[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/cudax/test/utility/ensure_current_device.cu

@@ -0,0 +1,51 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental in CUDA C++ Core Libraries,
+// 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) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/experimental/__stream/stream.cuh>
+#include <cuda/experimental/__utility/ensure_current_device.cuh>
+#include <cuda/experimental/launch.cuh>
+
+#include <utility.cuh>
+
+namespace driver = cuda::__driver;
+
+void recursive_check_device_setter(int id)
+{
+  int cudart_id;
+  cudax::__ensure_current_device setter(cuda::device_ref{id});
+  REQUIRE(test::count_driver_stack() == cuda::devices.size() - id);
+  auto ctx = driver::__ctxGetCurrent();
+  REQUIRE_CUDART(cudaGetDevice(&cudart_id));
+  REQUIRE(cudart_id == id);
+
+  if (id != 0)
+  {
+    recursive_check_device_setter(id - 1);
+
+    REQUIRE(test::count_driver_stack() == cuda::devices.size() - id);
+    REQUIRE(ctx == driver::__ctxGetCurrent());
+    REQUIRE_CUDART(cudaGetDevice(&cudart_id));
+    REQUIRE(cudart_id == id);
+  }
+}
+
+C2H_TEST("ensure current device", "[device]")
+{
+  test::empty_driver_stack();
+  // If possible use something different than REQUIRE_CUDART default 0
+  int target_device = static_cast<int>(cuda::devices.size() - 1);
+
+  SECTION("device setter")
+  {
+    recursive_check_device_setter(target_device);
+
+    REQUIRE(test::count_driver_stack() == 0);
+  }
+}

cccl_upstream/cudax/test/utility/optionally_static.cu

@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental in CUDA C++ Core Libraries,
+// 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) 2022-2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/experimental/__utility/optionally_static.cuh>
+
+#include <cstddef>
+
+#include <c2h/catch2_test_helper.h>
+
+namespace cudax = cuda::experimental;
+
+TEST_CASE("optionally_static static value", "[utility]")
+{
+  constexpr cudax::optionally_static<::std::size_t(42), 0> v1;
+  static_assert(v1.get() == 42);
+  static_assert(v1 == v1);
+  static_assert(v1 == 42UL);
+
+  constexpr cudax::optionally_static<::std::size_t(43), 0> v2;
+  static_assert(v2.get() == 43UL);
+}
+
+TEST_CASE("optionally_static dynamic value", "[utility]")
+{
+  cudax::optionally_static<::std::size_t(0), 0> v3;
+  REQUIRE(v3.get() == 0);
+  v3 = 44;
+  REQUIRE(v3.get() == 44UL);
+}
+
+TEST_CASE("optionally_static multiplication", "[utility]")
+{
+  constexpr cudax::optionally_static<::std::size_t(42), 0> v1;
+  constexpr cudax::optionally_static<::std::size_t(43), 0> v2;
+
+  static_assert(v1 * v1 == 42UL * 42UL);
+  static_assert(v1 * v2 == 42UL * 43UL);
+  static_assert(v1 * 44 == 42UL * 44UL);
+  static_assert(44 * v1 == 42UL * 44UL);
+
+  cudax::optionally_static<::std::size_t(0), 0> v3;
+  v3 = 44;
+  REQUIRE(v1 * v3 == 42 * 44);
+}
+
+TEST_CASE("optionally_static reserved product", "[utility]")
+{
+  constexpr cudax::optionally_static<3, 18> v4;
+  constexpr cudax::optionally_static<6, 18> v5;
+  static_assert(v4 * v5 == 18UL);
+  static_assert(v4 * v5 == (cudax::optionally_static<18, 18>(18)));
+}

