[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/cub/benchmarks/bench/merge/keys.cu

@@ -0,0 +1,78 @@
+// SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include <cub/device/device_merge.cuh>
+
+#include <thrust/detail/raw_pointer_cast.h>
+
+#include <cuda/std/utility>
+
+#include <cstdint>
+
+#include <nvbench_helper.cuh>
+
+#include "merge_common.cuh"
+
+// %RANGE% TUNE_TRANSPOSE trp 0:1:1
+// %RANGE% TUNE_LOAD ld 0:3:1
+// %RANGE% TUNE_ITEMS_PER_THREAD ipt 7:24:1
+// %RANGE% TUNE_THREADS_PER_BLOCK_POW2 tpb 6:10:1
+
+template <typename KeyT>
+void keys(nvbench::state& state, nvbench::type_list<KeyT>)
+{
+  using offset_t     = int64_t;
+  using compare_op_t = less_t;
+
+  // Retrieve axis parameters
+  const auto elements       = static_cast<std::size_t>(state.get_int64("Elements{io}"));
+  const bit_entropy entropy = str_to_entropy(state.get_string("Entropy"));
+
+  const auto num_items_lhs  = elements / 2;
+  const auto num_items_rhs  = elements - num_items_lhs;
+  auto [keys_lhs, keys_rhs] = generate_lhs_rhs<KeyT>(num_items_lhs, num_items_rhs, entropy);
+
+  thrust::device_vector<KeyT> keys_out(elements, thrust::no_init);
+  KeyT* d_keys_lhs = thrust::raw_pointer_cast(keys_lhs.data());
+  KeyT* d_keys_rhs = thrust::raw_pointer_cast(keys_rhs.data());
+  KeyT* d_keys_out = thrust::raw_pointer_cast(keys_out.data());
+
+  // Enable throughput calculations and add "Size" column to results.
+  state.add_element_count(elements);
+  state.add_global_memory_reads<KeyT>(elements);
+  state.add_global_memory_writes<KeyT>(elements);
+
+  caching_allocator_t alloc;
+  state.exec(nvbench::exec_tag::gpu | nvbench::exec_tag::no_batch, [&](nvbench::launch& launch) {
+    auto env = cub_bench_env(
+      alloc,
+      launch
+#if !TUNE_BASE
+      ,
+      cuda::execution::tune(bench_policy_selector<key_t>{})
+#endif // !TUNE_BASE
+    );
+    _CCCL_TRY_CUDA_API(
+      cub::DeviceMerge::MergeKeys,
+      "MergePairs failed",
+      d_keys_lhs,
+      static_cast<offset_t>(num_items_lhs),
+      d_keys_rhs,
+      static_cast<offset_t>(num_items_rhs),
+      d_keys_out,
+      compare_op_t{},
+      env);
+  });
+}
+
+#ifdef TUNE_KeyT
+using key_types = nvbench::type_list<TUNE_KeyT>;
+#else // !defined(TUNE_KeyT)
+using key_types = fundamental_types;
+#endif // TUNE_KeyT
+
+NVBENCH_BENCH_TYPES(keys, NVBENCH_TYPE_AXES(key_types))
+  .set_name("base")
+  .set_type_axes_names({"KeyT{ct}"})
+  .add_int64_power_of_two_axis("Elements{io}", nvbench::range(16, 28, 4))
+  .add_string_axis("Entropy", {"1.000", "0.201"});

