[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/threads/axpy-threads-2.cu

@@ -0,0 +1,115 @@
+//===----------------------------------------------------------------------===//
+//
+// 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.
+//
+//===----------------------------------------------------------------------===//
+
+/**
+ * @file
+ *
+ * @brief An AXPY kernel implemented with a task of the CUDA stream backend
+ *
+ */
+
+#include <cuda/experimental/__stf/stream/stream_ctx.cuh>
+
+#include <mutex>
+#include <thread>
+
+using namespace cuda::experimental::stf;
+
+static __global__ void cuda_sleep_kernel(long long int clock_cnt)
+{
+  long long int start_clock  = clock64();
+  long long int clock_offset = 0;
+  while (clock_offset < clock_cnt)
+  {
+    clock_offset = clock64() - start_clock;
+  }
+}
+
+void cuda_sleep(double ms, cudaStream_t stream)
+{
+  int device;
+  cudaGetDevice(&device);
+
+  // cudaDevAttrClockRate: Peak clock frequency in kilohertz;
+  int clock_rate;
+  cudaDeviceGetAttribute(&clock_rate, cudaDevAttrClockRate, device);
+
+  long long int clock_cnt = (long long int) (ms * clock_rate);
+  cuda_sleep_kernel<<<1, 1, 0, stream>>>(clock_cnt);
+}
+
+__global__ void axpy(double a, slice<const double> x, slice<double> y)
+{
+  int tid      = blockIdx.x * blockDim.x + threadIdx.x;
+  int nthreads = gridDim.x * blockDim.x;
+
+  for (int i = tid; i < x.size(); i += nthreads)
+  {
+    y(i) += a * x(i);
+  }
+}
+
+double X0(int i)
+{
+  return sin((double) i);
+}
+
+double Y0(int i)
+{
+  return cos((double) i);
+}
+
+void mytask(stream_ctx ctx, int /*id*/, logical_data<slice<double>> lX)
+{
+  // std::cout << "Thread " << id << " is executing.\n";
+
+  const size_t N = 16;
+
+  double alpha = 3.14;
+
+  auto lY = ctx.logical_data<double>(N);
+  ctx.task(lY.write())->*[](cudaStream_t, auto) {};
+
+  /* Compute Y = Y + alpha X */
+  for (size_t i = 0; i < 10; i++)
+  {
+    ctx.task(lX.read(), lY.rw())->*[&](cudaStream_t s, auto dX, auto dY) {
+      axpy<<<16, 128, 0, s>>>(alpha, dX, dY);
+      cuda_sleep(100.0, s);
+    };
+  }
+}
+
+int main()
+{
+  stream_ctx ctx;
+  const size_t N = 16;
+  auto lX        = ctx.logical_data<double>(N);
+  ctx.task(lX.write())->*[](cudaStream_t, auto) {};
+
+  std::vector<std::thread> threads;
+  // Launch 8 threads.
+  for (int i = 0; i < 10; ++i)
+  {
+    threads.emplace_back(mytask, ctx, i, lX);
+  }
+
+  // Wait for all threads to complete.
+  for (auto& th : threads)
+  {
+    th.join();
+  }
+
+  ctx.task(lX.rw())->*[&](cudaStream_t s, auto) {
+    cuda_sleep(100.0, s);
+  };
+
+  ctx.finalize();
+}

cccl_upstream/cudax/test/stf/threads/axpy-threads-graph-capture.cu

@@ -0,0 +1,85 @@
+//===----------------------------------------------------------------------===//
+//
+// 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) 2022-2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+/**
+ * @file
+ *
+ * @brief Ensure a graph_ctx can be used concurrently
+ *
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+#include <mutex>
+#include <thread>
+
+using namespace cuda::experimental::stf;
+
+__global__ void axpy(slice<int> y, slice<const int> x, int a)
+{
+  int tid      = blockIdx.x * blockDim.x + threadIdx.x;
+  int nthreads = gridDim.x * blockDim.x;
+
+  for (int i = tid; i < x.size(); i += nthreads)
+  {
+    y(i) += a * x(i);
+  }
+}
+
+void mytask(graph_ctx ctx, int /*id*/)
+{
+  const size_t N = 16;
+
+  int alpha = 3;
+
+  auto lX = ctx.logical_data<int>(N);
+  auto lY = ctx.logical_data<int>(N);
+
+  ctx.parallel_for(lX.shape(), lX.write(), lY.write())->*[] __device__(size_t i, auto x, auto y) {
+    x(i) = (1 + i);
+    y(i) = (2 + i * i);
+  };
+
+  /* Compute Y = Y + alpha X */
+  for (size_t i = 0; i < 200; i++)
+  {
+    ctx.task(lY.rw(), lX.read())->*[alpha](cudaStream_t stream, auto dY, auto dX) {
+      axpy<<<128, 64, 0, stream>>>(dY, dX, alpha);
+    };
+  }
+
+  ctx.host_launch(lX.read(), lY.read())->*[alpha](auto x, auto y) {
+    for (size_t i = 0; i < N; i++)
+    {
+      EXPECT(x(i) == 1 + i);
+      EXPECT(y(i) == 2 + i * i + 200 * alpha * x(i));
+    }
+  };
+}
+
+int main()
+{
+  graph_ctx ctx;
+
+  ::std::vector<::std::thread> threads;
+  // Launch threads
+  for (int i = 0; i < 10; ++i)
+  {
+    threads.emplace_back(mytask, ctx, i);
+  }
+
+  // Wait for all threads to complete.
+  for (auto& th : threads)
+  {
+    th.join();
+  }
+
+  ctx.finalize();
+}

