[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/thrust/testing/cuda/reduce.cu

@@ -0,0 +1,118 @@
+#include <thrust/execution_policy.h>
+#include <thrust/reduce.h>
+
+#include <cuda/iterator>
+
+#include <unittest/unittest.h>
+
+template <typename ExecutionPolicy, typename Iterator, typename T, typename Iterator2>
+__global__ void reduce_kernel(ExecutionPolicy exec, Iterator first, Iterator last, T init, Iterator2 result)
+{
+  *result = thrust::reduce(exec, first, last, init);
+}
+
+#ifdef THRUST_TEST_DEVICE_SIDE
+template <typename T, typename ExecutionPolicy>
+void TestReduceDevice(ExecutionPolicy exec, const size_t n)
+{
+  thrust::host_vector<T> h_data   = unittest::random_integers<T>(n);
+  thrust::device_vector<T> d_data = h_data;
+
+  thrust::device_vector<T> d_result(1);
+
+  T init = 13;
+
+  T h_result = thrust::reduce(h_data.begin(), h_data.end(), init);
+
+  reduce_kernel<<<1, 1>>>(exec, d_data.begin(), d_data.end(), init, d_result.begin());
+  cudaError_t const err = cudaDeviceSynchronize();
+  ASSERT_EQUAL(cudaSuccess, err);
+
+  ASSERT_EQUAL(h_result, d_result[0]);
+}
+
+template <typename T>
+struct TestReduceDeviceSeq
+{
+  void operator()(const size_t n)
+  {
+    TestReduceDevice<T>(thrust::seq, n);
+  }
+};
+VariableUnitTest<TestReduceDeviceSeq, IntegralTypes> TestReduceDeviceSeqInstance;
+
+template <typename T>
+struct TestReduceDeviceDevice
+{
+  void operator()(const size_t n)
+  {
+    TestReduceDevice<T>(thrust::device, n);
+  }
+};
+VariableUnitTest<TestReduceDeviceDevice, IntegralTypes> TestReduceDeviceDeviceInstance;
+
+template <typename T>
+struct TestReduceDeviceNoSync
+{
+  void operator()(const size_t n)
+  {
+    TestReduceDevice<T>(thrust::cuda::par_nosync, n);
+  }
+};
+VariableUnitTest<TestReduceDeviceNoSync, IntegralTypes> TestReduceDeviceNoSyncInstance;
+#endif
+
+template <typename ExecutionPolicy>
+void TestReduceCudaStreams(ExecutionPolicy policy)
+{
+  using Vector = thrust::device_vector<int>;
+
+  Vector v(3);
+  v[0] = 1;
+  v[1] = -2;
+  v[2] = 3;
+
+  cudaStream_t s;
+  cudaStreamCreate(&s);
+
+  auto streampolicy = policy.on(s);
+
+  // no initializer
+  ASSERT_EQUAL(thrust::reduce(streampolicy, v.begin(), v.end()), 2);
+
+  // with initializer
+  ASSERT_EQUAL(thrust::reduce(streampolicy, v.begin(), v.end(), 10), 12);
+
+  cudaStreamDestroy(s);
+}
+
+void TestReduceCudaStreamsSync()
+{
+  TestReduceCudaStreams(thrust::cuda::par);
+}
+DECLARE_UNITTEST(TestReduceCudaStreamsSync);
+
+void TestReduceCudaStreamsNoSync()
+{
+  TestReduceCudaStreams(thrust::cuda::par_nosync);
+}
+DECLARE_UNITTEST(TestReduceCudaStreamsNoSync);
+
+#if defined(THRUST_RDC_ENABLED)
+void TestReduceLargeInput()
+{
+  using T                 = unsigned long long;
+  using OffsetT           = std::size_t;
+  const OffsetT num_items = 1ull << 32;
+
+  cuda::constant_iterator<T> d_data(T{1});
+  thrust::device_vector<T> d_result(1);
+
+  reduce_kernel<<<1, 1>>>(thrust::device, d_data, d_data + num_items, T{}, d_result.begin());
+  cudaError_t const err = cudaDeviceSynchronize();
+  ASSERT_EQUAL(cudaSuccess, err);
+
+  ASSERT_EQUAL(num_items, d_result[0]);
+}
+DECLARE_UNITTEST(TestReduceLargeInput);
+#endif