[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/stf/allocators/adapter.cu

@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDASTF 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) 2022-2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/experimental/__stf/allocators/adapters.cuh>
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+int main()
+{
+  double* d_ptrA;
+  const size_t N     = 128 * 1024;
+  const size_t NITER = 10;
+
+  // User allocated memory
+  cuda_safe_call(cudaMalloc(&d_ptrA, N * sizeof(double)));
+
+  async_resources_handle handle;
+
+  cudaStream_t stream;
+  cuda_safe_call(cudaStreamCreate(&stream));
+
+  for (size_t i = 0; i < NITER; i++)
+  {
+    graph_ctx ctx(stream, handle);
+
+    // The uncached allocator of the context will be using cudaMallocAsync(...,
+    // stream) to avoid creating memory nodes in the graph (because they are
+    // costly and caching the graph also keeps memory allocated)
+    auto wrapper = stream_adapter(ctx, stream);
+
+    ctx.set_allocator(block_allocator<buddy_allocator>(ctx, wrapper.allocator()));
+
+    auto A = ctx.logical_data(make_slice(d_ptrA, N), data_place::current_device());
+
+    for (size_t k = 0; k < 4; k++)
+    {
+      auto tmp  = ctx.logical_data(A.shape());
+      auto tmp2 = ctx.logical_data(A.shape());
+      // Test device and managed memory
+      ctx.parallel_for(A.shape(), A.read(), tmp.write(), tmp2.write(data_place::managed()))
+          ->*[] __device__(size_t i, auto a, auto tmp, auto tmp2) {
+                tmp(i)  = a(i);
+                tmp2(i) = a(i);
+              };
+    }
+
+    ctx.finalize();
+
+    wrapper.clear();
+  }
+  cuda_safe_call(cudaStreamSynchronize(stream));
+}

cccl_upstream/cudax/test/stf/allocators/buddy_allocator.cu

@@ -0,0 +1,43 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDASTF 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) 2022-2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/experimental/stf.cuh>
+
+/**
+ * @brief Ensure the buddy allocation is working properly on the different backends
+ */
+
+using namespace cuda::experimental::stf;
+
+template <typename ctx_t>
+void test_buddy()
+{
+  ctx_t ctx;
+  ctx.set_allocator(block_allocator<buddy_allocator>(ctx));
+
+  std::vector<logical_data<slice<char>>> data;
+  for (size_t i = 0; i < 10; i++)
+  {
+    size_t s = (1 + i % 8) * 1024ULL * 1024ULL;
+    auto l   = ctx.logical_data(shape_of<slice<char>>(s));
+    data.push_back(l);
+
+    ctx.task(l.write())->*[](cudaStream_t, auto) {};
+  }
+
+  ctx.finalize();
+}
+
+int main(int, char**)
+{
+  test_buddy<stream_ctx>();
+  test_buddy<graph_ctx>();
+  test_buddy<context>();
+}

cccl_upstream/cudax/test/stf/allocators/cap_tmp_buffers.cu

@@ -0,0 +1,56 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDASTF 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) 2022-2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+int main(int, char**)
+{
+  context ctx;
+
+  const size_t PART_SIZE = 1024;
+  const size_t PART_CNT  = 64;
+
+  pooled_allocator_config config;
+  config.max_entries_per_place = 8;
+  auto fixed_alloc             = block_allocator<pooled_allocator>(ctx, config);
+
+  /* Create a large device buffer which will be used part by part. */
+  double* dA;
+  cuda_safe_call(cudaMalloc(&dA, PART_SIZE * PART_CNT * sizeof(double)));
+
+  for (size_t p = 0; p < PART_CNT; p++)
+  {
+    /* Create a logical data from a subset of the existing device buffer */
+    auto Ap = ctx.logical_data(make_slice(&dA[p * PART_SIZE], PART_SIZE), data_place::current_device());
+
+    ctx.parallel_for(Ap.shape(), Ap.write()).set_symbol("init_Ap")->*[p, PART_SIZE] __device__(size_t i, auto ap) {
+      ap(i) = 1.0 * (i + p * PART_SIZE);
+    };
+
+    auto tmp = ctx.logical_data(Ap.shape());
+    tmp.set_allocator(fixed_alloc);
+
+    ctx.parallel_for(Ap.shape(), Ap.read(), tmp.write()).set_symbol("set_tmp")->*
+      [] __device__(size_t i, auto ap, auto tmp) {
+        tmp(i) = 2.0 * ap(i);
+      };
+
+    ctx.parallel_for(Ap.shape(), Ap.write(), tmp.read()).set_symbol("update_Ap")
+        ->*[] __device__(size_t i, auto ap, auto tmp) {
+              ap(i) = tmp(i);
+            };
+  }
+
+  ctx.finalize();
+
+  cuda_safe_call(cudaFree(dA));
+}