aboutsummaryrefslogtreecommitdiff
path: root/src/armnn/backends/test/NormTestImpl.hpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/armnn/backends/test/NormTestImpl.hpp')
-rw-r--r--src/armnn/backends/test/NormTestImpl.hpp241
1 files changed, 0 insertions, 241 deletions
diff --git a/src/armnn/backends/test/NormTestImpl.hpp b/src/armnn/backends/test/NormTestImpl.hpp
deleted file mode 100644
index 2690313655..0000000000
--- a/src/armnn/backends/test/NormTestImpl.hpp
+++ /dev/null
@@ -1,241 +0,0 @@
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "armnn/Exceptions.hpp"
-#include "armnn/LayerSupport.hpp"
-
-#include "backends/CpuTensorHandle.hpp"
-#include "backends/WorkloadFactory.hpp"
-
-LayerTestResult<float,4> SimpleNormalizationTestImpl(armnn::IWorkloadFactory& workloadFactory,
- armnn::NormalizationAlgorithmChannel normChannel,
- armnn::NormalizationAlgorithmMethod normMethod)
-{
- const unsigned int inputHeight = 2;
- const unsigned int inputWidth = 2;
- const unsigned int inputChannels = 1;
- const unsigned int inputNum = 2;
-
- unsigned int outputHeight = inputHeight;
- unsigned int outputWidth = inputWidth;
- unsigned int outputChannels = inputChannels;
- unsigned int outputNum = inputNum;
-
- unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth };
- unsigned int outputShape[] = { outputNum, outputChannels, outputHeight, outputWidth };
-
- auto inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32);
- auto outputTensorInfo = armnn::TensorInfo(4, outputShape, armnn::DataType::Float32);
-
- LayerTestResult<float,4> ret(outputTensorInfo);
-
- auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({
- // Batch #0
- 1.0f, 2.0f,
- 3.0f, 4.0f,
- // Batch #1
- 5.0f, 6.0f,
- 7.0f, 8.0f
- }));
-
- float alpha = 1.f;
- float beta = 1.f;
- float kappa = 1.f;
- uint32_t normSize = 3;
-
- std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
- std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
-
- armnn::NormalizationQueueDescriptor data;
- armnn::WorkloadInfo info;
- AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
- AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
- data.m_Parameters.m_NormChannelType = normChannel;
- data.m_Parameters.m_NormMethodType = normMethod;
- data.m_Parameters.m_NormSize = normSize;
- data.m_Parameters.m_Alpha = alpha;
- data.m_Parameters.m_Beta = beta;
- data.m_Parameters.m_K = kappa;
-
- armnn::PassthroughCpuTensorHandle refHandle(outputTensorInfo, &ret.outputExpected[0][0][0][0]);
- armnn::NormalizationQueueDescriptor refData = data;
- armnn::WorkloadInfo refInfo = info;
- SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, &refHandle);
-
- std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateNormalization(data, info);
-
- inputHandle->Allocate();
- outputHandle->Allocate();
-
- CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]);
-
- workloadFactory.Finalize();
- workload->Execute();
-
- CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());
-
- switch (normMethod)
- {
- case armnn::NormalizationAlgorithmMethod::LocalBrightness:
- {
- switch (normChannel)
- {
- case armnn::NormalizationAlgorithmChannel::Within:
- {
- // When normalising within channels, the 3x3 kernel covers the entire 2x2 input at every index.
- // Therefore, all output values should equal the inputs, but divided by:
- // pow((kappa + (accumulatedScale * alpha)), beta)
- // ...where accumulatedScale is the sum of every element squared.
- float divisor[inputNum];
- for(int i = 0; i < boost::numeric_cast<int>(inputNum); i++)
- {
- float accumulatedScale = input[i][0][0][0]*input[i][0][0][0] +
- input[i][0][0][1]*input[i][0][0][1] +
- input[i][0][1][0]*input[i][0][1][0] +
- input[i][0][1][1]*input[i][0][1][1];
- divisor[i] = powf((kappa + accumulatedScale * alpha), beta);
- }
- ret.outputExpected = MakeTensor<float, 4>(outputTensorInfo,
- std::vector<float>({input[0][0][0][0]/divisor[0],
- input[0][0][0][1]/divisor[0],
- input[0][0][1][0]/divisor[0],
- input[0][0][1][1]/divisor[0],
- input[1][0][0][0]/divisor[1],
- input[1][0][0][1]/divisor[1],
- input[1][0][1][0]/divisor[1],
- input[1][0][1][1]/divisor[1]}));
- break;
- }
- case armnn::NormalizationAlgorithmChannel::Across:
- {
- // When normalising across channels, all output values should equal the inputs, but multiplied by:
- // pow((kappa + (accumulatedScale * alpha)), -beta)
- // ...where accumulatedScale is the sum of the inputs for adjacent channels for this element squared
- // ...where adjacent channels means within half the normSize for the channel
- // The test data has only one channel, so this is simplified below.