cccl_upstream/cub/benchmarks/bench/merge/merge_common.cuh

@@ -0,0 +1,135 @@
+// SPDX-FileCopyrightText: Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3-Clause
+
+#pragma once
+
+#include <thrust/copy.h>
+#include <thrust/count.h>
+#include <thrust/device_vector.h>
+#include <thrust/sort.h>
+
+#include <cuda/iterator>
+
+#include <nvbench_helper.cuh>
+
+#if !TUNE_BASE
+#  if TUNE_LOAD == 0
+#    define TUNE_LOAD_MODIFIER cub::LOAD_DEFAULT
+#    define TUNE_USE_BL2SH     false
+#  elif TUNE_LOAD == 1
+#    define TUNE_LOAD_MODIFIER cub::LOAD_LDG
+#    define TUNE_USE_BL2SH     false
+#  elif TUNE_LOAD == 2
+#    define TUNE_LOAD_MODIFIER cub::LOAD_CA
+#    define TUNE_USE_BL2SH     false
+#  else // TUNE_LOAD == 3
+#    define TUNE_LOAD_MODIFIER cub::LOAD_DEFAULT
+#    define TUNE_USE_BL2SH     true
+#  endif // TUNE_LOAD
+
+template <typename KeyT>
+struct bench_policy_selector
+{
+  [[nodiscard]] _CCCL_HOST_DEVICE constexpr auto operator()(cuda::compute_capability) const -> cub::MergePolicy
+  {
+    return cub::MergePolicy{
+      (1 << TUNE_THREADS_PER_BLOCK_POW2),
+      cub::Nominal4BItemsToItems<KeyT>(TUNE_ITEMS_PER_THREAD),
+      TUNE_LOAD_MODIFIER,
+      TUNE_TRANSPOSE == 0 ? cub::BLOCK_STORE_DIRECT : cub::BLOCK_STORE_WARP_TRANSPOSE,
+      TUNE_USE_BL2SH};
+  }
+};
+#endif // TUNE_BASE
+
+struct select_if_less_than_t
+{
+  bool negate;
+  uint8_t threshold;
+
+  __device__ __forceinline__ bool operator()(uint8_t val) const
+  {
+    return negate ? !(val < threshold) : val < threshold;
+  }
+};
+
+template <typename OffsetT>
+struct write_pivot_point_t
+{
+  OffsetT threshold;
+  OffsetT* pivot_point;
+
+  __device__ void operator()(OffsetT output_index, OffsetT input_index) const
+  {
+    if (output_index == threshold)
+    {
+      *pivot_point = input_index;
+    }
+  }
+};
+
+template <typename KeyT>
+std::pair<thrust::device_vector<KeyT>, thrust::device_vector<KeyT>>
+generate_lhs_rhs(std::size_t num_items_lhs, std::size_t num_items_rhs, bit_entropy entropy)
+{
+  using offset_t = std::size_t;
+
+  const auto elements = num_items_lhs + num_items_rhs;
+
+  // We generate data distributions in the range [0, 255], which, with lower entropy, get skewed towards 0.
+  // We use this to generate increasingly large *consecutive* segments of data that are getting selected from the lhs
+  thrust::device_vector<uint8_t> rnd_selector_val = generate(elements, entropy);
+  uint8_t threshold                               = 128;
+  select_if_less_than_t select_lhs_op{false, threshold};
+  select_if_less_than_t select_rhs_op{true, threshold};
+
+  // The following algorithm only works under the precondition that there's at least 50% of the data in the lhs
+  // If that's not the case, we simply swap the logic for selecting into lhs and rhs
+  const auto num_items_selected_into_lhs =
+    static_cast<offset_t>(thrust::count_if(rnd_selector_val.begin(), rnd_selector_val.end(), select_lhs_op));
+  if (num_items_selected_into_lhs < num_items_lhs)
+  {
+    using ::cuda::std::swap;
+    swap(select_lhs_op, select_rhs_op);
+  }
+
+  // We want lhs and rhs to be of equal size. We also want to have skewed distributions, such that we put different
+  // workloads on the binary search part. For this reason, we identify the index from the input, referred to as pivot
+  // point, after which the lhs is "full". We compose the rhs by selecting all items up to the pivot point that were not
+  // selected for lhs and *all* items after the pivot point.
+  constexpr std::size_t num_pivot_points = 1;
+  thrust::device_vector<offset_t> pivot_point(num_pivot_points);
+  auto counting_it            = thrust::make_counting_iterator(offset_t{0});
+  using counting_difference_t = typename decltype(counting_it)::difference_type;
+  thrust::copy_if(
+    counting_it,
+    counting_it + static_cast<counting_difference_t>(elements),
+    rnd_selector_val.begin(),
+    cuda::make_tabulate_output_iterator(write_pivot_point_t<offset_t>{
+      static_cast<offset_t>(num_items_lhs), thrust::raw_pointer_cast(pivot_point.data())}),
+    select_lhs_op);
+
+  thrust::device_vector<KeyT> keys_lhs(num_items_lhs);
+  thrust::device_vector<KeyT> keys_rhs(num_items_rhs);
+
+  thrust::device_vector<KeyT> increasing_input = generate(elements);
+  thrust::sort(increasing_input.begin(), increasing_input.end());
+
+  offset_t pivot_point_val = pivot_point[0];
+  auto const end_lhs       = thrust::copy_if(
+    increasing_input.cbegin(),
+    increasing_input.cbegin() + pivot_point_val,
+    rnd_selector_val.cbegin(),
+    keys_lhs.begin(),
+    select_lhs_op);
+
+  auto const end_rhs = thrust::copy_if(
+    increasing_input.cbegin(),
+    increasing_input.cbegin() + pivot_point_val,
+    rnd_selector_val.cbegin(),
+    keys_rhs.begin(),
+    select_rhs_op);
+  thrust::copy(increasing_input.cbegin() + pivot_point_val, increasing_input.cbegin() + elements, end_rhs);
+
+  return {keys_lhs, keys_rhs};
+}