cccl_upstream/cudax/test/stf/threads/axpy-threads-graph.cu

@@ -0,0 +1,74 @@
+//===----------------------------------------------------------------------===//
+//
+// 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) 2022-2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+/**
+ * @file
+ *
+ * @brief Ensure a graph_ctx can be used concurrently
+ *
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+#include <mutex>
+#include <thread>
+
+using namespace cuda::experimental::stf;
+
+void mytask(graph_ctx ctx, int /*id*/)
+{
+  const size_t N = 16;
+
+  int alpha = 3;
+
+  auto lX = ctx.logical_data<int>(N);
+  auto lY = ctx.logical_data<int>(N);
+
+  ctx.parallel_for(lX.shape(), lX.write(), lY.write())->*[] __device__(size_t i, auto x, auto y) {
+    x(i) = (1 + i);
+    y(i) = (2 + i * i);
+  };
+
+  /* Compute Y = Y + alpha X */
+  for (size_t i = 0; i < 200; i++)
+  {
+    ctx.parallel_for(lY.shape(), lY.rw(), lX.read())->*[alpha] __device__(size_t i, auto dY, auto dX) {
+      dY(i) += alpha * dX(i);
+    };
+  }
+
+  ctx.host_launch(lX.read(), lY.read())->*[alpha](auto x, auto y) {
+    for (size_t i = 0; i < N; i++)
+    {
+      EXPECT(x(i) == 1 + i);
+      EXPECT(y(i) == 2 + i * i + 200 * alpha * x(i));
+    }
+  };
+}
+
+int main()
+{
+  graph_ctx ctx;
+
+  ::std::vector<::std::thread> threads;
+  // Launch threads
+  for (int i = 0; i < 10; ++i)
+  {
+    threads.emplace_back(mytask, ctx, i);
+  }
+
+  // Wait for all threads to complete.
+  for (auto& th : threads)
+  {
+    th.join();
+  }
+
+  ctx.finalize();
+}

cccl_upstream/cudax/test/stf/threads/axpy-threads-pfor.cu

@@ -0,0 +1,114 @@
+//===----------------------------------------------------------------------===//
+//
+// 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.
+//
+//===----------------------------------------------------------------------===//
+
+/**
+ * @file
+ *
+ * @brief An AXPY kernel implemented with a task of the CUDA stream backend
+ *
+ */
+
+#include <cuda/experimental/__stf/stream/stream_ctx.cuh>
+
+#include <mutex>
+#include <thread>
+
+using namespace cuda::experimental::stf;
+
+static __global__ void cuda_sleep_kernel(long long int clock_cnt)
+{
+  long long int start_clock  = clock64();
+  long long int clock_offset = 0;
+  while (clock_offset < clock_cnt)
+  {
+    clock_offset = clock64() - start_clock;
+  }
+}
+
+void cuda_sleep(double ms, cudaStream_t stream)
+{
+  int device;
+  cudaGetDevice(&device);
+
+  // cudaDevAttrClockRate: Peak clock frequency in kilohertz;
+  int clock_rate;
+  cudaDeviceGetAttribute(&clock_rate, cudaDevAttrClockRate, device);
+
+  long long int clock_cnt = (long long int) (ms * clock_rate);
+  cuda_sleep_kernel<<<1, 1, 0, stream>>>(clock_cnt);
+}
+
+__global__ void axpy(double a, slice<const double> x, slice<double> y)
+{
+  int tid      = blockIdx.x * blockDim.x + threadIdx.x;
+  int nthreads = gridDim.x * blockDim.x;
+
+  for (int i = tid; i < x.size(); i += nthreads)
+  {
+    y(i) += a * x(i);
+  }
+}
+
+double X0(int i)
+{
+  return sin((double) i);
+}
+
+double Y0(int i)
+{
+  return cos((double) i);
+}
+
+void mytask(stream_ctx ctx, int /*id*/)
+{
+  // std::cout << "Thread " << id << " is executing.\n";
+
+  const size_t N = 16;
+
+  double alpha = 3.14;
+
+  auto lX = ctx.logical_data<double>(N);
+  auto lY = ctx.logical_data<double>(N);
+
+  ctx.parallel_for(lX.shape(), lX.write())->*[] __device__(size_t i, auto x) {
+    x(i) = 1.0;
+  };
+
+  ctx.task(lY.write())->*[](cudaStream_t, auto) {};
+
+  /* Compute Y = Y + alpha X */
+  for (size_t i = 0; i < 10; i++)
+  {
+    ctx.task(lX.read(), lY.rw())->*[&](cudaStream_t s, auto dX, auto dY) {
+      axpy<<<16, 128, 0, s>>>(alpha, dX, dY);
+      cuda_sleep(100.0, s);
+    };
+  }
+}
+
+int main()
+{
+  stream_ctx ctx;
+
+  std::vector<std::thread> threads;
+  // Launch 8 threads.
+  for (int i = 0; i < 10; ++i)
+  {
+    threads.emplace_back(mytask, ctx, i);
+  }
+
+  // Wait for all threads to complete.
+  for (auto& th : threads)
+  {
+    th.join();
+  }
+
+  ctx.finalize();
+}

