1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonWorkloadFactoryHelper.hpp"
7
8
9 #include <test/TensorHelpers.hpp>
10
11 #include <backendsCommon/CpuTensorHandle.hpp>
12 #include <backendsCommon/WorkloadFactory.hpp>
13
14 #include <neon/NeonTimer.hpp>
15 #include <neon/NeonWorkloadFactory.hpp>
16
17 #include <backendsCommon/test/LayerTests.hpp>
18 #include <backendsCommon/test/TensorCopyUtils.hpp>
19 #include <backendsCommon/test/WorkloadTestUtils.hpp>
20
21 #include <boost/test/unit_test.hpp>
22
23 #include <cstdlib>
24 #include <algorithm>
25
26 using namespace armnn;
27
28 BOOST_AUTO_TEST_SUITE(NeonTimerInstrument)
29
30
BOOST_AUTO_TEST_CASE(NeonTimerGetName)31 BOOST_AUTO_TEST_CASE(NeonTimerGetName)
32 {
33 NeonTimer neonTimer;
34 BOOST_CHECK_EQUAL(neonTimer.GetName(), "NeonKernelTimer");
35 }
36
BOOST_AUTO_TEST_CASE(NeonTimerMeasure)37 BOOST_AUTO_TEST_CASE(NeonTimerMeasure)
38 {
39 NeonWorkloadFactory workloadFactory =
40 NeonWorkloadFactoryHelper::GetFactory(NeonWorkloadFactoryHelper::GetMemoryManager());
41
42 unsigned int inputWidth = 2000u;
43 unsigned int inputHeight = 2000u;
44 unsigned int inputChannels = 1u;
45 unsigned int inputBatchSize = 1u;
46
47 float upperBound = 1.0f;
48 float lowerBound = -1.0f;
49
50 size_t inputSize = inputWidth * inputHeight * inputChannels * inputBatchSize;
51 std::vector<float> inputData(inputSize, 0.f);
52 std::generate(inputData.begin(), inputData.end(), [](){
53 return (static_cast<float>(rand()) / static_cast<float>(RAND_MAX / 3)) + 1.f; });
54
55 unsigned int outputWidth = inputWidth;
56 unsigned int outputHeight = inputHeight;
57 unsigned int outputChannels = inputChannels;
58 unsigned int outputBatchSize = inputBatchSize;
59
60 armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth },
61 armnn::DataType::Float32);
62
63 armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth },
64 armnn::DataType::Float32);
65
66 LayerTestResult<float, 4> result(inputTensorInfo);
67
68 auto input = MakeTensor<float, 4>(inputTensorInfo, inputData);
69
70 ARMNN_NO_DEPRECATE_WARN_BEGIN
71 std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
72 std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
73 ARMNN_NO_DEPRECATE_WARN_END
74
75 // Setup bounded ReLu
76 armnn::ActivationQueueDescriptor descriptor;
77 armnn::WorkloadInfo workloadInfo;
78 AddInputToWorkload(descriptor, workloadInfo, inputTensorInfo, inputHandle.get());
79 AddOutputToWorkload(descriptor, workloadInfo, outputTensorInfo, outputHandle.get());
80
81 descriptor.m_Parameters.m_Function = armnn::ActivationFunction::BoundedReLu;
82 descriptor.m_Parameters.m_A = upperBound;
83 descriptor.m_Parameters.m_B = lowerBound;
84
85 std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateActivation(descriptor, workloadInfo);
86
87 inputHandle->Allocate();
88 outputHandle->Allocate();
89
90 CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]);
91
92 NeonTimer neonTimer;
93 // Start the timer.
94 neonTimer.Start();
95 // Execute the workload.
96 workload->Execute();
97 // Stop the timer.
98 neonTimer.Stop();
99
100 std::vector<Measurement> measurements = neonTimer.GetMeasurements();
101
102 BOOST_CHECK(measurements.size() <= 2);
103 if (measurements.size() > 1)
104 {
105 BOOST_CHECK_EQUAL(measurements[0].m_Name, "NeonKernelTimer/0: NEFillBorderKernel");
106 BOOST_CHECK(measurements[0].m_Value > 0.0);
107 }
108 std::ostringstream oss;
109 oss << "NeonKernelTimer/" << measurements.size()-1 << ": NEActivationLayerKernel";
110 BOOST_CHECK_EQUAL(measurements[measurements.size()-1].m_Name, oss.str());
111 BOOST_CHECK(measurements[measurements.size()-1].m_Value > 0.0);
112 }
113
114 BOOST_AUTO_TEST_SUITE_END()
115