cccl_upstream/cub/benchmarks/bench/merge/pairs.cu

@@ -0,0 +1,103 @@
+// SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3-Clause
+
+#include <cub/device/device_merge.cuh>
+
+#include <thrust/detail/raw_pointer_cast.h>
+
+#include <cuda/std/utility>
+
+#include <cstdint>
+
+#include <nvbench_helper.cuh>
+
+#include "merge_common.cuh"
+
+// %RANGE% TUNE_TRANSPOSE trp 0:1:1
+// %RANGE% TUNE_LOAD ld 0:3:1
+// %RANGE% TUNE_ITEMS_PER_THREAD ipt 7:24:1
+// %RANGE% TUNE_THREADS_PER_BLOCK_POW2 tpb 6:10:1
+
+template <typename KeyT, typename ValueT>
+void pairs(nvbench::state& state, nvbench::type_list<KeyT, ValueT>)
+{
+  using offset_t     = int64_t;
+  using compare_op_t = less_t;
+
+  // Retrieve axis parameters
+  const auto elements       = static_cast<std::size_t>(state.get_int64("Elements{io}"));
+  const bit_entropy entropy = str_to_entropy(state.get_string("Entropy"));
+
+  const auto num_items_lhs = elements / 2;
+  const auto num_items_rhs = elements - num_items_lhs;
+
+  thrust::device_vector<KeyT> keys_out(elements, thrust::no_init);
+  thrust::device_vector<ValueT> values_lhs(num_items_lhs, thrust::no_init);
+  thrust::device_vector<ValueT> values_rhs(num_items_rhs, thrust::no_init);
+  thrust::device_vector<ValueT> values_out(elements, thrust::no_init);
+
+  auto [keys_lhs, keys_rhs] = generate_lhs_rhs<KeyT>(num_items_lhs, num_items_rhs, entropy);
+
+  KeyT* d_keys_lhs     = thrust::raw_pointer_cast(keys_lhs.data());
+  KeyT* d_keys_rhs     = thrust::raw_pointer_cast(keys_rhs.data());
+  KeyT* d_keys_out     = thrust::raw_pointer_cast(keys_out.data());
+  ValueT* d_values_lhs = thrust::raw_pointer_cast(values_lhs.data());
+  ValueT* d_values_rhs = thrust::raw_pointer_cast(values_rhs.data());
+  ValueT* d_values_out = thrust::raw_pointer_cast(values_out.data());
+
+  // Enable throughput calculations and add "Size" column to results.
+  state.add_element_count(elements);
+  state.add_global_memory_reads<KeyT>(elements);
+  state.add_global_memory_reads<ValueT>(elements);
+  state.add_global_memory_writes<KeyT>(elements);
+  state.add_global_memory_writes<ValueT>(elements);
+
+  caching_allocator_t alloc;
+  state.exec(nvbench::exec_tag::gpu | nvbench::exec_tag::no_batch, [&](nvbench::launch& launch) {
+    auto env = cub_bench_env(
+      alloc,
+      launch
+#if !TUNE_BASE
+      ,
+      cuda::execution::tune(bench_policy_selector<key_t>{})
+#endif // !TUNE_BASE
+    );
+    _CCCL_TRY_CUDA_API(
+      cub::DeviceMerge::MergePairs,
+      "MergePairs failed",
+      d_keys_lhs,
+      d_values_lhs,
+      static_cast<offset_t>(num_items_lhs),
+      d_keys_rhs,
+      d_values_rhs,
+      static_cast<offset_t>(num_items_rhs),
+      d_keys_out,
+      d_values_out,
+      compare_op_t{},
+      env);
+  });
+}
+
+#ifdef TUNE_KeyT
+using key_types = nvbench::type_list<TUNE_KeyT>;
+#else // !defined(TUNE_KeyT)
+using key_types = fundamental_types;
+#endif // TUNE_KeyT
+
+#ifdef TUNE_ValueT
+using value_types = nvbench::type_list<TUNE_ValueT>;
+#else // !defined(TUNE_ValueT)
+using value_types = nvbench::type_list<int8_t, int16_t, int32_t, int64_t
+#  if _CCCL_HAS_INT128()
+// nvcc currently hangs for __int128 value type with the fallback policy of {CTA: 64, IPT: 1}. NVBug 4384075
+//  ,
+//  int128_t
+#  endif
+                                       >;
+#endif // TUNE_ValueT
+
+NVBENCH_BENCH_TYPES(pairs, NVBENCH_TYPE_AXES(key_types, value_types))
+  .set_name("base")
+  .set_type_axes_names({"KeyT{ct}", "ValueT{ct}"})
+  .add_int64_power_of_two_axis("Elements{io}", nvbench::range(16, 28, 4))
+  .add_string_axis("Entropy", {"1.000", "0.201"});