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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ClWorkloadFactoryHelper.hpp"
#include <Network.hpp>
#include <GraphUtils.hpp>
#include <cl/ClWorkloadFactory.hpp>
#include <cl/ClBackendContext.hpp>
#include <armnnUtils/Filesystem.hpp>
#include <doctest/doctest.h>
TEST_SUITE("ClOptimizedNetwork")
{
TEST_CASE("OptimizeValidateGpuDeviceSupportLayerNoFallback")
{
// build up the structure of the network
armnn::INetworkPtr net(armnn::INetwork::Create());
armnn::IConnectableLayer* input = net->AddInputLayer(0);
armnn::IConnectableLayer* output = net->AddOutputLayer(0);
input->GetOutputSlot(0).Connect(output->GetInputSlot(0));
input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
CHECK(optNet);
// validate workloads
armnn::ClWorkloadFactory fact =
ClWorkloadFactoryHelper::GetFactory(ClWorkloadFactoryHelper::GetMemoryManager());
const armnn::Graph& theGraph = GetGraphForTesting(optNet.get());
for (auto&& layer : theGraph)
{
CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc);
CHECK_NOTHROW(
layer->CreateWorkload(fact));
}
}
TEST_CASE("FP16TurboModeTestOnGpuAcc")
{
// Test to check when Fp16 Turbo mode set
// it converts the Fp32 network to Fp16 Network
// add Fp32ToFp16 conversion layer after the InputLayer
// add Fp16ToFp32 conversion layer after the OutputLayer
// checks the other layers if they are supported in Fp16
// if they are not put the conversion layers before and after
// if they are not supported in Fp16 use Fp32 instead
// if there are inverse conversion layers remove them with optimization
// at the moment FloorLayer is not supported in Fp16 so it rolls back to Fp32
// and inverse conversion layers are removed by the optimizer
armnn::INetworkPtr net(armnn::INetwork::Create());
// Defines layers.
auto input = net->AddInputLayer(0, "input layer");
// ReLu1
armnn::ActivationDescriptor activation1Descriptor;
activation1Descriptor.m_Function = armnn::ActivationFunction::BoundedReLu;
activation1Descriptor.m_A = 1.f;
activation1Descriptor.m_B = -1.f;
auto activation = net->AddActivationLayer(activation1Descriptor, "activation layer");
auto output = net->AddOutputLayer(0, "output layer");
// Connects layers.
input->GetOutputSlot(0).Connect(activation->GetInputSlot(0));
activation->GetOutputSlot(0).Connect(output->GetInputSlot(0));
armnn::TensorShape shape({4});
armnn::TensorInfo info(shape, armnn::DataType::Float32);
input->GetOutputSlot(0).SetTensorInfo(info);
activation->GetOutputSlot(0).SetTensorInfo(info);
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
armnn::OptimizerOptions optimizerOptions;
optimizerOptions.m_ReduceFp32ToFp16 = true;
armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(
*net, backends, runtime->GetDeviceSpec(), optimizerOptions);
const armnn::Graph& graph = GetGraphForTesting(optimizedNet.get());
// Tests that all layers are present in the graph.
CHECK(graph.GetNumLayers() == 5);
// Tests that the vertices exist and have correct names.
