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
|
//
// Copyright © 2020 STMicroelectronics and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <algorithm>
#include <getopt.h>
#include <numeric>
#include <signal.h>
#include <string>
#include <sys/time.h>
#include <vector>
#include <armnn/BackendId.hpp>
#include <armnn/BackendRegistry.hpp>
#include <armnn/IRuntime.hpp>
#include <armnn/utility/NumericCast.hpp>
#include <armnnTfLiteParser/ITfLiteParser.hpp>
// Application parameters
std::vector<armnn::BackendId> preferred_backends_order = {armnn::Compute::CpuAcc, armnn::Compute::CpuRef};
std::string model_file_str;
std::string preferred_backend_str;
int nb_loops = 1;
double get_us(struct timeval t)
{
return (armnn::numeric_cast<double>(t.tv_sec) *
armnn::numeric_cast<double>(1000000) +
armnn::numeric_cast<double>(t.tv_usec));
}
double get_ms(struct timeval t)
{
return (armnn::numeric_cast<double>(t.tv_sec) *
armnn::numeric_cast<double>(1000) +
armnn::numeric_cast<double>(t.tv_usec) / 1000);
}
static void print_help(char** argv)
{
std::cout <<
"Usage: " << argv[0] << " -m <model .tflite>\n"
"\n"
"-m --model_file <.tflite file path>: .tflite model to be executed\n"
"-b --backend <device>: preferred backend device to run layers on by default. Possible choices: "
<< armnn::BackendRegistryInstance().GetBackendIdsAsString() << "\n"
"-l --loops <int>: provide the number of time the inference will be executed\n"
" (by default nb_loops=1)\n"
"--help: show this help\n";
exit(1);
}
void process_args(int argc, char** argv)
{
const char* const short_opts = "m:b:l:h";
const option long_opts[] = {
{"model_file", required_argument, nullptr, 'm'},
{"backend", required_argument, nullptr, 'b'},
{"loops", required_argument, nullptr, 'l'},
{"help", no_argument, nullptr, 'h'},
{nullptr, no_argument, nullptr, 0}
};
while (true)
{
const auto opt = getopt_long(argc, argv, short_opts, long_opts, nullptr);
if (-1 == opt)
{
break;
}
switch (opt)
{
case 'm':
model_file_str = std::string(optarg);
std::cout << "model file set to: " << model_file_str << std::endl;
break;
case 'b':
preferred_backend_str = std::string(optarg);
// Overwrite the prefered backend order
if (preferred_backend_str == "CpuAcc")
{
preferred_backends_order = {armnn::Compute::CpuAcc, armnn::Compute::CpuRef};
}
else if (preferred_backend_str == "CpuRef")
{
preferred_backends_order = {armnn::Compute::CpuRef, armnn::Compute::CpuAcc};
}
std::cout << "preferred backend device set to:";
for (unsigned int i = 0; i < preferred_backends_order.size(); i++)
{
std::cout << " " << preferred_backends_order.at(i);
}
std::cout << std::endl;
break;
case 'l':
nb_loops = std::stoi(optarg);
std::cout << "benchmark will execute " << nb_loops << " inference(s)" << std::endl;
break;
case 'h': // -h or --help
case '?': // Unrecognized option
default:
print_help(argv);
break;
}
}
if (model_file_str.empty())
{
print_help(argv);
}
}
int main(int argc, char* argv[])
{
std::vector<double> inferenceTimes;
// Get options
process_args(argc, argv);
// Create the runtime
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
// Create Parser
armnnTfLiteParser::ITfLiteParserPtr armnnparser(armnnTfLiteParser::ITfLiteParser::Create());
// Create a network
armnn::INetworkPtr network = armnnparser->CreateNetworkFromBinaryFile(model_file_str.c_str());
if (!network)
{
throw armnn::Exception("Failed to create an ArmNN network");
}
// Optimize the network
armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*network,
preferred_backends_order,
runtime->GetDeviceSpec());
