aboutsummaryrefslogtreecommitdiff
path: root/src/backends/backendsCommon
diff options
context:
space:
mode:
Diffstat (limited to 'src/backends/backendsCommon')
-rw-r--r--src/backends/backendsCommon/LayerSupportBase.cpp11
-rw-r--r--src/backends/backendsCommon/LayerSupportBase.hpp8
-rw-r--r--src/backends/backendsCommon/WorkloadData.hpp37
-rw-r--r--src/backends/backendsCommon/WorkloadFactory.cpp76
-rw-r--r--src/backends/backendsCommon/WorkloadFactory.hpp3
-rw-r--r--src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp77
6 files changed, 207 insertions, 5 deletions
diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index e843423a92..f202fedb4f 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -347,6 +347,17 @@ bool LayerSupportBase::IsQuantizeSupported(const armnn::TensorInfo& input,
return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
}
+bool LayerSupportBase::IsQuantizedLstmSupported(const TensorInfo& input,
+ const TensorInfo& previousCellStateIn,
+ const TensorInfo& previousOutputIn,
+ const TensorInfo& cellStateOut,
+ const TensorInfo& output,
+ const QuantizedLstmInputParamsInfo& paramsInfo,
+ Optional<std::string&> reasonIfUnsupported) const
+{
+ return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
bool LayerSupportBase::IsReshapeSupported(const TensorInfo& input,
const ReshapeDescriptor& descriptor,
Optional<std::string&> reasonIfUnsupported) const
diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp
index d49fc3e371..c860e34874 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -214,6 +214,14 @@ public:
const TensorInfo& output,
Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+ bool IsQuantizedLstmSupported(const TensorInfo& input,
+ const TensorInfo& previousCellStateIn,
+ const TensorInfo& previousOutputIn,
+ const TensorInfo& cellStateOut,
+ const TensorInfo& output,
+ const QuantizedLstmInputParamsInfo& paramsInfo,
+ Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
bool IsReshapeSupported(const TensorInfo& input,
const ReshapeDescriptor& descriptor,
Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index f3d50699e6..d790dafd58 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -477,4 +477,41 @@ struct TransposeConvolution2dQueueDescriptor : QueueDescriptorWithParameters<Tra
void Validate(const WorkloadInfo& workloadInfo) const;
};
+struct QuantizedLstmQueueDescriptor : QueueDescriptor
+{
+ QuantizedLstmQueueDescriptor()
+ : m_InputToInputWeights(nullptr)
+ , m_InputToForgetWeights(nullptr)
+ , m_InputToCellWeights(nullptr)
+ , m_InputToOutputWeights(nullptr)
+
+ , m_RecurrentToInputWeights(nullptr)
+ , m_RecurrentToForgetWeights(nullptr)
+ , m_RecurrentToCellWeights(nullptr)
+ , m_RecurrentToOutputWeights(nullptr)
+
+ , m_InputGateBias(nullptr)
+ , m_ForgetGateBias(nullptr)
+ , m_CellBias(nullptr)
+ , m_OutputGateBias(nullptr)
+ {}
+
+ const ConstCpuTensorHandle* m_InputToInputWeights;
+ const ConstCpuTensorHandle* m_InputToForgetWeights;
+ const ConstCpuTensorHandle* m_InputToCellWeights;
+ const ConstCpuTensorHandle* m_InputToOutputWeights;
+
+ const ConstCpuTensorHandle* m_RecurrentToInputWeights;
+ const ConstCpuTensorHandle* m_RecurrentToForgetWeights;
+ const ConstCpuTensorHandle* m_RecurrentToCellWeights;
+ const ConstCpuTensorHandle* m_RecurrentToOutputWeights;
+
+ const ConstCpuTensorHandle* m_InputGateBias;
+ const ConstCpuTensorHandle* m_ForgetGateBias;
+ const ConstCpuTensorHandle* m_CellBias;
+ const ConstCpuTensorHandle* m_OutputGateBias;
+
+ void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
} //namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index a24a325b2d..cbaae4075c 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -631,6 +631,76 @@ bool IWorkloadFactory::IsLayerSupported(const BackendId& backendId,
result = layerSupportObject->IsQuantizeSupported(input, output, reason);
break;
}
+ case LayerType::QuantizedLstm:
+ {
+ auto cLayer = boost::polymorphic_downcast<const QuantizedLstmLayer*>(&layer);
+
+ // Inputs
