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
|
//
// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ExecuteNetworkParams.hpp"
#include "NetworkExecutionUtils/NetworkExecutionUtils.hpp"
#include <armnn/Logging.hpp>
#include <fmt/format.h>
#include <armnnUtils/Filesystem.hpp>
void CheckClTuningParameter(const int& tuningLevel,
const std::string& tuningPath,
const std::vector<armnn::BackendId> computeDevices)
{
if (!tuningPath.empty())
{
if (tuningLevel == 0)
{
ARMNN_LOG(info) << "Using cl tuning file: " << tuningPath << "\n";
if (!ValidatePath(tuningPath, true))
{
throw armnn::InvalidArgumentException("The tuning path is not valid");
}
}
else if ((1 <= tuningLevel) && (tuningLevel <= 3))
{
ARMNN_LOG(info) << "Starting execution to generate a cl tuning file: " << tuningPath << "\n"
<< "Tuning level in use: " << tuningLevel << "\n";
}
else if ((0 < tuningLevel) || (tuningLevel > 3))
{
throw armnn::InvalidArgumentException(fmt::format("The tuning level {} is not valid.",
tuningLevel));
}
// Ensure that a GpuAcc is enabled. Otherwise no tuning data are used or genereted
// Only warn if it's not enabled
auto it = std::find(computeDevices.begin(), computeDevices.end(), "GpuAcc");
if (it == computeDevices.end())
{
ARMNN_LOG(warning) << "To use Cl Tuning the compute device GpuAcc needs to be active.";
}
}
}
void ExecuteNetworkParams::ValidateParams()
{
if (m_DynamicBackendsPath == "")
{
// Check compute devices are valid unless they are dynamically loaded at runtime
std::string invalidBackends;
if (!CheckRequestedBackendsAreValid(m_ComputeDevices, armnn::Optional<std::string&>(invalidBackends)))
{
ARMNN_LOG(fatal) << "The list of preferred devices contains invalid backend IDs: "
<< invalidBackends;
}
}
CheckClTuningParameter(m_TuningLevel, m_TuningPath, m_ComputeDevices);
if (m_EnableBf16TurboMode && !m_EnableFastMath)
{
throw armnn::InvalidArgumentException("To use BF16 please use --enable-fast-math. ");
}
// Check input tensor shapes
if ((m_InputTensorShapes.size() != 0) &&
(m_InputTensorShapes.size() != m_InputNames.size()))
{
throw armnn::InvalidArgumentException("input-name and input-tensor-shape must have "
"the same amount of elements. ");
}
if (m_InputTensorDataFilePaths.size() != 0)
{
if (!ValidatePaths(m_InputTensorDataFilePaths, true))
{
throw armnn::InvalidArgumentException("One or more input data file paths are not valid.");
}
if (m_InputTensorDataFilePaths.size() < m_InputNames.size())
{
throw armnn::InvalidArgumentException(
fmt::format("According to the number of input names the user provided the network has {} "
"inputs. But only {} input-tensor-data file paths were provided. Each input of the "
"model is expected to be stored in it's own file.",
m_InputNames.size(),
m_InputTensorDataFilePaths.size()));
}
}
// Check that threshold time is not less than zero
if (m_ThresholdTime < 0)
{
throw armnn::InvalidArgumentException("Threshold time supplied as a command line argument is less than zero.");
}
// Warn if ExecuteNetwork will generate dummy input data
if (m_GenerateTensorData)
{
ARMNN_LOG(warning) << "No input files provided, input tensors will be filled with 0s.";
}
if (m_AllowExpandedDims && m_InferOutputShape)
{
throw armnn::InvalidArgumentException("infer-output-shape and allow-expanded-dims cannot be used together.");
}
}
#if defined(ARMNN_TFLITE_DELEGATE)
/**
* A utility method that populates a DelegateOptions object from this ExecuteNetworkParams.
*
* @return a populated armnnDelegate::DelegateOptions object.
*/
armnnDelegate::DelegateOptions ExecuteNetworkParams::ToDelegateOptions() const
{
armnnDelegate::DelegateOptions delegateOptions(m_ComputeDevices);
delegateOptions.SetDynamicBackendsPath(m_DynamicBackendsPath);
delegateOptions.SetGpuProfilingState(m_EnableProfiling);
delegateOptions.SetInternalProfilingParams(m_EnableProfiling, armnn::ProfilingDetailsMethod::Undefined);
if (m_OutputDetailsOnlyToStdOut)
{
delegateOptions.SetInternalProfilingParams(m_EnableProfiling, armnn::ProfilingDetailsMethod::DetailsOnly);
}
else if (m_OutputDetailsToStdOut)
{
delegateOptions.SetInternalProfilingParams(m_EnableProfiling, armnn::ProfilingDetailsMethod::DetailsWithEvents);
}
// GPU Backend options first.
{
armnn::BackendOptions gpuOption("GpuAcc", {{"TuningLevel", m_TuningLevel}});
delegateOptions.AddBackendOption(gpuOption);
}
{
armnn::BackendOptions gpuOption("GpuAcc", {{"TuningFile", m_TuningPath.c_str()}});
delegateOptions.AddBackendOption(gpuOption);
}
{
armnn::BackendOptions gpuOption("GpuAcc", {{"KernelProfilingEnabled", m_EnableProfiling}});
delegateOptions.AddBackendOption(gpuOption);
}
// Optimizer options next.
armnn::OptimizerOptionsOpaque optimizerOptions;
optimizerOptions.SetReduceFp32ToFp16(m_EnableFp16TurboMode);
optimizerOptions.SetDebugEnabled(m_PrintIntermediate);
optimizerOptions.SetDebugToFileEnabled(m_PrintIntermediateOutputsToFile);
optimizerOptions.SetProfilingEnabled(m_EnableProfiling);
optimizerOptions.SetShapeInferenceMethod(armnn::ShapeInferenceMethod::ValidateOnly);
if (m_InferOutputShape)
{
optimizerOptions.SetShapeInferenceMethod(armnn::ShapeInferenceMethod::InferAndValidate);
armnn::BackendOptions networkOption("ShapeInferenceMethod",
{
{"InferAndValidate", true}
});
optimizerOptions.AddModelOption(networkOption);
}
{
armnn::BackendOptions option("GpuAcc", {{"FastMathEnabled", m_EnableFastMath}});
optimizerOptions.AddModelOption(option);
}
{
armnn::BackendOptions option("GpuAcc", {{"CachedNetworkFilePath", m_CachedNetworkFilePath}});
optimizerOptions.AddModelOption(option);
}
{
armnn::BackendOptions option("GpuAcc", {{"MLGOTuningFilePath", m_MLGOTuningFilePath}});
optimizerOptions.AddModelOption(option);
}
armnn::BackendOptions cpuAcc("CpuAcc",
{
{ "FastMathEnabled", m_EnableFastMath },
{ "NumberOfThreads", m_NumberOfThreads }
});
optimizerOptions.AddModelOption(cpuAcc);
if (m_AllowExpandedDims)
{
armnn::BackendOptions networkOption("AllowExpandedDims",
{
{"AllowExpandedDims", true}
});
optimizerOptions.AddModelOption(networkOption);
}
delegateOptions.SetOptimizerOptions(optimizerOptions);
return delegateOptions;
}
#endif
|
