[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/c/parallel.v2/src/reduce.cu

@@ -0,0 +1,167 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental 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) 2026 NVIDIA CORPORATION.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cstdio>
+#include <cstring>
+#include <filesystem>
+#include <string>
+
+#include <cccl/c/reduce.h>
+#include <hostjit/codegen/cub_call.hpp>
+#include <util/build_utils.h>
+
+using namespace hostjit::codegen;
+
+// (temp_storage, temp_bytes, d_in, d_out, num_items, op_state, init_state, stream)
+using reduce_fn_t = int (*)(void*, size_t*, void*, void*, unsigned long long, void*, void*, void*);
+
+CUresult cccl_device_reduce_build_ex(
+  cccl_device_reduce_build_result_t* build,
+  cccl_iterator_t input_it,
+  cccl_iterator_t output_it,
+  cccl_op_t op,
+  cccl_value_t init,
+  cccl_determinism_t determinism,
+  int cc_major,
+  int cc_minor,
+  const char* cub_path,
+  const char* thrust_path,
+  const char* libcudacxx_path,
+  const char* ctk_path,
+  cccl_build_config* build_config)
+try
+{
+  std::string cccl_include_str  = cccl::detail::parse_cccl_include_path(libcudacxx_path);
+  std::string ctk_root_str      = cccl::detail::parse_ctk_root(ctk_path);
+  const char* cccl_include_path = cccl_include_str.empty() ? nullptr : cccl_include_str.c_str();
+  const char* ctk_root          = ctk_root_str.empty() ? nullptr : ctk_root_str.c_str();
+  cccl::detail::MergedBuildConfig merged(build_config, cub_path, thrust_path);
+
+  auto result =
+    CubCall::from("cub/device/device_reduce.cuh")
+      .run("cub::DeviceReduce::Reduce")
+      .name("cccl_jit_reduce")
+      .with(temp_storage, temp_bytes, in(input_it), out(output_it), num_items, op, init, stream)
+      .compile(cc_major, cc_minor, merged.get(), ctk_root, cccl_include_path);
+
+  build->cc = cc_major * 10 + cc_minor;
+  cccl::detail::copy_cubin(result.cubin, build->payload, build->payload_size);
+  build->jit_compiler     = result.compiler;
+  build->reduce_fn        = reinterpret_cast<void*>(result.fn_ptr);
+  build->accumulator_size = init.type.size;
+  build->determinism      = determinism;
+
+  return CUDA_SUCCESS;
+}
+catch (const std::exception& exc)
+{
+  fprintf(stderr, "\nEXCEPTION in cccl_device_reduce_build(): %s\n", exc.what());
+  return CUDA_ERROR_UNKNOWN;
+}
+
+CUresult cccl_device_reduce(
+  cccl_device_reduce_build_result_t build,
+  void* d_temp_storage,
+  size_t* temp_storage_bytes,
+  cccl_iterator_t d_in,
+  cccl_iterator_t d_out,
+  uint64_t num_items,
+  cccl_op_t op,
+  cccl_value_t init,
+  CUstream stream)
+try
+{
+  if (!build.reduce_fn)
+  {
+    return CUDA_ERROR_INVALID_VALUE;
+  }
+  auto reduce_fn = reinterpret_cast<reduce_fn_t>(build.reduce_fn);
+
+  // Parameter order matches CubCall::with() order: ..., num_items, op.state, init.state, stream
+  const int status = reduce_fn(
+    d_temp_storage,
+    temp_storage_bytes,
+    d_in.state,
+    d_out.state,
+    num_items,
+    op.state,
+    init.state,
+    reinterpret_cast<void*>(stream));
+
+  return (status == 0) ? CUDA_SUCCESS : CUDA_ERROR_UNKNOWN;
+}
+catch (const std::exception& exc)
+{
+  fprintf(stderr, "\nEXCEPTION in cccl_device_reduce(): %s\n", exc.what());
+  return CUDA_ERROR_UNKNOWN;
+}
+
+CUresult cccl_device_reduce_nondeterministic(
+  cccl_device_reduce_build_result_t build,
+  void* d_temp_storage,
+  size_t* temp_storage_bytes,
+  cccl_iterator_t d_in,
+  cccl_iterator_t d_out,
+  uint64_t num_items,
+  cccl_op_t op,
+  cccl_value_t init,
+  CUstream stream)
+{
+  return cccl_device_reduce(build, d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, op, init, stream);
+}
+
+CUresult cccl_device_reduce_cleanup(cccl_device_reduce_build_result_t* build_ptr)
+try
+{
+  if (build_ptr == nullptr)
+  {
+    return CUDA_ERROR_INVALID_VALUE;
+  }
+
+  cccl::detail::release_jit_artifacts(build_ptr);
+  build_ptr->reduce_fn = nullptr;
+
+  return CUDA_SUCCESS;
+}
+catch (const std::exception& exc)
+{
+  fprintf(stderr, "\nEXCEPTION in cccl_device_reduce_cleanup(): %s\n", exc.what());
+  return CUDA_ERROR_UNKNOWN;
+}
+
+CUresult cccl_device_reduce_build(
+  cccl_device_reduce_build_result_t* build,
+  cccl_iterator_t d_in,
+  cccl_iterator_t d_out,
+  cccl_op_t op,
+  cccl_value_t init,
+  cccl_determinism_t determinism,
+  int cc_major,
+  int cc_minor,
+  const char* cub_path,
+  const char* thrust_path,
+  const char* libcudacxx_path,
+  const char* ctk_path)
+{
+  return cccl_device_reduce_build_ex(
+    build,
+    d_in,
+    d_out,
+    op,
+    init,
+    determinism,
+    cc_major,
+    cc_minor,
+    cub_path,
+    thrust_path,
+    libcudacxx_path,
+    ctk_path,
+    nullptr);
+}