+ const TensorInfo& input = OverrideDataType(
+ layer.GetInputSlot(0).GetConnection()->GetTensorInfo(), dataType);
+ const TensorInfo& previousCellStateIn = OverrideDataType(
+ layer.GetInputSlot(1).GetConnection()->GetTensorInfo(), dataType);
+ const TensorInfo& previousOutputIn = OverrideDataType(
+ layer.GetInputSlot(2).GetConnection()->GetTensorInfo(), dataType);
+
+ // Outputs
+ const TensorInfo& cellStateOut = OverrideDataType(layer.GetOutputSlot(0).GetTensorInfo(), dataType);
+ const TensorInfo& output = OverrideDataType(layer.GetOutputSlot(1).GetTensorInfo(), dataType);
+
+ // QuantizedLstm parameters
+ const TensorInfo& inputToInputWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_InputToInputWeights->GetTensorInfo(), dataType);
+ const TensorInfo& inputToForgetWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_InputToForgetWeights->GetTensorInfo(), dataType);
+ const TensorInfo& inputToCellWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_InputToCellWeights->GetTensorInfo(), dataType);
+ const TensorInfo& inputToOutputWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_InputToOutputWeights->GetTensorInfo(), dataType);
+
+ const TensorInfo& recurrentToInputWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_RecurrentToInputWeights->GetTensorInfo(), dataType);
+ const TensorInfo& recurrentToForgetWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_RecurrentToForgetWeights->GetTensorInfo(), dataType);
+ const TensorInfo& recurrentToCellWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_RecurrentToCellWeights->GetTensorInfo(), dataType);
+ const TensorInfo& recurrentToOutputWeights = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_RecurrentToOutputWeights->GetTensorInfo(), dataType);
+
+ const TensorInfo& inputGateBias = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_InputGateBias->GetTensorInfo(), dataType);
+ const TensorInfo& forgetGateBias = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_ForgetGateBias->GetTensorInfo(), dataType);
+ const TensorInfo& cellBias = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_CellBias->GetTensorInfo(), dataType);
+ const TensorInfo& outputGateBias = OverrideDataType(
+ cLayer->m_QuantizedLstmParameters.m_OutputGateBias->GetTensorInfo(), dataType);
+
+ QuantizedLstmInputParamsInfo paramsInfo;
+
+ paramsInfo.m_InputToInputWeights = &inputToInputWeights;
+ paramsInfo.m_InputToForgetWeights = &inputToForgetWeights;
+ paramsInfo.m_InputToCellWeights = &inputToCellWeights;
+ paramsInfo.m_InputToOutputWeights = &inputToOutputWeights;
+
+ paramsInfo.m_RecurrentToInputWeights = &recurrentToInputWeights;
+ paramsInfo.m_RecurrentToForgetWeights = &recurrentToForgetWeights;
+ paramsInfo.m_RecurrentToCellWeights = &recurrentToCellWeights;
+ paramsInfo.m_RecurrentToOutputWeights = &recurrentToOutputWeights;
+
+ paramsInfo.m_InputGateBias = &inputGateBias;
+ paramsInfo.m_ForgetGateBias = &forgetGateBias;
+ paramsInfo.m_CellBias = &cellBias;
+ paramsInfo.m_OutputGateBias = &outputGateBias;
+
+ result = layerSupportObject->IsQuantizedLstmSupported(input,
+ previousCellStateIn,
+ previousOutputIn,
+ cellStateOut,
+ output,
+ paramsInfo,
+ reason);
+ break;
+ }
case LayerType::Division:
{
const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
@@ -1109,6 +1179,12 @@ std::unique_ptr<IWorkload> IWorkloadFactory::CreateQuantize(const QuantizeQueueD
return std::unique_ptr<IWorkload>();
}
+std::unique_ptr<IWorkload> IWorkloadFactory::CreateQuantizedLstm(const QuantizedLstmQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const
+{
+ return std::unique_ptr<IWorkload>();
+}
+
std::unique_ptr<IWorkload> IWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor,
const WorkloadInfo& info) const
{
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index 749a258a9d..6d03da74fc 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -161,6 +161,9 @@ public:
virtual std::unique_ptr<IWorkload> CreateQuantize(const QuantizeQueueDescriptor& descriptor,
const WorkloadInfo& Info) const;
