IVGCVSW-6124 ConstTensorsAsInput: Conv2d - FrontEnd
* Update Front-end and Tools.
* Updated Serializer, Deserializer and unit tests to reflect this.
* Updated TfLiteDelegate, TfLiteParser and OnnxParser.
* Updated Ref.
* Fixed resulting Neon / CL tests
* Unified optimizers for conv2d ops
* Optimizer Fix - Fp32ToBf16
* Partial implementation for ACL backends to fix VTS failures
!android-nn-driver:7477
Signed-off-by: Keith Davis <keith.davis@arm.com>
Change-Id: I5fb18877f7ee32643e15a9818945356274bb401b
diff --git a/src/armnnTestUtils/CreateWorkload.hpp b/src/armnnTestUtils/CreateWorkload.hpp
index 2590ae8..7700a55 100644
--- a/src/armnnTestUtils/CreateWorkload.hpp
+++ b/src/armnnTestUtils/CreateWorkload.hpp
@@ -483,28 +483,37 @@
layerDesc.m_PadBottom = 1;
layerDesc.m_StrideX = 2;
layerDesc.m_StrideY = 4;
- layerDesc.m_BiasEnabled = true;
+ layerDesc.m_BiasEnabled = false;
layerDesc.m_DataLayout = dataLayout;
+ float inputsQScale = DataType == armnn::DataType::QAsymmU8 ? 1.0f : 0.0;
+ float outputQScale = DataType == armnn::DataType::QAsymmU8 ? 2.0f : 0.0;
+
Convolution2dLayer* const layer = graph.AddLayer<Convolution2dLayer>(layerDesc, "layer");
TensorShape weightShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 5, 3} : TensorShape{2, 5, 3, 3};
TensorShape inputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 8, 16} : TensorShape{2, 8, 16, 3};
TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 2, 2, 10} : TensorShape{2, 2, 10, 2};
+ // As optimization isn't run member variables need to be updated.
layer->m_Weight = std::make_unique<ScopedTensorHandle>(TensorInfo(weightShape, DataType));
- layer->m_Bias = std::make_unique<ScopedTensorHandle>(TensorInfo({2}, GetBiasDataType(DataType)));
-
layer->m_Weight->Allocate();
- layer->m_Bias->Allocate();
+
+ armnn::TensorInfo weightsTensorInfo(weightShape, DataType, inputsQScale);
+ weightsTensorInfo.SetConstant();
// Creates extra layers.
Layer* const input = graph.AddLayer<InputLayer>(0, "input");
+ auto const weights = graph.AddLayer<ConstantLayer>("weights");
Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
+ weights->m_LayerOutput = std::make_unique<ScopedTensorHandle>(weightsTensorInfo);
+ weights->m_LayerOutput->Allocate();
+
// Connects up.
- Connect(input, layer, TensorInfo(inputShape, DataType));
- Connect(layer, output, TensorInfo(outputShape, DataType));
+ Connect(input, layer, TensorInfo(inputShape, DataType, inputsQScale));
+ Connect(weights, layer, weightsTensorInfo, 0, 1);
+ Connect(layer, output, TensorInfo(outputShape, DataType, outputQScale));
CreateTensorHandles(graph, factory);
// Makes the workload and checks it.
@@ -517,14 +526,11 @@
CHECK(queueDescriptor.m_Parameters.m_PadRight == 3);
CHECK(queueDescriptor.m_Parameters.m_PadTop == 1);
CHECK(queueDescriptor.m_Parameters.m_PadBottom == 1);
- CHECK(queueDescriptor.m_Parameters.m_BiasEnabled);
+ CHECK(!queueDescriptor.m_Parameters.m_BiasEnabled);
CHECK((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
- CHECK(queueDescriptor.m_Inputs.size() == 1);
+ CHECK(queueDescriptor.m_Inputs.size() == 2);
CHECK(queueDescriptor.m_Outputs.size() == 1);
- CHECK((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo(weightShape, DataType)));
- CHECK((queueDescriptor.m_Bias->GetTensorInfo() ==
- TensorInfo({2}, GetBiasDataType(DataType))));
// Returns so we can do extra, backend-specific tests.
return workload;
@@ -548,19 +554,26 @@
layerDesc.m_BiasEnabled = true;
layerDesc.m_DataLayout = dataLayout;
+ float inputsQScale = DataType == armnn::DataType::QAsymmU8 ? 1.0f : 0.0;
+ float outputQScale = DataType == armnn::DataType::QAsymmU8 ? 2.0f : 0.0;
Convolution2dLayer* const layer = graph.AddLayer<Convolution2dLayer>(layerDesc, "layer");
TensorShape weightShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 5, 3} : TensorShape{2, 5, 3, 3};
TensorShape inputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 8, 16} : TensorShape{2, 8, 16, 3};
TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 2, 2, 10} : TensorShape{2, 2, 10, 2};
-
+ // As optimization isn't run member variables need to be updated.
layer->m_Weight = std::make_unique<ScopedTensorHandle>(TensorInfo(weightShape, DataType));
layer->m_Bias = std::make_unique<ScopedTensorHandle>(TensorInfo({2}, GetBiasDataType(DataType)));
layer->m_Weight->Allocate();
layer->m_Bias->Allocate();
+ armnn::TensorInfo weightsTensorInfo(weightShape, DataType, inputsQScale);
+ weightsTensorInfo.SetConstant();
+ armnn::TensorInfo biasTensorInfo({2}, DataType, inputsQScale);
+ biasTensorInfo.SetConstant();
+
auto activationDesc = std::make_shared<ActivationDescriptor>();
activationDesc->m_A = 10.0f;
activationDesc->m_B = 5.0f;
@@ -579,11 +592,20 @@
// Creates extra layers.
