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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <armnn/ArmNN.hpp>
#include <armnn/Network.hpp>
#include <armnn/test/GraphUtils.hpp>
#include <backends/cl/ClWorkloadFactory.hpp>
#include <boost/test/unit_test.hpp>
BOOST_AUTO_TEST_SUITE(ClOptimizedNetwork)
BOOST_AUTO_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());
BOOST_CHECK(optNet);
// validate workloads
armnn::ClWorkloadFactory fact;
for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
{
BOOST_CHECK(layer->GetBackendId() == armnn::Compute::GpuAcc);
BOOST_CHECK_NO_THROW(
layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact));
}
}
BOOST_AUTO_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::Network net;
// 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 = static_cast<armnn::OptimizedNetwork*>(optimizedNet.get())->GetGraph();
// Tests that all layers are present in the graph.
BOOST_TEST(graph.GetNumLayers() == 5);
// Tests that the vertices exist and have correct names.
BOOST_TEST(GraphHasNamedLayer(graph, "input layer"));
BOOST_TEST(GraphHasNamedLayer(graph, "convert_fp32_to_fp16-0-input layer"));
BOOST_TEST(GraphHasNamedLayer(graph, "activation layer"));
BOOST_TEST(GraphHasNamedLayer(graph, "convert_fp16_to_fp32-0-output layer"));
BOOST_TEST(GraphHasNamedLayer(graph, "output layer"));
}
BOOST_AUTO_TEST_SUITE_END();
|
