56fd68e7dd
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
104 lines
3.4 KiB
Plaintext
104 lines
3.4 KiB
Plaintext
// 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"});
|