Layer* const input = graph.AddLayer<InputLayer>(0, "input");
+ auto const weights = graph.AddLayer<ConstantLayer>("weights");
+ auto const bias = graph.AddLayer<ConstantLayer>("bias");
Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
+ weights->m_LayerOutput = std::make_unique<ScopedTensorHandle>(weightsTensorInfo);
+ weights->m_LayerOutput->Allocate();
+ bias->m_LayerOutput = std::make_unique<ScopedTensorHandle>(biasTensorInfo);
+ bias->m_LayerOutput->Allocate();
+
// Connects up.
- Connect(input, layer, TensorInfo(inputShape, DataType));
- Connect(layer, output, TensorInfo(outputShape, DataType));
+ Connect(input, layer, TensorInfo(inputShape, DataType, inputsQScale));
+ Connect(weights, layer, weightsTensorInfo, 0, 1);
+ Connect(bias, layer, biasTensorInfo, 0, 2);
+ Connect(layer, output, TensorInfo(outputShape, DataType, outputQScale));
CreateTensorHandles(graph, factory);
// Makes the workload and checks it.
@@ -606,11 +628,9 @@
CHECK(queueDescriptor.m_Parameters.m_PadBottom == 1);
CHECK(queueDescriptor.m_Parameters.m_BiasEnabled);
CHECK((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
+
CHECK(queueDescriptor.m_Outputs.size() == 1);
- CHECK((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo(weightShape, DataType)));
- CHECK((queueDescriptor.m_Bias->GetTensorInfo() ==
- TensorInfo({2}, GetBiasDataType(DataType))));
- CHECK(queueDescriptor.m_Inputs.size() == 1);
+ CHECK(queueDescriptor.m_Inputs.size() == 3);
// Returns so we can do extra, backend-specific tests.
return workload;
@@ -630,28 +650,41 @@
layerDesc.m_PadBottom = 0;
layerDesc.m_StrideX = 1;
layerDesc.m_StrideY = 1;
- layerDesc.m_BiasEnabled = false;
+ layerDesc.m_BiasEnabled = true;
layerDesc.m_DataLayout = dataLayout;
+ float inputsQScale = DataType == armnn::DataType::QAsymmU8 ? 1.0f : 0.0;
+ float outputQScale = DataType == armnn::DataType::QAsymmU8 ? 2.0f : 0.0;
+
Convolution2dLayer* const layer = graph.AddLayer<Convolution2dLayer>(layerDesc, "layer");
- TensorShape weightShape = TensorShape{32, 32, 3, 3};
- TensorShape inputShape = TensorShape{1, 32, 149, 149};
- TensorShape outputShape = TensorShape{1, 32, 147, 147};
-
+ TensorShape weightShape = TensorShape{ 32, 32, 3, 3 };
+ TensorShape biasShape = TensorShape{ 32 };
+ TensorShape inputShape = TensorShape{ 1, 32, 149, 149 };
+ TensorShape outputShape = TensorShape{ 1, 32, 147, 147 };
+ // As optimization isn't run member variables need to be updated.
layer->m_Weight = std::make_unique<ScopedTensorHandle>(TensorInfo(weightShape, DataType));
- layer->m_Bias = std::make_unique<ScopedTensorHandle>(TensorInfo({2}, GetBiasDataType(DataType)));
+ layer->m_Bias = std::make_unique<ScopedTensorHandle>(TensorInfo(biasShape, GetBiasDataType(DataType)));
layer->m_Weight->Allocate();
layer->m_Bias->Allocate();
+ armnn::TensorInfo weightsTensorInfo(weightShape, DataType, inputsQScale);
+ weightsTensorInfo.SetConstant();
+ armnn::TensorInfo biasTensorInfo(biasShape, DataType, inputsQScale);
+ biasTensorInfo.SetConstant();
+
// Creates extra layers.
Layer* const input = graph.AddLayer<InputLayer>(0, "input");
+ auto const weights = graph.AddLayer<ConstantLayer>("weights");
+ auto const bias = graph.AddLayer<ConstantLayer>("bias");
Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
// Connects up.
Connect(input, layer, TensorInfo(inputShape, DataType));
- Connect(layer, output, TensorInfo(outputShape, DataType));
+ Connect(weights, layer, weightsTensorInfo, 0, 1);
+ Connect(bias, layer, biasTensorInfo, 0, 2);
+ Connect(layer, output, TensorInfo(outputShape, DataType, outputQScale));
CreateTensorHandles(graph, factory);
// Makes the workload and checks it.
