[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/dot/basic.cu

@@ -0,0 +1,48 @@
+//===----------------------------------------------------------------------===//
+//
+// 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 This test makes sure we can generate a dot file
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+int main()
+{
+// TODO (miscco): Make it work for windows
+#if !_CCCL_COMPILER(MSVC)
+  // Generate a random filename
+  int r = rand();
+
+  char filename[64];
+  snprintf(filename, 64, "output_%d.dot", r);
+  // fprintf(stderr, "filename %s\n", filename);
+  setenv("CUDASTF_DOT_FILE", filename, 1);
+
+  stream_ctx ctx;
+
+  auto lA = ctx.logical_data(shape_of<slice<char>>(64));
+  ctx.task(lA.write())->*[](cudaStream_t, auto) {};
+  ctx.task(lA.rw())->*[](cudaStream_t, auto) {};
+  ctx.finalize();
+
+  // Call this explicitly for the purpose of the test
+  reserved::dot::instance().finish();
+
+  // Make sure the file exists, and erase it
+  // fprintf(stderr, "ERASE. ...\n");
+  EXPECT(access(filename, F_OK) != -1);
+
+  EXPECT(unlink(filename) == 0);
+#endif // !_CCCL_COMPILER(MSVC)
+}

cccl_upstream/cudax/test/stf/dot/graph_print_to_dot.cu

@@ -0,0 +1,50 @@
+//===----------------------------------------------------------------------===//
+//
+// 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 This test makes sure we can generate a dot file
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+__global__ void dummy() {}
+
+int main()
+{
+// TODO (miscco): Make it work for windows
+#if !_CCCL_COMPILER(MSVC)
+  // Generate a random filename
+  int r = rand();
+
+  char filename[64];
+  snprintf(filename, 64, "output_%d.dot", r);
+  // fprintf(stderr, "filename %s\n", filename);
+
+  graph_ctx ctx;
+
+  auto lA = ctx.logical_data(shape_of<slice<char>>(64));
+  ctx.task(lA.write())->*[](cudaStream_t s, auto) {
+    dummy<<<1, 1, 0, s>>>();
+  };
+  ctx.task(lA.rw())->*[](cudaStream_t s, auto) {
+    dummy<<<1, 1, 0, s>>>();
+  };
+  ctx.print_to_dot(filename, cudaGraphDebugDotFlagsVerbose);
+  ctx.finalize();
+
+  // Make sure the file exists, and erase it
+  EXPECT(access(filename, F_OK) != -1);
+
+  EXPECT(unlink(filename) == 0);
+#endif // !_CCCL_COMPILER(MSVC)
+}

cccl_upstream/cudax/test/stf/dot/section_movable.cu

@@ -0,0 +1,47 @@
+//===----------------------------------------------------------------------===//
+//
+// 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 Test if dot section guards are movable and if we can them as an optional value
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+#include <optional>
+#include <vector>
+
+using namespace cuda::experimental::stf;
+
+int main()
+{
+  context ctx;
+
+  // Ensure the guard is movable
+  {
+    auto g  = ctx.dot_section("foo");
+    auto g2 = mv(g);
+  }
+
+  // Ensure the guard can be stored as an optional
+  ::std::vector<::std::optional<reserved::dot_section::guard>> nested_sections;
+
+  for (size_t depth = 0; depth < 3; depth++)
+  {
+    nested_sections.emplace_back(ctx.dot_section("foo"));
+  }
+
+  for (size_t depth = 0; depth < 3; depth++)
+  {
+    nested_sections.pop_back();
+  }
+
+  ctx.finalize();
+}

cccl_upstream/cudax/test/stf/dot/sections.cu

@@ -0,0 +1,63 @@
+//===----------------------------------------------------------------------===//
+//
+// 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 This test makes sure we can generate a dot file with sections
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+int main()
+{
+// TODO (miscco): Make it work for windows
+#if !_CCCL_COMPILER(MSVC)
+  // Generate a random filename
+  int r = rand();
+
+  char filename[64];
+  snprintf(filename, 64, "output_%d.dot", r);
+  // fprintf(stderr, "filename %s\n", filename);
+  setenv("CUDASTF_DOT_FILE", filename, 1);
+
+  context ctx;
+
+  auto lA = ctx.logical_data(shape_of<slice<char>>(64));
+  auto lB = ctx.logical_data(shape_of<slice<char>>(64));
+  auto lC = ctx.logical_data(shape_of<slice<char>>(64));
+  ctx.task(lA.write()).set_symbol("initA")->*[](cudaStream_t, auto) {};
+  ctx.task(lB.write()).set_symbol("initB")->*[](cudaStream_t, auto) {};
+  ctx.task(lC.write()).set_symbol("initC")->*[](cudaStream_t, auto) {};
+  for (size_t j = 0; j < 3; j++)
+  {
+    ctx.task(lA.rw()).set_symbol("f1")->*[](cudaStream_t, auto) {};
+    auto guard = ctx.dot_section("sec_loop " + ::std::to_string(j));
+    for (size_t i = 0; i < 2; i++)
+    {
+      auto guard_inner = ctx.dot_section("sec_inner_loop " + ::std::to_string(i));
+      ctx.task(lA.read(), lB.rw()).set_symbol("f2")->*[](cudaStream_t, auto, auto) {};
+      ctx.task(lA.read(), lC.rw()).set_symbol("f2")->*[](cudaStream_t, auto, auto) {};
+      ctx.task(lB.read(), lC.read(), lA.rw()).set_symbol("f3")->*[](cudaStream_t, auto, auto, auto) {};
+    }
+  }
+  ctx.finalize();
+
+  // Call this explicitly for the purpose of the test
+  reserved::dot::instance().finish();
+
+  // Make sure the file exists, and erase it
+  // fprintf(stderr, "ERASE. ...\n");
+  EXPECT(access(filename, F_OK) != -1);
+
+  EXPECT(unlink(filename) == 0);
+#endif // !_CCCL_COMPILER(MSVC)
+}

