[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/debugging/buffer/CMakeLists.txt

@@ -0,0 +1,15 @@
+libcudacxx_add_pretty_printer_test(
+  NAME buffer
+  SOURCES source.cu
+  CASES
+    # gersemi: off
+    --case inspect_normal 1 buffer.normal normal_values
+    --case inspect_alias 1 buffer.alias aliased_values
+    --case inspect_vector 1 buffer.vector.0 "buffer_vector[0]"
+    --case inspect_vector 1 buffer.vector.1 "buffer_vector[1]"
+    --case inspect_host_device 1 buffer.host_device host_device_values
+    --case inspect_empty 1 buffer.empty empty_values
+    --case inspect_before_update 1 buffer.update.before updated_values
+    --case inspect_after_update 1 buffer.update.after updated_values
+    # gersemi: on
+)

cccl_upstream/libcudacxx/test/debugging/buffer/gdb.expected

@@ -0,0 +1,63 @@
+=============== buffer.normal begin ===============
+cuda::buffer<int, cuda::mr::device_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=10, align=4, data=<address> (device) = {
+  [0] = -56,
+  [1] = 22,
+  [2] = 94,
+  [3] = -13,
+  [4] = 7,
+  [5] = 41,
+  [6] = -82,
+  [7] = 0,
+  [8] = 63,
+  [9] = -5
+}
+=============== buffer.normal end ===============
+=============== buffer.alias begin ===============
+cuda::buffer<int, cuda::mr::device_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (device) = {
+  [0] = 17,
+  [1] = -31,
+  [2] = 8,
+  [3] = 55
+}
+=============== buffer.alias end ===============
+=============== buffer.vector.0 begin ===============
+cuda::buffer<int, cuda::mr::device_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=3, align=4, data=<address> (device) = {
+  [0] = -2,
+  [1] = 4,
+  [2] = 6
+}
+=============== buffer.vector.0 end ===============
+=============== buffer.vector.1 begin ===============
+cuda::buffer<int, cuda::mr::device_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=3, align=4, data=<address> (device) = {
+  [0] = 11,
+  [1] = -9,
+  [2] = 27
+}
+=============== buffer.vector.1 end ===============
+=============== buffer.host_device begin ===============
+cuda::buffer<int, cuda::mr::device_accessible, cuda::mr::host_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible, cuda::mr::host_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (host/device) = {
+  [0] = 3,
+  [1] = 14,
+  [2] = -15,
+  [3] = 92
+}
+=============== buffer.host_device end ===============
+=============== buffer.empty begin ===============
+cuda::buffer<int, cuda::mr::device_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=0, align=4, data=0x0 (device)
+=============== buffer.empty end ===============
+=============== buffer.update.before begin ===============
+cuda::buffer<int, cuda::mr::device_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (device) = {
+  [0] = 1,
+  [1] = 2,
+  [2] = 3,
+  [3] = 4
+}
+=============== buffer.update.before end ===============
+=============== buffer.update.after begin ===============
+cuda::buffer<int, cuda::mr::device_accessible> mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (device) = {
+  [0] = -8,
+  [1] = 13,
+  [2] = 21,
+  [3] = -34
+}
+=============== buffer.update.after end ===============

cccl_upstream/libcudacxx/test/debugging/buffer/lldb.expected

@@ -0,0 +1,63 @@
+=============== buffer.normal begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=10, align=4, data=<address> (device) {
+  [0] = -56
+  [1] = 22
+  [2] = 94
+  [3] = -13
+  [4] = 7
+  [5] = 41
+  [6] = -82
+  [7] = 0
+  [8] = 63
+  [9] = -5
+}
+=============== buffer.normal end ===============
+=============== buffer.alias begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (device) {
+  [0] = 17
+  [1] = -31
+  [2] = 8
+  [3] = 55
+}
+=============== buffer.alias end ===============
+=============== buffer.vector.0 begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=3, align=4, data=<address> (device) {
+  [0] = -2
+  [1] = 4
+  [2] = 6
+}
+=============== buffer.vector.0 end ===============
+=============== buffer.vector.1 begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=3, align=4, data=<address> (device) {
+  [0] = 11
+  [1] = -9
+  [2] = 27
+}
+=============== buffer.vector.1 end ===============
+=============== buffer.host_device begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible, cuda::mr::host_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible, cuda::mr::host_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (host/device) {
+  [0] = 3
+  [1] = 14
+  [2] = -15
+  [3] = 92
+}
+=============== buffer.host_device end ===============
+=============== buffer.empty begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=0, align=4, data=0x0 (device)
+=============== buffer.empty end ===============
+=============== buffer.update.before begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (device) {
+  [0] = 1
+  [1] = 2
+  [2] = 3
+  [3] = 4
+}
+=============== buffer.update.before end ===============
+=============== buffer.update.after begin ===============
+(cuda::buffer<int, cuda::mr::device_accessible>) mr=cuda::mr::any_resource<cuda::mr::device_accessible> @ <address>, stream=<address>, size=4, align=4, data=<address> (device) {
+  [0] = -8
+  [1] = 13
+  [2] = 21
+  [3] = -34
+}
+=============== buffer.update.after end ===============

