diff options
Diffstat (limited to 'src/backends/backendsCommon')
4 files changed, 211 insertions, 142 deletions
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp index 3fe0823b03..319cdb106b 100644 --- a/src/backends/backendsCommon/WorkloadData.cpp +++ b/src/backends/backendsCommon/WorkloadData.cpp @@ -1041,15 +1041,12 @@ void FullyConnectedQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) c { const std::string descriptorName{"FullyConnectedQueueDescriptor"}; - uint32_t numInputs = 1; - if (!m_Parameters.m_ConstantWeights) + uint32_t numInputs = 2; + if (m_Parameters.m_BiasEnabled) { - numInputs = 2; - if (m_Parameters.m_BiasEnabled) - { - numInputs = 3; - } + numInputs = 3; } + ValidateNumInputs(workloadInfo, descriptorName, numInputs); ValidateNumOutputs(workloadInfo, descriptorName, 1); @@ -1063,30 +1060,12 @@ void FullyConnectedQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) c throw InvalidArgumentException(descriptorName + ": Input tensor must have 2 or 4 dimensions."); } - TensorInfo weightTensorInfo; - if (m_Parameters.m_ConstantWeights) - { - ValidatePointer(m_Weight, descriptorName, "weight"); - weightTensorInfo = m_Weight->GetTensorInfo(); - } - else - { - weightTensorInfo = workloadInfo.m_InputTensorInfos[1]; - } + TensorInfo weightTensorInfo = workloadInfo.m_InputTensorInfos[1]; ValidateTensorNumDimensions(weightTensorInfo, descriptorName, 2, "weight"); if (m_Parameters.m_BiasEnabled) { - TensorInfo biasTensorInfo; - if (m_Parameters.m_ConstantWeights) - { - ValidatePointer(m_Bias, descriptorName, "bias"); - biasTensorInfo = m_Bias->GetTensorInfo(); - } - else - { - biasTensorInfo = workloadInfo.m_InputTensorInfos[2]; - } + TensorInfo biasTensorInfo = workloadInfo.m_InputTensorInfos[2]; // Validates type and quantization values. ValidateBiasTensorQuantization(biasTensorInfo, inputTensorInfo, weightTensorInfo, descriptorName); ValidateTensorDataType(biasTensorInfo, GetBiasDataType(inputTensorInfo.GetDataType()), descriptorName, "bias"); @@ -1894,11 +1873,9 @@ void FloorQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const }; ValidateDataTypes(inputTensorInfo, supportedTypes, descriptorName); - - if (inputTensorInfo != outputTensorInfo) - { - throw InvalidArgumentException(descriptorName + ": Input and output tensor infos do not match."); - } + ValidateTensorDataTypesMatch(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output"); + ValidateTensorShapesMatch(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output"); + ValidateTensorQuantizationSpace(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output"); } void LstmQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp index 1c18551679..3f5972dab6 100644 --- a/src/backends/backendsCommon/WorkloadFactory.cpp +++ b/src/backends/backendsCommon/WorkloadFactory.cpp @@ -36,7 +36,11 @@ const TensorInfo OverrideDataType(const TensorInfo& info, Optional<DataType> typ return info; } - return TensorInfo(info.GetShape(), type.value(), info.GetQuantizationScale(), info.GetQuantizationOffset()); + return TensorInfo(info.GetShape(), + type.value(), + info.GetQuantizationScale(), + info.GetQuantizationOffset(), + info.IsConstant()); } } // anonymous namespace @@ -364,16 +368,7 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, TensorInfo weightsInfo; const TensorInfo* weightsInfoPtr = nullptr; - if (descriptor.m_ConstantWeights) - { - ARMNN_ASSERT(cLayer->m_Weight.get() != nullptr); - weightsInfo = OverrideDataType(cLayer->m_Weight->GetTensorInfo(), dataType); - } - else - { - weightsInfo = OverrideDataType(layer.GetInputSlot(1).GetConnection()->GetTensorInfo(), dataType); - - } + weightsInfo = OverrideDataType(layer.GetInputSlot(1).GetConnection()->GetTensorInfo(), dataType); weightsInfoPtr = &weightsInfo; TensorInfo biasInfo; @@ -385,17 +380,8 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId, if (descriptor.m_BiasEnabled) { - if(descriptor.m_ConstantWeights) - { - ARMNN_ASSERT(cLayer->m_Bias.get() != nullptr); - biasInfo = OverrideDataType(cLayer->m_Bias->GetTensorInfo(), GetBiasTypeFromWeightsType(dataType)); - biasInfoPtr = &biasInfo; - } - else - { - biasInfo = OverrideDataType(layer.GetInputSlot(2).GetConnection()->GetTensorInfo(), dataType); - biasInfoPtr = &biasInfo; - } + biasInfo = OverrideDataType(layer.GetInputSlot(2).GetConnection()->GetTensorInfo(), dataType); + biasInfoPtr = &biasInfo; } else { diff --git a/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp b/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp index 923d6f3641..af6b56852a 100644 --- a/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp +++ b/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp @@ -28,10 +28,7 @@ armnn::INetworkPtr CreateFullyConnectedNetworkNonConstWeights(const armnn::Tenso armnn::IConnectableLayer* inputLayer = network->AddInputLayer(0, "Input"); armnn::IConnectableLayer* weightsInputLayer = network->AddInputLayer(1, "Weights_Input"); - armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, - armnn::EmptyOptional(), - armnn::EmptyOptional(), - "Fully_Connected"); + armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, "Fully_Connected"); armnn::IConnectableLayer* outputLayer = network->AddOutputLayer(0, "Output"); Connect(inputLayer, fullyConnectedLayer, inputTensorInfo, 0, 0); @@ -41,6 +38,52 @@ armnn::INetworkPtr CreateFullyConnectedNetworkNonConstWeights(const armnn::Tenso return network; } +armnn::INetworkPtr CreateFullyConnectedNetworkNonConstWeightsConstBias(const armnn::TensorInfo& inputTensorInfo, + const armnn::TensorInfo& outputTensorInfo, + const armnn::TensorInfo& weightsTensorInfo, + const armnn::TensorInfo& biasTensorInfo, + const armnn::ConstTensor& biasConstantTensor, + armnn::FullyConnectedDescriptor descriptor) +{ + armnn::INetworkPtr network(armnn::INetwork::Create()); + + armnn::IConnectableLayer* inputLayer = network->AddInputLayer(0, "Input"); + armnn::IConnectableLayer* weightsInputLayer = network->AddInputLayer(1, "Weights_Input"); + armnn::IConnectableLayer* biasLayer = network->AddConstantLayer(biasConstantTensor, "Weights"); + armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, "Fully_Connected"); + armnn::IConnectableLayer* outputLayer = network->AddOutputLayer(0, "Output"); + + Connect(inputLayer, fullyConnectedLayer, inputTensorInfo, 0, 0); + Connect(weightsInputLayer, fullyConnectedLayer, weightsTensorInfo, 0, 1); + Connect(biasLayer, fullyConnectedLayer, biasTensorInfo, 0, 2); + Connect(fullyConnectedLayer, outputLayer, outputTensorInfo, 0, 0); + + return network; +} + +armnn::INetworkPtr CreateFullyConnectedNetworkConstWeightsNonConstBias(const armnn::TensorInfo& inputTensorInfo, + const armnn::TensorInfo& outputTensorInfo, + const armnn::TensorInfo& weightsTensorInfo, + const armnn::TensorInfo& biasTensorInfo, + const armnn::ConstTensor& weightsConstantTensor, + armnn::FullyConnectedDescriptor descriptor) +{ + armnn::INetworkPtr network(armnn::INetwork::Create()); + + armnn::IConnectableLayer* inputLayer = network->AddInputLayer(0, "Input"); + armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weightsConstantTensor, "Weights"); + armnn::IConnectableLayer* biasLayer = network->AddInputLayer(2, "Bias_Input"); + armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, "Fully_Connected"); + armnn::IConnectableLayer* outputLayer = network->AddOutputLayer(0, "Output"); + + Connect(inputLayer, fullyConnectedLayer, inputTensorInfo, 0, 0); + Connect(weightsLayer, fullyConnectedLayer, weightsTensorInfo, 0, 1); + Connect(biasLayer, fullyConnectedLayer, biasTensorInfo, 0, 2); + Connect(fullyConnectedLayer, outputLayer, outputTensorInfo, 0, 0); + + return network; +} + template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> void FullyConnectedWithDynamicWeightsEndToEnd(const std::vector<armnn::BackendId>& backends) { @@ -94,4 +137,123 @@ void FullyConnectedWithDynamicWeightsEndToEnd(const std::vector<armnn::BackendId backends, 1.0f); } + +template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> +void FullyConnectedWithDynamicOrConstantInputsEndToEnd(const std::vector<armnn::BackendId>& backends, + const bool transposeWeights, + const bool constantWeightsOrBias) +{ + unsigned int inputWidth = 1; + unsigned int inputHeight = 1; + unsigned int inputChannels = 5; + unsigned int inputNum = 2; + + unsigned int outputChannels = 3; + unsigned int outputNum = 2; + + unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth }; + unsigned int outputShape[] = { outputNum, outputChannels }; + unsigned int weightsShape[] = { inputChannels, outputChannels }; + + if (transposeWeights) + { + std::swap(weightsShape[0], weightsShape[1]); + } + + unsigned int biasShape[] = { outputChannels }; + + armnn::TensorInfo inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32); + armnn::TensorInfo outputTensorInfo = armnn::TensorInfo(2, outputShape, armnn::DataType::Float32); + armnn::TensorInfo weightsDesc = armnn::TensorInfo(2, weightsShape, armnn::DataType::Float32); + armnn::TensorInfo biasesDesc = armnn::TensorInfo(1, biasShape, armnn::DataType::Float32); + + std::vector<float> input = + { + 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, + 5.0f, 4.0f, 3.0f, 2.0f, 1.0f + }; + + std::vector<float> weights = + { + .5f, 2.f, .5f, + .5f, 2.f, 1.f, + .5f, 2.f, 2.f, + .5f, 2.f, 3.f, + .5f, 2.f, 4.f + }; + + if (transposeWeights) + { + weights = + { + .5f, .5f, .5f, .5f, .5f, + 2.f, 2.f, 2.f, 2.f, 2.f, + .5f, 1.f, 2.f, 3.f, 4.f + }; + } + + std::vector<float> biasValues = std::vector<float>({10.f, 20.f, 30.f}); + + std::vector<float> expectedOutput = + { + 0.5f + 1.0f + 1.5f + 2.0f + 2.5f + biasValues[0], + 2.0f + 4.0f + 6.0f + 8.0f + 10.f + biasValues[1], + 0.5f + 2.0f + 6.0f + 12.f + 20.f + biasValues[2], + + 2.5f + 2.0f + 1.5f + 1.0f + 0.5f + biasValues[0], + 10.0f + 8.0f + 6.0f + 4.0f + 2.f + biasValues[1], + 2.5f + 4.0f + 6.0f + 6.f + 4.f + biasValues[2] + }; + + FullyConnectedDescriptor descriptor; + descriptor.m_BiasEnabled = true; + descriptor.m_TransposeWeightMatrix = transposeWeights; + descriptor.m_ConstantWeights = constantWeightsOrBias; + + if (!constantWeightsOrBias) + { + // Tests non constant weights and constant bias. + ConstTensor biasConstantTensor(biasesDesc, biasValues.data()); + + armnn::INetworkPtr network = CreateFullyConnectedNetworkNonConstWeightsConstBias(inputTensorInfo, + outputTensorInfo, + weightsDesc, + biasesDesc, + biasConstantTensor, + descriptor); + CHECK(network); + + std::map<int, std::vector<T>> inputTensorData = {{ 0, input }, {1, weights}}; + std::map<int, std::vector<T>> expectedOutputTensorData = {{ 0, expectedOutput }}; + + EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(network), + inputTensorData, + expectedOutputTensorData, + backends, + 1.0f); + } + else + { + // Tests constant weights and non constant bias. + ConstTensor weightsConstantTensor(weightsDesc, weights.data()); + + armnn::INetworkPtr network = CreateFullyConnectedNetworkConstWeightsNonConstBias(inputTensorInfo, + outputTensorInfo, + weightsDesc, + biasesDesc, + weightsConstantTensor, + descriptor); + CHECK(network); + + std::map<int, std::vector<T>> inputTensorData = {{ 0, input }, {2, biasValues}}; + std::map<int, std::vector<T>> expectedOutputTensorData = {{ 0, expectedOutput }}; + + EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(network), + inputTensorData, + expectedOutputTensorData, + backends, + 1.0f); + } +} + } // anonymous namespace diff --git a/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp index c47048e566..dcf87fe92b 100644 --- a/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp +++ b/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp @@ -22,56 +22,6 @@ template<typename T, typename B> LayerTestResult<T, 2> SimpleFullyConnectedTestImpl( - armnn::IWorkloadFactory& workloadFactory, - const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, - const armnn::ITensorHandleFactory& tensorHandleFactory, - armnn::TensorInfo inputTensorInfo, - armnn::TensorInfo outputTensorInfo, - armnn::TensorInfo weightsDesc, - armnn::TensorInfo biasesDesc, - std::vector<T>& weights, - std::vector<B>& bias, - std::vector<T>& input, - bool biasEnabled, - bool transposeWeights) -{ - std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); - std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo); - - armnn::FullyConnectedQueueDescriptor data; - armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(weightsDesc); - armnn::ScopedTensorHandle biasTensor(biasesDesc); - - std::vector<T> actualOutput(outputTensorInfo.GetNumElements()); - - AllocateAndCopyDataToITensorHandle(&weightsTensor, weights.