1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
|
// MIT License
//
// Copyright (c) 2022-2024 Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in all
// copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
// SOFTWARE.
#include "benchmark_utils.hpp"
// CmdParser
#include "cmdparser.hpp"
// Google Benchmark
#include <benchmark/benchmark.h>
// HIP API
#include <hip/hip_runtime.h>
// rocPRIM
#include <rocprim/block/block_load_func.hpp>
#include <rocprim/block/block_radix_rank.hpp>
#include <rocprim/block/block_store_func.hpp>
#include <chrono>
#include <limits>
#include <string>
#ifndef DEFAULT_N
const size_t DEFAULT_BYTES = 1024 * 1024 * 128 * 4;
#endif
namespace rp = rocprim;
template<typename T,
unsigned int BlockSize,
unsigned int ItemsPerThread,
unsigned int RadixBits,
bool Descending,
rp::block_radix_rank_algorithm Algorithm,
unsigned int Trials>
__global__ __launch_bounds__(BlockSize) void rank_kernel(const T* keys_input,
unsigned int* ranks_output)
{
using rank_type = rp::block_radix_rank<BlockSize, RadixBits, Algorithm>;
const unsigned int lid = threadIdx.x;
const unsigned int block_offset = blockIdx.x * ItemsPerThread * BlockSize;
T keys[ItemsPerThread];
rp::block_load_direct_striped<BlockSize>(lid, keys_input + block_offset, keys);
unsigned int ranks[ItemsPerThread];
ROCPRIM_NO_UNROLL
for(unsigned int trial = 0; trial < Trials; ++trial)
{
ROCPRIM_SHARED_MEMORY typename rank_type::storage_type storage;
unsigned int begin_bit = 0;
const unsigned int end_bit = sizeof(T) * 8;
while(begin_bit < end_bit)
{
const unsigned pass_bits = min(RadixBits, end_bit - begin_bit);
if ROCPRIM_IF_CONSTEXPR(Descending)
{
rank_type().rank_keys_desc(keys, ranks, storage, begin_bit, pass_bits);
}
else
{
rank_type().rank_keys(keys, ranks, storage, begin_bit, pass_bits);
}
begin_bit += RadixBits;
}
}
rp::block_store_direct_striped<BlockSize>(lid, ranks_output + block_offset, ranks);
}
template<typename T,
unsigned int BlockSize,
unsigned int ItemsPerThread,
rp::block_radix_rank_algorithm Algorithm,
unsigned int RadixBits = 4,
bool Descending = false,
unsigned int Trials = 10>
void run_benchmark(benchmark::State& state, size_t bytes, const managed_seed& seed, hipStream_t stream)
{
// Calculate the number of elements N
size_t N = bytes / sizeof(T);
constexpr unsigned int items_per_block = BlockSize * ItemsPerThread;
const unsigned int grid_size = ((N + items_per_block - 1) / items_per_block);
const unsigned int size = items_per_block * grid_size;
std::vector<T> input = get_random_data<T>(size,
generate_limits<T>::min(),
generate_limits<T>::max(),
seed.get_0());
T* d_input;
unsigned int* d_output;
HIP_CHECK(hipMalloc(&d_input, size * sizeof(T)));
HIP_CHECK(hipMalloc(&d_output, size * sizeof(unsigned int)));
HIP_CHECK(hipMemcpy(d_input, input.data(), size * sizeof(T), hipMemcpyHostToDevice));
HIP_CHECK(hipDeviceSynchronize());
for(auto _ : state)
{
auto start = std::chrono::steady_clock::now();
hipLaunchKernelGGL(HIP_KERNEL_NAME(rank_kernel<T,
BlockSize,
ItemsPerThread,
RadixBits,
Descending,
Algorithm,
Trials>),
dim3(grid_size),
dim3(BlockSize),
0,
stream,
d_input,
d_output);
HIP_CHECK(hipPeekAtLastError());