cccl_upstream/libcudacxx/test/debugging/buffer/source.cu

@@ -0,0 +1,102 @@
+// Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+//
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include <cuda/buffer>
+#include <cuda/memory_resource>
+#include <cuda/std/array>
+#include <cuda/stream>
+
+#include <vector>
+
+#include <cuda_runtime_api.h>
+
+template <class T>
+[[gnu::noinline]] void keep_for_debugger(const T& value)
+{
+  asm volatile("" : : "g"(&value) : "memory");
+}
+
+[[gnu::noinline]] void inspect_normal(const cuda::device_buffer<int>& values)
+{
+  keep_for_debugger(values);
+}
+
+using device_buffer_alias = cuda::buffer<int, cuda::mr::device_accessible>;
+
+[[gnu::noinline]] void inspect_alias(const device_buffer_alias& values)
+{
+  keep_for_debugger(values);
+}
+
+[[gnu::noinline]] void inspect_vector(const std::vector<cuda::device_buffer<int>>& values)
+{
+  keep_for_debugger(values[0]);
+  keep_for_debugger(values[1]);
+}
+
+template <class Buffer>
+[[gnu::noinline]] void inspect_host_device(const Buffer& values)
+{
+  keep_for_debugger(values);
+}
+
+[[gnu::noinline]] void inspect_empty(const cuda::device_buffer<int>& values)
+{
+  keep_for_debugger(values);
+}
+
+[[gnu::noinline]] void inspect_before_update(const cuda::device_buffer<int>& values)
+{
+  keep_for_debugger(values);
+}
+
+[[gnu::noinline]] void inspect_after_update(const cuda::device_buffer<int>& values)
+{
+  keep_for_debugger(values);
+}
+
+int main()
+{
+  constexpr cuda::device_ref device{0};
+  cuda::stream stream{device};
+
+  const cuda::std::array normal_host_values{-56, 22, 94, -13, 7, 41, -82, 0, 63, -5};
+  const auto normal_values = cuda::make_device_buffer<int>(stream, device, normal_host_values);
+
+  const cuda::std::array alias_host_values{17, -31, 8, 55};
+  const device_buffer_alias aliased_values = cuda::make_device_buffer<int>(stream, device, alias_host_values);
+
+  std::vector<cuda::device_buffer<int>> buffer_vector;
+  buffer_vector.emplace_back(cuda::make_device_buffer<int>(stream, device, cuda::std::array{-2, 4, 6}));
+  buffer_vector.emplace_back(cuda::make_device_buffer<int>(stream, device, cuda::std::array{11, -9, 27}));
+
+  cuda::mr::legacy_managed_memory_resource managed_resource;
+  const cuda::std::array host_device_host_values{3, 14, -15, 92};
+  const auto host_device_values = cuda::make_buffer<int>(stream, managed_resource, host_device_host_values);
+  const auto empty_values       = cuda::make_device_buffer<int>(stream, device);
+
+  const cuda::std::array initial_updated_host_values{1, 2, 3, 4};
+  auto updated_values = cuda::make_device_buffer<int>(stream, device, initial_updated_host_values);
+
+  stream.sync();
+  inspect_normal(normal_values);
+  inspect_alias(aliased_values);
+  inspect_vector(buffer_vector);
+  inspect_host_device(host_device_values);
+  inspect_empty(empty_values);
+  inspect_before_update(updated_values);
+
+  const cuda::std::array replacement_host_values{-8, 13, 21, -34};
+  if (cudaMemcpyAsync(updated_values.data(),
+                      replacement_host_values.data(),
+                      replacement_host_values.size() * sizeof(*updated_values.data()),
+                      cudaMemcpyDefault,
+                      stream.get())
+      != cudaSuccess)
+  {
+    return 1;
+  }
+  stream.sync();
+  inspect_after_update(updated_values);
+}