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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <armnn/ArmNN.hpp>
#include <armnn/Descriptors.hpp>
#include <armnn/IRuntime.hpp>
#include <armnn/INetwork.hpp>
#include <armnnUtils/Half.hpp>
#include <Graph.hpp>
#include <Optimizer.hpp>
#include <backends/CpuTensorHandle.hpp>
#include <backends/test/QuantizeHelper.hpp>
#include <boost/core/ignore_unused.hpp>
#include <boost/test/unit_test.hpp>
#include <set>
using namespace armnn;
BOOST_AUTO_TEST_SUITE(Fp16Support)
BOOST_AUTO_TEST_CASE(Fp16DataTypeSupport)
{
Graph graph;
Layer* const inputLayer1 = graph.AddLayer<InputLayer>(1, "input1");
Layer* const inputLayer2 = graph.AddLayer<InputLayer>(2, "input2");
Layer* const additionLayer = graph.AddLayer<AdditionLayer>("addition");
Layer* const outputLayer = graph.AddLayer<armnn::OutputLayer>(0, "output");
TensorInfo fp16TensorInfo({1, 2, 3, 5}, armnn::DataType::Float16);
inputLayer1->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(0));
inputLayer2->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(1));
additionLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
inputLayer1->GetOutputSlot().SetTensorInfo(fp16TensorInfo);
inputLayer2->GetOutputSlot().SetTensorInfo(fp16TensorInfo);
additionLayer->GetOutputSlot().SetTensorInfo(fp16TensorInfo);
BOOST_CHECK(inputLayer1->GetOutputSlot(0).GetTensorInfo().GetDataType() == armnn::DataType::Float16);
BOOST_CHECK(inputLayer2->GetOutputSlot(0).GetTensorInfo().GetDataType() == armnn::DataType::Float16);
BOOST_CHECK(additionLayer->GetOutputSlot(0).GetTensorInfo().GetDataType() == armnn::DataType::Float16);
}
BOOST_AUTO_TEST_CASE(Fp16AdditionTest)
{
using namespace half_float::literal;
// Create runtime in which test will run
IRuntime::CreationOptions options;
IRuntimePtr runtime(IRuntime::Create(options));
// Builds up the structure of the network.
INetworkPtr net(INetwork::Create());
IConnectableLayer* inputLayer1 = net->AddInputLayer(0);
IConnectableLayer* inputLayer2 = net->AddInputLayer(1);
IConnectableLayer* additionLayer = net->AddAdditionLayer();
IConnectableLayer* outputLayer = net->AddOutputLayer(0);
inputLayer1->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(0));
inputLayer2->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(1));
additionLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
//change to float16
TensorInfo fp16TensorInfo(TensorShape({4}), DataType::Float16);
inputLayer1->GetOutputSlot(0).SetTensorInfo(fp16TensorInfo);
inputLayer2->GetOutputSlot(0).SetTensorInfo(fp16TensorInfo);
additionLayer->GetOutputSlot(0).SetTensorInfo(fp16TensorInfo);
// optimize the network
std::vector<Compute> backends = {Compute::GpuAcc};
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
// Loads it into the runtime.
NetworkId netId;
runtime->LoadNetwork(netId, std::move(optNet));
std::vector<Half> input1Data
{
1.0_h, 2.0_h, 3.0_h, 4.0_h
};
std::vector<Half> input2Data
{
100.0_h, 200.0_h, 300.0_h, 400.0_h
};
InputTensors inputTensors
{
{0,ConstTensor(runtime->GetInputTensorInfo(netId, 0), input1Data.data())},
{1,ConstTensor(runtime->GetInputTensorInfo(netId, 0), input2Data.data())}
};
std::vector<Half> outputData(input1Data.size());
OutputTensors outputTensors
{
{0,Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
};
// Does the inference.
runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
// Checks the results.
BOOST_TEST(outputData == std::vector<Half>({ 101.0_h, 202.0_h, 303.0_h, 404.0_h})); // Add
}
BOOST_AUTO_TEST_SUITE_END()
|