CHECK(GraphHasNamedLayer(graph, "input layer"));
CHECK(GraphHasNamedLayer(graph, "convert_fp32_to_fp16-0-input layer"));
CHECK(GraphHasNamedLayer(graph, "activation layer"));
CHECK(GraphHasNamedLayer(graph, "convert_fp16_to_fp32-0-output layer"));
CHECK(GraphHasNamedLayer(graph, "output layer"));
}
TEST_CASE("FastMathEnabledTestOnGpuAcc")
{
armnn::INetworkPtr net(armnn::INetwork::Create());
armnn::IConnectableLayer* input = net->AddInputLayer(0);
armnn::IConnectableLayer* output = net->AddOutputLayer(0);
input->GetOutputSlot(0).Connect(output->GetInputSlot(0));
input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
armnn::OptimizerOptions optimizerOptions;
armnn::BackendOptions modelOptions("GpuAcc", {{"FastMathEnabled", true}});
optimizerOptions.m_ModelOptions.push_back(modelOptions);
armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(
*net, backends, runtime->GetDeviceSpec(), optimizerOptions);
CHECK(optimizedNet);
auto modelOptionsOut = GetModelOptionsForTesting(optimizedNet.get());
CHECK(modelOptionsOut.size() == 1);
CHECK(modelOptionsOut[0].GetOption(0).GetName() == "FastMathEnabled");
CHECK(modelOptionsOut[0].GetOption(0).GetValue().AsBool() == true);
}
TEST_CASE("CheckMLGOTuningFile")
{
class ClBackendContextTestClass : public armnn::ClBackendContext
{
public:
ClBackendContextTestClass(const armnn::IRuntime::CreationOptions &options) : ClBackendContext(options)
{}
bool call_reload_from_file()
{
return m_MLGOTuner.reload_from_file(m_MLGOTuningFile);
}
};
const std::string validText{
"<header>\n"
"gemm-version, [1,2,1]\n"
"ip-type,gpu\n"
"</header>\n"
"<heuristics-table>\n"
"0, g71 , 8, f32, best-performance, static, gemm-type, [m,n,k,n]\n"
"1, g71 , 8, f32, best-performance, static, gemm-config-reshaped-only-rhs, [m,n,k,n]\n"
"2, g71 , 8, f32, best-performance, static, gemm-config-reshaped, [m,n,k,n]\n"
"3, g71 , 8, qasymm8, best-performance, static, gemm-type, [m,n,k,n]\n"
"4, g71 , 8, qasymm8, best-performance, static, gemm-config-reshaped-only-rhs, [m,n,k,n]\n"
"5, g71 , 8, qasymm8, best-performance, static, gemm-config-native, [m,n,k,n]\n"
"</heuristics-table>\n"
"<heuristic, 0>\n"
"b , 0, var, r_mn, >=, num, 2., 1, 2\n"
"l , 1, gemm-type, reshaped\n"
"l , 2, gemm-type, reshaped-only-rhs\n"
"</heuristic>\n"
"<heuristic, 1>\n"
"l ,0,gemm-config-reshaped-only-rhs, [2, 4,4,4,1,1,0]\n"
"</heuristic>\n"
"<heuristic, 2>\n"
"l ,0,gemm-config-reshaped,[4,2,8,16,16,1,0,1,0]\n"
"</heuristic>\n"
"<heuristic, 3>\n"
"l , 0, gemm-type, native\n"
"</heuristic>\n"
"<heuristic, 4>\n"
"l ,0,gemm-config-reshaped-only-rhs, [2, 4,4,4,1,1,0]\n"
"</heuristic>\n"
"<heuristic, 5>\n"
"l ,0,gemm-config-native,[4,2,8]\n"
"</heuristic>\n"};
const std::string invalidText{"ʕノ•ᴥ•ʔノ ︵ ┻━┻"};
fs::path validFile = armnnUtils::Filesystem::NamedTempFile("validFile.mlgo");
fs::path invalidFile = armnnUtils::Filesystem::NamedTempFile("invalidFile.mlgo");
try
{
std::ofstream ofs1{validFile};
ofs1 << validText << std::endl;
ofs1.close();
std::ofstream ofs2{invalidFile};
ofs2 << invalidText << std::endl;
ofs2.close();
}
catch (std::exception &e)
{
std::cerr << "Unable to write to file at location [" << validFile.c_str() << "] : " << e.what() << std::endl;
CHECK(false);
}
armnn::IRuntime::CreationOptions creationOptions1;
armnn::BackendOptions validOptions
{
"GpuAcc",
{
{"MLGOTuningFilePath", validFile.c_str()}
}
};
creationOptions1.m_BackendOptions.emplace_back(validOptions);
ClBackendContextTestClass clBackendContext1(creationOptions1);
CHECK(clBackendContext1.call_reload_from_file());
armnn::BackendOptions invalidOptions
{
"GpuAcc",
{
{"MLGOTuningFilePath", invalidFile.c_str()}
}
};
armnn::IRuntime::CreationOptions creationOptions2;
creationOptions2.m_BackendOptions.emplace_back(invalidOptions);
ClBackendContextTestClass clBackendContext2(creationOptions2);
CHECK(clBackendContext2.call_reload_from_file() == false);
armnn::BackendOptions invalidPathOptions
{
"GpuAcc",
{
{"MLGOTuningFilePath", "not_a_real_file_path"}
}
};
armnn::IRuntime::CreationOptions creationOptions3;
creationOptions3.m_BackendOptions.emplace_back(invalidPathOptions);
ClBackendContextTestClass clBackendContext3(creationOptions3);
CHECK(clBackendContext3.call_reload_from_file() == false);
}
}
|