HIP_CHECK(hipDeviceSynchronize());
auto end = std::chrono::steady_clock::now();
auto elapsed_seconds
= std::chrono::duration_cast<std::chrono::duration<double>>(end - start);
state.SetIterationTime(elapsed_seconds.count());
}
state.SetBytesProcessed(state.iterations() * Trials * size * sizeof(T));
state.SetItemsProcessed(state.iterations() * Trials * size);
HIP_CHECK(hipFree(d_input));
HIP_CHECK(hipFree(d_output));
}
#define CREATE_BENCHMARK(T, BS, IPT, KIND) \
benchmark::RegisterBenchmark( \
bench_naming::format_name("{lvl:block,algo:radix_rank,key_type:" #T ",cfg:{bs:" #BS \
",ipt:" #IPT ",method:" #KIND "}}") \
.c_str(), \
run_benchmark<T, BS, IPT, KIND>, \
bytes, \
seed, \
stream)
// clang-format off
#define CREATE_BENCHMARK_KINDS(type, block, ipt) \
CREATE_BENCHMARK(type, block, ipt, rp::block_radix_rank_algorithm::basic), \
CREATE_BENCHMARK(type, block, ipt, rp::block_radix_rank_algorithm::basic_memoize), \
CREATE_BENCHMARK(type, block, ipt, rp::block_radix_rank_algorithm::match)
#define BENCHMARK_TYPE(type, block) \
CREATE_BENCHMARK_KINDS(type, block, 1), \
CREATE_BENCHMARK_KINDS(type, block, 4), \
CREATE_BENCHMARK_KINDS(type, block, 8), \
CREATE_BENCHMARK_KINDS(type, block, 12), \
CREATE_BENCHMARK_KINDS(type, block, 16), \
CREATE_BENCHMARK_KINDS(type, block, 20)
// clang-format on
void add_benchmarks(std::vector<benchmark::internal::Benchmark*>& benchmarks,
size_t bytes,
const managed_seed& seed,
hipStream_t stream)
{
std::vector<benchmark::internal::Benchmark*> bs = {
BENCHMARK_TYPE(int, 128),
BENCHMARK_TYPE(int, 256),
BENCHMARK_TYPE(int, 512),
BENCHMARK_TYPE(uint8_t, 128),
BENCHMARK_TYPE(uint8_t, 256),
BENCHMARK_TYPE(uint8_t, 512),
BENCHMARK_TYPE(long long, 128),
BENCHMARK_TYPE(long long, 256),
BENCHMARK_TYPE(long long, 512),
};
benchmarks.insert(benchmarks.end(), bs.begin(), bs.end());
}
int main(int argc, char* argv[])
{
cli::Parser parser(argc, argv);
parser.set_optional<size_t>("size", "size", DEFAULT_BYTES, "number of bytes");
parser.set_optional<int>("trials", "trials", -1, "number of iterations");
parser.set_optional<std::string>("name_format",
"name_format",
"human",
"either: json,human,txt");
parser.set_optional<std::string>("seed", "seed", "random", get_seed_message());
parser.run_and_exit_if_error();
// Parse argv
benchmark::Initialize(&argc, argv);
const size_t bytes = parser.get<size_t>("size");
const int trials = parser.get<int>("trials");
bench_naming::set_format(parser.get<std::string>("name_format"));
const std::string seed_type = parser.get<std::string>("seed");
const managed_seed seed(seed_type);
// HIP
hipStream_t stream = 0; // default
// Benchmark info
add_common_benchmark_info();
benchmark::AddCustomContext("bytes", std::to_string(bytes));
benchmark::AddCustomContext("seed", seed_type);
// Add benchmarks
std::vector<benchmark::internal::Benchmark*> benchmarks;
add_benchmarks(benchmarks, bytes, seed, stream);
// Use manual timing
for(auto& b : benchmarks)
{
b->UseManualTime();
b->Unit(benchmark::kMillisecond);
}
// Force number of iterations
if(trials > 0)
{
for(auto& b : benchmarks)
{
b->Iterations(trials);
}
}
// Run benchmarks
benchmark::RunSpecifiedBenchmarks();
return 0;
}
|
