[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/libcudacxx/test/nvtarget/CMakeLists.txt

@@ -0,0 +1,5 @@
+add_custom_target(libcudacxx.test.nvtarget)
+
+add_subdirectory(arch_specific)
+add_subdirectory(dialect)
+add_subdirectory(family_specific)

cccl_upstream/libcudacxx/test/nvtarget/arch_specific/CMakeLists.txt

@@ -0,0 +1,20 @@
+if (NOT "${CMAKE_CUDA_COMPILER_ID}" STREQUAL "NVIDIA")
+  return()
+endif()
+
+# arch-specific features are supported since 12.9
+if ("${CMAKE_CUDA_COMPILER_VERSION}" VERSION_LESS "12.9")
+  return()
+endif()
+
+set_directory_properties(
+  PROPERTIES INCLUDE_DIRECTORIES "${libcudacxx_SOURCE_DIR}/include"
+)
+
+set(target_name "libcudacxx.test.nvtarget.arch_specific")
+add_library(${target_name} OBJECT arch_specific.cu)
+set_target_properties(
+  ${target_name}
+  PROPERTIES CUDA_ARCHITECTURES "103a-virtual"
+)
+add_dependencies(libcudacxx.test.nvtarget ${target_name})

cccl_upstream/libcudacxx/test/nvtarget/arch_specific/arch_specific.cu

@@ -0,0 +1,63 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// 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 & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+// This test checks if arch-specific NV target macros work properly.
+
+#include <nv/target>
+
+// Currently, nvcc is the only compiler that supports arch-specific architectures.
+#if !defined(__NVCC__)
+#  error "This test works with nvcc only."
+#endif // !__NVCC__
+
+#if defined(__CUDA_ARCH__)
+#  if __CUDA_ARCH_SPECIFIC__ != 1030
+#    error "This test must be compiled for sm_103a target."
+#  endif // __CUDA_ARCH_SPECIFIC__ != 1030
+#endif // __CUDA_ARCH__
+
+#define CHECK_TRUE(_PRED)                                                 \
+  do                                                                      \
+  {                                                                       \
+    NV_IF_ELSE_TARGET(_PRED, static_assert(true);, static_assert(false);) \
+  } while (0)
+#define CHECK_FALSE(_PRED)                                                \
+  do                                                                      \
+  {                                                                       \
+    NV_IF_ELSE_TARGET(_PRED, static_assert(false);, static_assert(true);) \
+  } while (0)
+
+#ifdef __CUDACC_TILE__
+__tile__
+#endif // __CUDACC_TILE__
+  __host__ __device__ void
+  fn()
+{
+#if defined(__CUDA_ARCH__)
+  CHECK_TRUE(NV_IS_EXACTLY_SM_103);
+
+  CHECK_TRUE(NV_HAS_FEATURE_SM_103a);
+
+  CHECK_TRUE(NV_HAS_FEATURE_SM_100f);
+  CHECK_TRUE(NV_HAS_FEATURE_SM_103f);
+#else // ^^^ __CUDA_ARCH__ ^^^ / vvv !__CUDA_ARCH__ vvv
+  CHECK_TRUE(NV_IS_HOST);
+
+  CHECK_FALSE(NV_HAS_FEATURE_SM_103a);
+
+  CHECK_FALSE(NV_HAS_FEATURE_SM_100f);
+  CHECK_FALSE(NV_HAS_FEATURE_SM_103f);
+#endif // ^^^ !__CUDA_ARCH__ ^^^
+
+  CHECK_FALSE(NV_HAS_FEATURE_SM_100a);
+  CHECK_FALSE(NV_HAS_FEATURE_SM_110a);
+
+  CHECK_FALSE(NV_HAS_FEATURE_SM_110f);
+}

cccl_upstream/libcudacxx/test/nvtarget/dialect/CMakeLists.txt

@@ -0,0 +1,47 @@
+# This unit just checks that nv/target can be used with the specified dialects.
+
+set(nvtarget_dialects 11 14 17 20)
+
+# cudafe++ 12.9 has problem compiling some parts of <type_traits> when using libstdc++ 14, so let's just filter out
+# dialect version 11.
+if (
+  "${CMAKE_CUDA_COMPILER_ID}" STREQUAL "NVIDIA"
+  AND "${CMAKE_CUDA_COMPILER_VERSION}" VERSION_GREATER_EQUAL "12.9"
+  AND "${CMAKE_CUDA_COMPILER_VERSION}" VERSION_LESS "13.0"
+  AND "${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU"
+  AND "${CMAKE_CXX_COMPILER_VERSION}" VERSION_GREATER_EQUAL "14.0"
+  AND "${CMAKE_CXX_COMPILER_VERSION}" VERSION_LESS "15.0"
+)
+  list(FILTER nvtarget_dialects EXCLUDE REGEX "^11$")
+endif()
+
+set_directory_properties(
+  PROPERTIES INCLUDE_DIRECTORIES "${libcudacxx_SOURCE_DIR}/include"
+)
+
+foreach (dialect ${nvtarget_dialects})
+  if ("cxx_std_${dialect}" IN_LIST CMAKE_CXX_COMPILE_FEATURES)
+    set(target_name "libcudacxx.test.nvtarget.dialect.cpp${dialect}")
+    add_library(${target_name} OBJECT nvtargettest.cpp)
+    # This is required to override CMAKE_CXX_STANDARD - target_compile_features is insufficient.
+    set_target_properties(
+      ${target_name}
+      PROPERTIES CXX_STANDARD "${dialect}" CXX_STANDARD_REQUIRED ON
+    )
+    add_dependencies(libcudacxx.test.nvtarget ${target_name})
+  endif()
+
+  if (
+    "cuda_std_${dialect}" IN_LIST CMAKE_CUDA_COMPILE_FEATURES
+    AND "cxx_std_${dialect}" IN_LIST CMAKE_CXX_COMPILE_FEATURES
+  )
+    set(target_name "libcudacxx.test.nvtarget.dialect.cuda${dialect}")
+    add_library(${target_name} OBJECT nvtargettest.cu)
+    # This is required to override CMAKE_CXX_STANDARD - target_compile_features is insufficient.
+    set_target_properties(
+      ${target_name}
+      PROPERTIES CUDA_STANDARD "${dialect}" CUDA_STANDARD_REQUIRED ON
+    )
+    add_dependencies(libcudacxx.test.nvtarget ${target_name})
+  endif()
+endforeach()