cccl_upstream/cudax/test/stf/threads/axpy-threads.cu

@@ -0,0 +1,112 @@
+//===----------------------------------------------------------------------===//
+//
+// 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.
+//
+//===----------------------------------------------------------------------===//
+
+/**
+ * @file
+ *
+ * @brief An AXPY kernel implemented with a task of the CUDA stream backend
+ *
+ */
+
+#include <cuda/experimental/__stf/stream/stream_ctx.cuh>
+
+#include <mutex>
+#include <thread>
+
+using namespace cuda::experimental::stf;
+
+static __global__ void cuda_sleep_kernel(long long int clock_cnt)
+{
+  long long int start_clock  = clock64();
+  long long int clock_offset = 0;
+  while (clock_offset < clock_cnt)
+  {
+    clock_offset = clock64() - start_clock;
+  }
+}
+
+void cuda_sleep(double ms, cudaStream_t stream)
+{
+  int device;
+  cudaGetDevice(&device);
+
+  // cudaDevAttrClockRate: Peak clock frequency in kilohertz;
+  int clock_rate;
+  cudaDeviceGetAttribute(&clock_rate, cudaDevAttrClockRate, device);
+
+  long long int clock_cnt = (long long int) (ms * clock_rate);
+  cuda_sleep_kernel<<<1, 1, 0, stream>>>(clock_cnt);
+}
+
+__global__ void axpy(double a, slice<const double> x, slice<double> y)
+{
+  int tid      = blockIdx.x * blockDim.x + threadIdx.x;
+  int nthreads = gridDim.x * blockDim.x;
+
+  for (int i = tid; i < x.size(); i += nthreads)
+  {
+    y(i) += a * x(i);
+  }
+}
+
+double X0(int i)
+{
+  return sin((double) i);
+}
+
+double Y0(int i)
+{
+  return cos((double) i);
+}
+
+void mytask(stream_ctx ctx, int /*id*/)
+{
+  // std::cout << "Thread " << id << " is executing.\n";
+
+  const size_t N = 16;
+
+  double alpha = 3.14;
+
+  auto lX = ctx.logical_data<double>(N);
+  auto lY = ctx.logical_data<double>(N);
+
+  ctx.task(lX.write())->*[](cudaStream_t, auto) {};
+
+  ctx.task(lY.write())->*[](cudaStream_t, auto) {};
+
+  /* Compute Y = Y + alpha X */
+  for (size_t i = 0; i < 10; i++)
+  {
+    ctx.task(lX.read(), lY.rw())->*[&](cudaStream_t s, auto dX, auto dY) {
+      axpy<<<16, 128, 0, s>>>(alpha, dX, dY);
+      cuda_sleep(100.0, s);
+    };
+  }
+}
+
+int main()
+{
+  stream_ctx ctx;
+
+  std::vector<std::thread> threads;
+  // Launch 8 threads.
+  for (int i = 0; i < 10; ++i)
+  {
+    threads.emplace_back(mytask, ctx, i);
+  }
+
+  // Wait for all threads to complete.
+  for (auto& th : threads)
+  {
+    th.join();
+  }
+
+  ctx.finalize();
+}