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); - - AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); - AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; - data.m_Parameters.m_BiasEnabled = biasEnabled; - data.m_Parameters.m_TransposeWeightMatrix = transposeWeights; - - std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateFullyConnected(data, info); - LayerTestResult<T, 2> result(outputTensorInfo); - - inputHandle->Allocate(); - outputHandle->Allocate(); - CopyDataToITensorHandle(inputHandle.get(), input.data()); - - ExecuteWorkload(*workload, memoryManager); - - CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get()); - result.m_ActualData = actualOutput; - - return result; -} - -template<typename T, typename B> -LayerTestResult<T, 2> SimpleFullyConnectedTestWeightsAsInputsImpl( armnn::IWorkloadFactory& workloadFactory, const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, const armnn::ITensorHandleFactory& tensorHandleFactory, @@ -83,7 +33,8 @@ LayerTestResult<T, 2> SimpleFullyConnectedTestWeightsAsInputsImpl( std::vector<B>& bias, std::vector<T>& input, bool biasEnabled, - bool transposeWeights) + bool transposeWeights, + bool constantWeights) { std::unique_ptr<armnn::ITensorHandle> input0Handle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); std::unique_ptr<armnn::ITensorHandle> input1Handle = tensorHandleFactory.CreateTensorHandle(weightsTensorInfo); @@ -93,13 +44,23 @@ LayerTestResult<T, 2> SimpleFullyConnectedTestWeightsAsInputsImpl( armnn::FullyConnectedQueueDescriptor data; armnn::WorkloadInfo info; + armnn::ScopedTensorHandle weightsTensor(weightsTensorInfo); + armnn::ScopedTensorHandle biasTensor(biasesTensorInfo); + + AllocateAndCopyDataToITensorHandle(&weightsTensor, weights.data()); + AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); AddInputToWorkload(data, info, inputTensorInfo, input0Handle.get()); AddInputToWorkload(data, info, weightsTensorInfo, input1Handle.get()); AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + // Need to set as layer members will be null when creating the workload because the optimization hasn't been run. + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; + data.m_Parameters.m_BiasEnabled = biasEnabled; data.m_Parameters.m_TransposeWeightMatrix = transposeWeights; - data.m_Parameters.m_ConstantWeights = false; + data.m_Parameters.m_ConstantWeights = constantWeights; std::unique_ptr<armnn::ITensorHandle> input2Handle = nullptr; if (biasEnabled) @@ -180,36 +141,19 @@ LayerTestResult<T, 2> FullyConnectedTest( std::vector<int32_t> bias = {9250, 67500}; - if (constantWeights) - { - result = SimpleFullyConnectedTestImpl<T>(workloadFactory, - memoryManager, - tensorHandleFactory, - inputTensorInfo, - outputTensorInfo, - weightsDesc, - biasesDesc, - weights, - bias, - input, - biasEnabled, - true); - } - else - { - result = SimpleFullyConnectedTestWeightsAsInputsImpl<T>(workloadFactory, - memoryManager, - tensorHandleFactory, - inputTensorInfo, - outputTensorInfo, - weightsDesc, - biasesDesc, - weights, - bias, - input, - biasEnabled, - true); - } + result = SimpleFullyConnectedTestImpl<T>(workloadFactory, + memoryManager, + tensorHandleFactory, + inputTensorInfo, + outputTensorInfo, + weightsDesc, + biasesDesc, + weights, + bias, + input, + biasEnabled, + true, + constantWeights); if (biasEnabled) { @@ -299,7 +243,7 @@ LayerTestResult<T, 2> FullyConnectedLargeTestCommon( inputTensorInfo, outputTensorInfo, weightsDesc, biasesDesc, weights, biasValues, input, - true, transposeWeights + true, transposeWeights, true ); result.m_ExpectedData = armnnUtils::QuantizedVector<T>({ 965432.0f }, qScale, qOffset); @@ -408,7 +352,7 @@ LayerTestResult<float, 2> FullyConnectedFloat32Test( inputTensorInfo, outputTensorInfo, weightsDesc, biasesDesc, weights, biasValues, input, - biasEnabled, transposeWeights + biasEnabled, transposeWeights, true ); std::vector<float> expectedOutput = |