cccl_upstream/cudax/test/utility/unstable_unique.cu

@@ -0,0 +1,101 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental in CUDA C++ Core Libraries,
+// 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) 2022-2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/std/functional>
+
+#include <cuda/experimental/__utility/unstable_unique.cuh>
+
+#include <algorithm>
+#include <iterator>
+#include <list>
+#include <vector>
+
+#include <c2h/catch2_test_helper.h>
+
+namespace cudax = cuda::experimental;
+
+TEST_CASE("unstable_unique empty range", "[utility]")
+{
+  std::vector<int> v;
+  auto new_end = cudax::unstable_unique(v.begin(), v.end());
+  REQUIRE(v.end() == new_end);
+}
+
+TEST_CASE("unstable_unique no duplicates", "[utility]")
+{
+  std::vector<int> v = {1, 2, 3, 4, 5};
+  auto new_end       = cudax::unstable_unique(v.begin(), v.end());
+  REQUIRE(v.end() == new_end);
+  REQUIRE(std::vector<int>({1, 2, 3, 4, 5}) == v);
+}
+
+TEST_CASE("unstable_unique leading duplicates", "[utility]")
+{
+  std::vector<int> v = {1, 1, 2, 3, 4, 5};
+  auto new_end       = cudax::unstable_unique(v.begin(), v.end());
+  REQUIRE(v.begin() + 5 == new_end);
+  REQUIRE(std::vector<int>({1, 5, 2, 3, 4, 5}) == v);
+}
+
+TEST_CASE("unstable_unique interleaved duplicates", "[utility]")
+{
+  std::vector<int> v = {1, 1, 2, 2, 3, 3, 4, 4, 5, 5};
+  auto new_end       = cudax::unstable_unique(v.begin(), v.end());
+  REQUIRE(v.begin() + 5 == new_end);
+  REQUIRE(std::vector<int>({1, 5, 2, 4, 3, 3, 4, 4, 5, 5}) == v);
+}
+
+TEST_CASE("unstable_unique all same", "[utility]")
+{
+  std::vector<int> v = {1, 1, 1, 1, 1};
+  auto new_end       = cudax::unstable_unique(v.begin(), v.end());
+  REQUIRE(1 + v.begin() == new_end);
+  REQUIRE(std::vector<int>({1, 1, 1, 1, 1}) == v);
+}
+
+TEST_CASE("unstable_unique trailing unique", "[utility]")
+{
+  std::vector<int> v = {1, 1, 1, 1, 1, 2};
+  auto new_end       = cudax::unstable_unique(v.begin(), v.end());
+  REQUIRE(v.begin() + 2 == new_end);
+  REQUIRE(std::vector<int>({1, 2, 1, 1, 1, 2}) == v);
+}
+
+TEST_CASE("unstable_unique with custom predicate", "[utility]")
+{
+  std::vector<int> v = {1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 3, 4, 5};
+  auto new_end       = cudax::unstable_unique(v.begin(), v.end(), cuda::std::equal_to<int>{});
+  REQUIRE(v.begin() + 5 == new_end);
+  REQUIRE(std::vector<int>{1, 5, 4, 3, 2, 1, 1, 1, 1, 2, 2, 2, 3, 4, 5} == v);
+}
+
+TEST_CASE("unstable_unique on bidirectional iterators (std::list)", "[utility]")
+{
+  // std::list has bidirectional (not random-access) iterators -- exercises the
+  // !=-based loop termination path.
+  std::list<int> l = {1, 1, 2, 2, 3, 3, 4, 4, 5, 5};
+  auto new_end     = cudax::unstable_unique(l.begin(), l.end());
+  REQUIRE(std::distance(l.begin(), new_end) == 5);
+
+  std::vector<int> deduped(l.begin(), new_end);
+  std::sort(deduped.begin(), deduped.end());
+  REQUIRE(std::vector<int>({1, 2, 3, 4, 5}) == deduped);
+}
+
+TEST_CASE("unstable_unique on bidirectional iterators with custom predicate", "[utility]")
+{
+  std::list<int> l = {1, 1, 1, 1, 1, 2, 2, 3};
+  auto new_end     = cudax::unstable_unique(l.begin(), l.end(), cuda::std::equal_to<int>{});
+  REQUIRE(std::distance(l.begin(), new_end) == 3);
+
+  std::vector<int> deduped(l.begin(), new_end);
+  std::sort(deduped.begin(), deduped.end());
+  REQUIRE(std::vector<int>({1, 2, 3}) == deduped);
+}