cccl_upstream/libcudacxx/test/nvtarget/dialect/nvtargettest.cpp

@@ -0,0 +1,13 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// 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 & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <nv/target>
+
+void fn() {}

cccl_upstream/libcudacxx/test/nvtarget/dialect/nvtargettest.cu

@@ -0,0 +1,18 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// 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 & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <nv/target>
+
+#ifdef __CUDACC_TILE__
+__tile__
+#endif // __CUDACC_TILE__
+  __host__ __device__ void
+  fn()
+{}

cccl_upstream/libcudacxx/test/nvtarget/family_specific/CMakeLists.txt

@@ -0,0 +1,20 @@
+if (NOT "${CMAKE_CUDA_COMPILER_ID}" STREQUAL "NVIDIA")
+  return()
+endif()
+
+# family-specific features are supported since 12.9
+if ("${CMAKE_CUDA_COMPILER_VERSION}" VERSION_LESS "12.9")
+  return()
+endif()
+
+set_directory_properties(
+  PROPERTIES INCLUDE_DIRECTORIES "${libcudacxx_SOURCE_DIR}/include"
+)
+
+set(target_name "libcudacxx.test.nvtarget.family_specific")
+add_library(${target_name} OBJECT family_specific.cu)
+set_target_properties(
+  ${target_name}
+  PROPERTIES CUDA_ARCHITECTURES "103f-virtual"
+)
+add_dependencies(libcudacxx.test.nvtarget ${target_name})

cccl_upstream/libcudacxx/test/nvtarget/family_specific/family_specific.cu

@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// 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 & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+// This test checks if family-specific NV target macros work properly.
+
+#include <nv/target>
+
+// Currently, nvcc is the only compiler that supports arch-specific architectures.
+#if !defined(__NVCC__)
+#  error "This test works with nvcc only."
+#endif // !__NVCC__
+
+#if defined(__CUDA_ARCH__)
+#  if __CUDA_ARCH_FAMILY_SPECIFIC__ != 1030
+#    error "This test must be compiled for sm_103f target."
+#  endif // __CUDA_ARCH_FAMILY_SPECIFIC__ != 1030
+#endif // __CUDA_ARCH__
+
+#define CHECK_TRUE(_PRED)                                                 \
+  do                                                                      \
+  {                                                                       \
+    NV_IF_ELSE_TARGET(_PRED, static_assert(true);, static_assert(false);) \
+  } while (0)
+#define CHECK_FALSE(_PRED)                                                \
+  do                                                                      \
+  {                                                                       \
+    NV_IF_ELSE_TARGET(_PRED, static_assert(false);, static_assert(true);) \
+  } while (0)
+
+#ifdef __CUDACC_TILE__
+__tile__
+#endif // __CUDACC_TILE__
+  __host__ __device__ void
+  fn()
+{
+#if defined(__CUDA_ARCH__)
+  CHECK_TRUE(NV_IS_EXACTLY_SM_103);
+
+  CHECK_TRUE(NV_HAS_FEATURE_SM_100f);
+  CHECK_TRUE(NV_HAS_FEATURE_SM_103f);
+#else // ^^^ __CUDA_ARCH__ ^^^ / vvv !__CUDA_ARCH__ vvv
+  CHECK_TRUE(NV_IS_HOST);
+
+  CHECK_FALSE(NV_HAS_FEATURE_SM_100f);
+  CHECK_FALSE(NV_HAS_FEATURE_SM_103f);
+#endif // ^^^ !__CUDA_ARCH__ ^^^
+
+  CHECK_FALSE(NV_HAS_FEATURE_SM_110f);
+
+  CHECK_FALSE(NV_HAS_FEATURE_SM_100a);
+  CHECK_FALSE(NV_HAS_FEATURE_SM_103a);
+  CHECK_FALSE(NV_HAS_FEATURE_SM_110a);
+}