[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/max_element.cu

@@ -0,0 +1,131 @@
+#include <thrust/execution_policy.h>
+#include <thrust/extrema.h>
+
+#include <unittest/unittest.h>
+
+#ifdef THRUST_TEST_DEVICE_SIDE
+template <typename ExecutionPolicy, typename Iterator, typename Iterator2>
+__global__ void max_element_kernel(ExecutionPolicy exec, Iterator first, Iterator last, Iterator2 result)
+{
+  *result = thrust::max_element(exec, first, last);
+}
+
+template <typename ExecutionPolicy, typename Iterator, typename BinaryPredicate, typename Iterator2>
+__global__ void
+max_element_kernel(ExecutionPolicy exec, Iterator first, Iterator last, BinaryPredicate pred, Iterator2 result)
+{
+  *result = thrust::max_element(exec, first, last, pred);
+}
+
+template <typename ExecutionPolicy>
+void TestMaxElementDevice(ExecutionPolicy exec)
+{
+  size_t n                          = 1000;
+  thrust::host_vector<int> h_data   = unittest::random_samples<int>(n);
+  thrust::device_vector<int> d_data = h_data;
+
+  using iter_type = typename thrust::device_vector<int>::iterator;
+
+  thrust::device_vector<iter_type> d_result(1);
+
+  typename thrust::host_vector<int>::iterator h_max = thrust::max_element(h_data.begin(), h_data.end());
+
+  max_element_kernel<<<1, 1>>>(exec, d_data.begin(), d_data.end(), d_result.begin());
+  {
+    cudaError_t const err = cudaDeviceSynchronize();
+    ASSERT_EQUAL(cudaSuccess, err);
+  }
+
+  ASSERT_EQUAL(h_max - h_data.begin(), (iter_type) d_result[0] - d_data.begin());
+
+  typename thrust::host_vector<int>::iterator h_min =
+    thrust::max_element(h_data.begin(), h_data.end(), ::cuda::std::greater<int>());
+
+  max_element_kernel<<<1, 1>>>(exec, d_data.begin(), d_data.end(), ::cuda::std::greater<int>(), d_result.begin());
+  {
+    cudaError_t const err = cudaDeviceSynchronize();
+    ASSERT_EQUAL(cudaSuccess, err);
+  }
+
+  ASSERT_EQUAL(h_min - h_data.begin(), (iter_type) d_result[0] - d_data.begin());
+}
+
+void TestMaxElementDeviceSeq()
+{
+  TestMaxElementDevice(thrust::seq);
+}
+DECLARE_UNITTEST(TestMaxElementDeviceSeq);
+
+void TestMaxElementDeviceDevice()
+{
+  TestMaxElementDevice(thrust::device);
+}
+DECLARE_UNITTEST(TestMaxElementDeviceDevice);
+
+void TestMaxElementDeviceNoSync()
+{
+  TestMaxElementDevice(thrust::cuda::par_nosync);
+}
+DECLARE_UNITTEST(TestMaxElementDeviceNoSync);
+#endif
+
+template <typename ExecutionPolicy>
+void TestMaxElementCudaStreams(ExecutionPolicy policy)
+{
+  using Vector = thrust::device_vector<int>;
+  using T      = Vector::value_type;
+
+  Vector data(6);
+  data[0] = 3;
+  data[1] = 5;
+  data[2] = 1;
+  data[3] = 2;
+  data[4] = 5;
+  data[5] = 1;
+
+  cudaStream_t s;
+  cudaStreamCreate(&s);
+
+  auto streampolicy = policy.on(s);
+
+  ASSERT_EQUAL(*thrust::max_element(streampolicy, data.begin(), data.end()), 5);
+  ASSERT_EQUAL(thrust::max_element(streampolicy, data.begin(), data.end()) - data.begin(), 1);
+
+  ASSERT_EQUAL(*thrust::max_element(streampolicy, data.begin(), data.end(), ::cuda::std::greater<T>()), 1);
+  ASSERT_EQUAL(thrust::max_element(streampolicy, data.begin(), data.end(), ::cuda::std::greater<T>()) - data.begin(),
+               2);
+
+  cudaStreamDestroy(s);
+}
+
+void TestMaxElementCudaStreamsSync()
+{
+  TestMaxElementCudaStreams(thrust::cuda::par);
+}
+DECLARE_UNITTEST(TestMaxElementCudaStreamsSync);
+
+void TestMaxElementCudaStreamsNoSync()
+{
+  TestMaxElementCudaStreams(thrust::cuda::par_nosync);
+}
+DECLARE_UNITTEST(TestMaxElementCudaStreamsNoSync);
+
+void TestMaxElementDevicePointer()
+{
+  using Vector = thrust::device_vector<int>;
+  using T      = Vector::value_type;
+
+  Vector data(6);
+  data[0] = 3;
+  data[1] = 5;
+  data[2] = 1;
+  data[3] = 2;
+  data[4] = 5;
+  data[5] = 1;
+
+  T* raw_ptr = thrust::raw_pointer_cast(data.data());
+  size_t n   = data.size();
+  ASSERT_EQUAL(thrust::max_element(thrust::device, raw_ptr, raw_ptr + n) - raw_ptr, 1);
+  ASSERT_EQUAL(thrust::max_element(thrust::device, raw_ptr, raw_ptr + n, ::cuda::std::greater<T>()) - raw_ptr, 2);
+}
+DECLARE_UNITTEST(TestMaxElementDevicePointer);