@@ -666,9 +699,8 @@
CHECK(queueDescriptor.m_Parameters.m_PadBottom == 0);
CHECK((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
- CHECK(queueDescriptor.m_Inputs.size() == 1);
+ CHECK(queueDescriptor.m_Inputs.size() == 3);
CHECK(queueDescriptor.m_Outputs.size() == 1);
- CHECK((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo(weightShape, DataType)));
// Returns so we can do extra, backend-specific tests.
return workload;
@@ -1074,9 +1106,9 @@
return workload;
}
-template <typename Convolution2dWorkload, armnn::DataType DataType>
+template<typename Convolution2dWorkload, armnn::DataType DataType>
std::unique_ptr<Convolution2dWorkload> CreateDirectConvolution2dWorkloadTest(armnn::IWorkloadFactory& factory,
- armnn::Graph& graph)
+ armnn::Graph& graph)
{
// Creates the layer we're testing.
Convolution2dDescriptor layerDesc;
@@ -1093,18 +1125,34 @@
float inputsQScale = DataType == armnn::DataType::QAsymmU8 ? 1.0f : 0.0;
float outputQScale = DataType == armnn::DataType::QAsymmU8 ? 2.0f : 0.0;
- layer->m_Weight = std::make_unique<ScopedTensorHandle>(TensorInfo({ 2, 3, 3, 3 }, DataType, inputsQScale));
- layer->m_Bias = std::make_unique<ScopedTensorHandle>
- (TensorInfo({2}, GetBiasDataType(DataType), inputsQScale));
+ TensorShape biasShape = TensorShape{ 2 };
+ TensorShape weightShape = TensorShape{ 2, 3, 3, 3 };
+ armnn::TensorInfo weightsTensorInfo(weightShape, DataType, inputsQScale);
+ weightsTensorInfo.SetConstant();
+ armnn::TensorInfo biasTensorInfo(biasShape, GetBiasDataType(DataType), inputsQScale);
+ biasTensorInfo.SetConstant();
+
+ layer->m_Weight = std::make_unique<ScopedTensorHandle>(weightsTensorInfo);
+ layer->m_Bias = std::make_unique<ScopedTensorHandle>(biasTensorInfo);
+
layer->m_Weight->Allocate();
layer->m_Bias->Allocate();
// Creates extra layers.
Layer* const input = graph.AddLayer<InputLayer>(0, "input");
+ auto const weights = graph.AddLayer<ConstantLayer>("weights");
+ auto const bias = graph.AddLayer<ConstantLayer>("bias");
Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
+ weights->m_LayerOutput = std::make_unique<ScopedTensorHandle>(weightsTensorInfo);
+ weights->m_LayerOutput->Allocate();
+ bias->m_LayerOutput = std::make_unique<ScopedTensorHandle>(biasTensorInfo);
+ bias->m_LayerOutput->Allocate();
+
// Connects up.
Connect(input, layer, TensorInfo({2, 3, 6, 6}, DataType, inputsQScale));
+ Connect(weights, layer, weightsTensorInfo, 0, 1);
+ Connect(bias, layer, biasTensorInfo, 0, 2);
Connect(layer, output, TensorInfo({2, 2, 6, 6}, DataType, outputQScale));
CreateTensorHandles(graph, factory);
@@ -1120,12 +1168,10 @@
CHECK(queueDescriptor.m_Parameters.m_PadBottom == 1);
CHECK(queueDescriptor.m_Parameters.m_BiasEnabled == true);
- CHECK(queueDescriptor.m_Inputs.size() == 1);
+ CHECK(queueDescriptor.m_Inputs.size() == 3);
CHECK(queueDescriptor.m_Outputs.size() == 1);
- CHECK((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo({2, 3, 3, 3},
- DataType, inputsQScale)));
- CHECK((queueDescriptor.m_Bias->GetTensorInfo()
- == TensorInfo({2}, GetBiasDataType(DataType), inputsQScale)));
+ CHECK((queueDescriptor.m_Weight->GetTensorInfo() == weightsTensorInfo));
+ CHECK((queueDescriptor.m_Bias->GetTensorInfo() == biasTensorInfo));
// Returns so we can do extra, backend-specific tests.
return workload;
@@ -2094,18 +2140,22 @@
armnn::ConstTensor biases(biasTensorInfo, biasData);
// Create convolution layer with biases
+ ARMNN_NO_DEPRECATE_WARN_BEGIN
convLayer = net->AddConvolution2dLayer(convDesc2d,
weights,
Optional<ConstTensor>(biases),
convLayerName.c_str());
+ ARMNN_NO_DEPRECATE_WARN_END
}
else
{
// Create convolution layer without biases
+ ARMNN_NO_DEPRECATE_WARN_BEGIN
convLayer = net->AddConvolution2dLayer(convDesc2d,
weights,
EmptyOptional(),
convLayerName.c_str());
+ ARMNN_NO_DEPRECATE_WARN_END
}
CHECK(convLayer);