armnn::NetworkId networkId;
// Load the network in to the runtime
runtime->LoadNetwork(networkId, std::move(optimizedNet));
// Check the number of subgraph
if (armnnparser->GetSubgraphCount() != 1)
{
std::cout << "Model with more than 1 subgraph is not supported by this benchmark application.\n";
exit(0);
}
size_t subgraphId = 0;
// Set up the input network
std::cout << "\nModel information:" << std::endl;
std::vector<armnnTfLiteParser::BindingPointInfo> inputBindings;
std::vector<armnn::TensorInfo> inputTensorInfos;
std::vector<std::string> inputTensorNames = armnnparser->GetSubgraphInputTensorNames(subgraphId);
for (unsigned int i = 0; i < inputTensorNames.size() ; i++)
{
std::cout << "inputTensorNames[" << i << "] = " << inputTensorNames[i] << std::endl;
armnnTfLiteParser::BindingPointInfo inputBinding = armnnparser->GetNetworkInputBindingInfo(
subgraphId,
inputTensorNames[i]);
armnn::TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(networkId, inputBinding.first);
inputBindings.push_back(inputBinding);
inputTensorInfos.push_back(inputTensorInfo);
}
// Set up the output network
std::vector<armnnTfLiteParser::BindingPointInfo> outputBindings;
std::vector<armnn::TensorInfo> outputTensorInfos;
std::vector<std::string> outputTensorNames = armnnparser->GetSubgraphOutputTensorNames(subgraphId);
for (unsigned int i = 0; i < outputTensorNames.size() ; i++)
{
std::cout << "outputTensorNames[" << i << "] = " << outputTensorNames[i] << std::endl;
armnnTfLiteParser::BindingPointInfo outputBinding = armnnparser->GetNetworkOutputBindingInfo(
subgraphId,
outputTensorNames[i]);
armnn::TensorInfo outputTensorInfo = runtime->GetOutputTensorInfo(networkId, outputBinding.first);
outputBindings.push_back(outputBinding);
outputTensorInfos.push_back(outputTensorInfo);
}
// Allocate input tensors
unsigned int nb_inputs = armnn::numeric_cast<unsigned int>(inputTensorInfos.size());
armnn::InputTensors inputTensors;
std::vector<std::vector<float>> in;
for (unsigned int i = 0 ; i < nb_inputs ; i++)
{
std::vector<float> in_data(inputTensorInfos.at(i).GetNumElements());
in.push_back(in_data);
inputTensors.push_back({ inputBindings[i].first, armnn::ConstTensor(inputBindings[i].second, in.data()) });
}
// Allocate output tensors
unsigned int nb_ouputs = armnn::numeric_cast<unsigned int>(outputTensorInfos.size());
armnn::OutputTensors outputTensors;
std::vector<std::vector<float>> out;
for (unsigned int i = 0; i < nb_ouputs ; i++)
{
std::vector<float> out_data(outputTensorInfos.at(i).GetNumElements());
out.push_back(out_data);
outputTensors.push_back({ outputBindings[i].first, armnn::Tensor(outputBindings[i].second, out[i].data()) });
}
// Run the inferences
std::cout << "\ninferences are running: " << std::flush;
for (int i = 0 ; i < nb_loops ; i++)
{
struct timeval start_time, stop_time;
gettimeofday(&start_time, nullptr);
runtime->EnqueueWorkload(networkId, inputTensors, outputTensors);
gettimeofday(&stop_time, nullptr);
inferenceTimes.push_back((get_us(stop_time) - get_us(start_time)));
std::cout << "# " << std::flush;
}
auto maxInfTime = *std::max_element(inferenceTimes.begin(), inferenceTimes.end());
auto minInfTime = *std::min_element(inferenceTimes.begin(), inferenceTimes.end());
auto avgInfTime = accumulate(inferenceTimes.begin(), inferenceTimes.end(), 0.0) /
armnn::numeric_cast<double>(inferenceTimes.size());
std::cout << "\n\ninference time: ";
std::cout << "min=" << minInfTime << "us ";
std::cout << "max=" << maxInfTime << "us ";
std::cout << "avg=" << avgInfTime << "us" << std::endl;
return 0;
}
|
