __start, int __width) noexcept
+{
+  uint32_t __ret;
+  asm("bfe.u32 %0, %1, %2, %3;" : "=r"(__ret) : "r"(__value), "r"(__start), "r"(__width));
+  return __ret;
+}
+
+[[nodiscard]] _CCCL_DEVICE_API inline uint64_t __bfe(uint64_t __value, int __start, int __width) noexcept
+{
+  uint64_t __ret;
+  asm("bfe.u64 %0, %1, %2, %3;" : "=l"(__ret) : "l"(__value), "r"(__start), "r"(__width));
+  return __ret;
+}
+
+#  endif // __cccl_ptx_isa >= 200
+#endif // !_CCCL_TILE_COMPILATION()
+
+template <typename _Tp>
+[[nodiscard]]
+_CCCL_API constexpr _Tp bitfield_insert(const _Tp __dest, const _Tp __source, int __start, int __width) noexcept
+{
+  static_assert(::cuda::std::__cccl_is_cv_unsigned_integer_v<_Tp>, "bitfield_insert() requires unsigned integer types");
+  [[maybe_unused]] constexpr auto __digits = ::cuda::std::numeric_limits<_Tp>::digits;
+  _CCCL_ASSERT(__width >= 0 && __width <= __digits, "width out of range");
+  _CCCL_ASSERT(__start >= 0 && __start <= __digits, "start position out of range");
+  _CCCL_ASSERT(__start + __width <= __digits, "start position + width out of range");
+#if !_CCCL_TILE_COMPILATION() // error: asm statement is unsupported in tile code
+  _CCCL_IF_NOT_CONSTEVAL_DEFAULT
+  {
+    if constexpr (sizeof(_Tp) <= sizeof(uint64_t))
+    {
+      // clang-format off
+      NV_DISPATCH_TARGET( // all SM < 70
+        NV_PROVIDES_SM_70, (;),
+        NV_IS_DEVICE,      (using _Up = ::cuda::std::_If<sizeof(_Tp) <= sizeof(uint32_t), uint32_t, uint64_t>;
+                            return ::cuda::__bfi(static_cast<_Up>(__dest), static_cast<_Up>(__source),
+                                                 __start, __width);))
+      // clang-format on
+    }
+  }
+#endif // !_CCCL_TILE_COMPILATION()
+  auto __mask = ::cuda::bitmask<_Tp>(__start, __width);
+  return (::cuda::std::shl(__source, static_cast<unsigned>(__start)) & __mask) | (__dest & ~__mask);
+}
+
+template <typename _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp bitfield_extract(const _Tp __value, int __start, int __width) noexcept
+{
+  static_assert(::cuda::std::__cccl_is_cv_unsigned_integer_v<_Tp>,
+                "bitfield_extract() requires unsigned integer types");
+  [[maybe_unused]] constexpr auto __digits = ::cuda::std::numeric_limits<_Tp>::digits;
+  _CCCL_ASSERT(__width >= 0 && __width <= __digits, "width out of range");
+  _CCCL_ASSERT(__start >= 0 && __start <= __digits, "start position out of range");
+  _CCCL_ASSERT(__start + __width <= __digits, "start position + width out of range");
+#if !_CCCL_TILE_COMPILATION() // error: asm statement is unsupported in tile code
+  _CCCL_IF_NOT_CONSTEVAL_DEFAULT
+  {
+    if constexpr (sizeof(_Tp) <= sizeof(uint32_t))
+    {
+      // clang-format off
+      NV_DISPATCH_TARGET( // all SM < 70
+        NV_PROVIDES_SM_70, (;),
+        NV_IS_DEVICE,      (using _Up = ::cuda::std::_If<sizeof(_Tp) <= sizeof(uint32_t), uint32_t, uint64_t>;
+                            return ::cuda::__bfe(static_cast<_Up>(__value), __start, __width);))
+      // clang-format on
+    }
+  }
+#endif // !_CCCL_TILE_COMPILATION()
+  return ::cuda::std::shr(__value, static_cast<unsigned>(__start)) & ::cuda::bitmask<_Tp>(0, __width);
+}
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___BIT_BITFILED_INSERT_EXTRACT_H

cccl_upstream/libcudacxx/include/cuda/__bit/bitmask.h

@@ -0,0 +1,62 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___BIT_BITMASK_H
+#define _CUDA___BIT_BITMASK_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/__bit/shl.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/is_unsigned_integer.h>
+#include <cuda/std/limits>
+
+#if _CCCL_CUDA_COMPILATION() && !_CCCL_TILE_COMPILATION()
+#  include <cuda/__ptx/instructions/bmsk.h>
+#endif // _CCCL_CUDA_COMPILATION() && !_CCCL_TILE_COMPILATION()
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+template <typename _Tp = uint32_t>
+[[nodiscard]] _CCCL_API constexpr _Tp bitmask(int __start, int __width) noexcept
+{
+  static_assert(::cuda::std::__cccl_is_unsigned_integer_v<_Tp>, "bitmask() requires unsigned integer types");
+  [[maybe_unused]] constexpr auto __digits = ::cuda::std::numeric_limits<_Tp>::digits;
+  _CCCL_ASSERT(__width >= 0 && __width <= __digits, "width out of range");
+  _CCCL_ASSERT(__start >= 0 && __start <= __digits, "start position out of range");
+  _CCCL_ASSERT(__start + __width <= __digits, "start position + width out of range");
+#if !_CCCL_TILE_COMPILATION() // error: asm statement is unsupported in tile code
+  _CCCL_IF_NOT_CONSTEVAL_DEFAULT
+  {
+    if constexpr (sizeof(_Tp) <= sizeof(uint32_t))
+    {
+      NV_IF_TARGET(NV_PROVIDES_SM_70, (return ::cuda::ptx::bmsk_clamp(__start, __width);))
+    }
+  }
+#endif // !_CCCL_TILE_COMPILATION()
+  return ::cuda::std::shl(static_cast<_Tp>(::cuda::std::shl(_Tp{1}, static_cast<unsigned>(__width)) - 1),
+                          static_cast<unsigned>(__start));
+}
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___BIT_BITMASK_H