- std::vector<float> outputVector;
- for (int n = 0; n < boost::numeric_cast<int>(inputNum); ++n)
- {
- for (int h = 0; h < boost::numeric_cast<int>(inputHeight); ++h)
- {
- for (int w = 0; w < boost::numeric_cast<int>(inputWidth); ++w)
- {
- float accumulatedScale = input[n][0][h][w]*input[n][0][h][w];
- float scale = powf((kappa + accumulatedScale * alpha), -beta);
- outputVector.push_back(input[n][0][h][w] * scale);
- }
- }
- }
- ret.outputExpected = MakeTensor<float, 4>(outputTensorInfo, outputVector);
- break;
- }
- default:
- {
- throw armnn::UnimplementedException("Unsupported normalisation channel type, "
- "only Across and Within are supported");
- }
- }
- break;
- }
- case armnn::NormalizationAlgorithmMethod::LocalContrast: // NOTE: intentional fallthrough.
- default:
- {
- throw armnn::UnimplementedException("Unsupported normalisation method type, "
- "only LocalBrightness is supported");
- }
- }
-
- return ret;
-}
-
-LayerTestResult<float,4> CompareNormalizationTestImpl(armnn::IWorkloadFactory& workloadFactory,
- armnn::IWorkloadFactory& refWorkloadFactory,
- armnn::NormalizationAlgorithmChannel normChannel,
- armnn::NormalizationAlgorithmMethod normMethod)
-{
- constexpr unsigned int inputNum = 5;
- constexpr unsigned int inputChannels = 3;
- constexpr unsigned int inputHeight = 32;
- constexpr unsigned int inputWidth = 24;
-
- constexpr unsigned int outputNum = inputNum;
- constexpr unsigned int outputChannels = inputChannels;
- constexpr unsigned int outputHeight = inputHeight;
- constexpr unsigned int outputWidth = inputWidth;
-
- armnn::TensorInfo inputTensorInfo;
- armnn::TensorInfo outputTensorInfo;
-
- unsigned int inputShape[] = {inputNum, inputChannels, inputHeight, inputWidth};
- unsigned int outputShape[] = {outputNum, outputChannels, outputHeight, outputWidth};
-
- inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32);
- outputTensorInfo = armnn::TensorInfo(4, outputShape, armnn::DataType::Float32);
-
- LayerTestResult<float,4> ret(outputTensorInfo);
-
- auto input = MakeRandomTensor<float, 4>(inputTensorInfo, 111234);
-
- constexpr float alpha = 1.f;
- constexpr float beta = 1.f;
- constexpr float kappa = 1.f;
- constexpr uint32_t normSize = 5;
-
- std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
- std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
-
- armnn::NormalizationQueueDescriptor data;
- armnn::WorkloadInfo info;
- AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
- AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
- data.m_Parameters.m_NormChannelType = normChannel;
- data.m_Parameters.m_NormMethodType = normMethod;
- data.m_Parameters.m_NormSize = normSize;
- data.m_Parameters.m_Alpha = alpha;
- data.m_Parameters.m_Beta = beta;
- data.m_Parameters.m_K = kappa;
-
- std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo);
- std::unique_ptr<armnn::ITensorHandle> inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo);
-
- armnn::NormalizationQueueDescriptor refData = data;
- armnn::WorkloadInfo refInfo = info;
- SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get());
- SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get());
-
- // Don't execute if Normalization is not supported for the method and channel types, as an exception will be raised.
- armnn::Compute compute = workloadFactory.GetCompute();
- const size_t reasonIfUnsupportedMaxLen = 255;
- char reasonIfUnsupported[reasonIfUnsupportedMaxLen+1];
- ret.supported = armnn::IsNormalizationSupported(compute, inputTensorInfo, outputTensorInfo, data.m_Parameters,
- reasonIfUnsupported, reasonIfUnsupportedMaxLen);
- if (!ret.supported)
- {
- return ret;
- }
-
- std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateNormalization(data, info);
- std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateNormalization(refData, refInfo);
-
- outputHandleRef->Allocate();
- inputHandleRef->Allocate();
-
- inputHandle->Allocate();
- outputHandle->Allocate();
-
- CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]);
- CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]);
-
- workloadFactory.Finalize();
- workload->Execute();
- refWorkloadFactory.Finalize();
- workloadRef->Execute();
-
- CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());
- CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get());
-
- return ret;
-}
-
generated by cgit v1.2.1 (git 2.23.0) at 2024-07-08 05:03:56 +0000