cccl_upstream/cudax/test/stf/dot/sections_2.cu

@@ -0,0 +1,49 @@
+//===----------------------------------------------------------------------===//
+//
+// 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 This test makes sure we can generate a dot file with sections
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+int main()
+{
+// TODO (miscco): Make it work for windows
+#if !_CCCL_COMPILER(MSVC)
+  context ctx;
+  auto lA = ctx.token().set_symbol("A");
+  auto lB = ctx.token().set_symbol("B");
+  auto lC = ctx.token().set_symbol("C");
+
+  // Begin a top-level section named "foo"
+  auto s_foo = ctx.dot_section("foo");
+  for (size_t i = 0; i < 2; i++)
+  {
+    // Section named "bar" using RAII
+    auto s_bar = ctx.dot_section("bar");
+    ctx.task(lA.read(), lB.rw()).set_symbol("t1")->*[](cudaStream_t) {};
+    for (size_t j = 0; j < 2; j++)
+    {
+      // Section named "baz" using RAII
+      auto s_bar = ctx.dot_section("baz");
+      ctx.task(lA.read(), lC.rw()).set_symbol("t2")->*[](cudaStream_t) {};
+      ctx.task(lB.read(), lC.read(), lA.rw()).set_symbol("t3")->*[](cudaStream_t) {};
+      // Implicit end of section "baz"
+    }
+    // Implicit end of section "bar"
+  }
+  s_foo.end(); // Explicit end of section "foo"
+  ctx.finalize();
+#endif // !_CCCL_COMPILER(MSVC)
+}

cccl_upstream/cudax/test/stf/dot/sections_stackable.cu

@@ -0,0 +1,61 @@
+//===----------------------------------------------------------------------===//
+//
+// 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 This test makes sure we can generate a dot file with sections and stackable contexts
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+int main()
+{
+// TODO (miscco): Make it work for windows
+#if !_CCCL_COMPILER(MSVC)
+  // Test configuration constants
+  constexpr size_t buffer_size    = 64;
+  constexpr size_t num_iterations = 3;
+
+  stackable_ctx ctx;
+
+  // Create logical data for the computation pipeline
+  auto lA = ctx.logical_data(shape_of<slice<char>>(buffer_size));
+  auto lB = ctx.logical_data(shape_of<slice<char>>(buffer_size));
+  auto lC = ctx.logical_data(shape_of<slice<char>>(buffer_size));
+
+  // Initialize all logical data in a dedicated DOT section
+  auto r_init = ctx.dot_section("init");
+  ctx.task(lA.write()).set_symbol("initA")->*[](cudaStream_t, auto) {};
+  ctx.task(lB.write()).set_symbol("initB")->*[](cudaStream_t, auto) {};
+  ctx.task(lC.write()).set_symbol("initC")->*[](cudaStream_t, auto) {};
+  r_init.end();
+
+  // Test nested DOT sections with stackable contexts
+  // This creates a hierarchical structure to verify DOT graph generation
+  for (size_t j = 0; j < num_iterations; j++)
+  {
+    auto r0 = ctx.dot_section("lvl0");
+    ctx.task(lA.rw()).set_symbol("f1")->*[](cudaStream_t, auto) {};
+
+    ctx.push();
+    {
+      auto r1 = ctx.dot_section("lvl1");
+      ctx.task(lA.read(), lB.rw()).set_symbol("f2")->*[](cudaStream_t, auto, auto) {};
+      ctx.task(lA.read(), lC.rw()).set_symbol("f2")->*[](cudaStream_t, auto, auto) {};
+      ctx.task(lB.read(), lC.read(), lA.rw()).set_symbol("f3")->*[](cudaStream_t, auto, auto, auto) {};
+    }
+    ctx.pop();
+  }
+  ctx.finalize();
+
+#endif // !_CCCL_COMPILER(MSVC)
+}

cccl_upstream/cudax/test/stf/dot/with_events.cu

@@ -0,0 +1,49 @@
+//===----------------------------------------------------------------------===//
+//
+// 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 This test makes sure we can generate a dot file with events
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+int main()
+{
+// TODO (miscco): Make it work for windows
+#if !_CCCL_COMPILER(MSVC)
+  // Generate a random filename
+  int r = rand();
+
+  char filename[64];
+  snprintf(filename, 64, "output_%d.dot", r);
+  // fprintf(stderr, "filename %s\n", filename);
+  setenv("CUDASTF_DOT_FILE", filename, 1);
+  setenv("CUDASTF_DOT_IGNORE_PREREQS", "0", 1);
+
+  stream_ctx ctx;
+
+  auto lA = ctx.logical_data(shape_of<slice<char>>(64));
+  ctx.task(lA.write())->*[](cudaStream_t, auto) {};
+  ctx.task(lA.rw())->*[](cudaStream_t, auto) {};
+  ctx.finalize();
+
+  // Call this explicitly for the purpose of the test
+  reserved::dot::instance().finish();
+
+  // Make sure the file exists, and erase it
+  // fprintf(stderr, "ERASE. ...\n");
+  EXPECT(access(filename, F_OK) != -1);
+
+  EXPECT(unlink(filename) == 0);
+#endif // !_CCCL_COMPILER(MSVC)
+}