+ virtual std::unique_ptr<IWorkload> CreateQuantizedLstm(const QuantizedLstmQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const;
+
virtual std::unique_ptr<IWorkload> CreateReshape(const ReshapeQueueDescriptor& descriptor,
const WorkloadInfo& info) const;
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index 6aff7596b5..451c585adc 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -27,14 +27,17 @@ template<armnn::DataType DataType>
armnn::WorkloadInfo MakeDummyWorkloadInfo(unsigned int numInputs, unsigned int numOutputs)
{
armnn::WorkloadInfo info;
+
for (unsigned int i=0; i < numInputs; i++)
{
info.m_InputTensorInfos.push_back(MakeDummyTensorInfo<DataType>());
}
+
for (unsigned int o=0; o < numOutputs; o++)
{
info.m_OutputTensorInfos.push_back(MakeDummyTensorInfo<DataType>());
}
+
return info;
}
@@ -46,10 +49,12 @@ struct DummyLayer
{
m_Layer = dummyGraph.AddLayer<LayerType>(DescType(), "");
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
LayerType* m_Layer;
};
@@ -61,10 +66,12 @@ struct DummyLayer<LayerType, void>
{
m_Layer = dummyGraph.AddLayer<LayerType>("");
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
LayerType* m_Layer;
};
@@ -83,12 +90,13 @@ struct DummyLayer<armnn::BatchNormalizationLayer>
m_Layer->m_Gamma = std::make_unique<armnn::ScopedCpuTensorHandle>(
armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32));
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
- armnn::BatchNormalizationLayer* m_Layer;
+ armnn::BatchNormalizationLayer* m_Layer;
};
template<>
@@ -98,10 +106,12 @@ struct DummyLayer<armnn::BatchToSpaceNdLayer>
{
m_Layer = dummyGraph.AddLayer<armnn::BatchToSpaceNdLayer>(armnn::BatchToSpaceNdDescriptor(), "");
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::BatchToSpaceNdLayer* m_Layer;
};
@@ -112,10 +122,12 @@ struct DummyLayer<armnn::ConstantLayer, void>
{
m_Layer = dummyGraph.AddLayer<armnn::ConstantLayer>("");
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::ConstantLayer* m_Layer;
};
@@ -125,12 +137,13 @@ struct DummyLayer<armnn::InputLayer, armnn::LayerBindingId>
DummyLayer()
{
m_Layer = dummyGraph.AddLayer<armnn::InputLayer>(armnn::LayerBindingId(), "");
-
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::InputLayer* m_Layer;
};
@@ -141,12 +154,13 @@ struct DummyLayer<armnn::ConcatLayer>
{
armnn::OriginsDescriptor desc(2);
m_Layer = dummyGraph.AddLayer<armnn::ConcatLayer>(desc, "");
-
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::ConcatLayer* m_Layer;
};
@@ -156,12 +170,13 @@ struct DummyLayer<armnn::OutputLayer, armnn::LayerBindingId>
DummyLayer()
{
m_Layer = dummyGraph.AddLayer<armnn::OutputLayer>(armnn::LayerBindingId(), "");
-
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::OutputLayer* m_Layer;
};
@@ -172,12 +187,13 @@ struct DummyLayer<armnn::SplitterLayer>
{
armnn::ViewsDescriptor desc(1);
m_Layer = dummyGraph.AddLayer<armnn::SplitterLayer>(desc, "");
-
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::SplitterLayer* m_Layer;
};
@@ -193,10 +209,12 @@ struct DummyConvolutionLayer
m_Layer->m_Bias = std::make_unique<armnn::ScopedCpuTensorHandle>(
armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32));
}
+
~DummyConvolutionLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
ConvolutionLayerType* m_Layer;
};
@@ -255,10 +273,12 @@ struct DummyLstmLayer
m_Layer->m_CifgParameters.m_InputGateBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32));
}
+
~DummyLstmLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::LstmLayer* m_Layer;
};
@@ -269,6 +289,49 @@ struct DummyLayer<armnn::LstmLayer>
};
template<>
+struct DummyLayer<armnn::QuantizedLstmLayer, void>
+{
+ DummyLayer()
+ {
+ m_Layer = dummyGraph.AddLayer<armnn::QuantizedLstmLayer>("");
+
+ m_Layer->m_QuantizedLstmParameters.m_InputToInputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+ m_Layer->m_QuantizedLstmParameters.m_InputToForgetWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+ m_Layer->m_QuantizedLstmParameters.m_InputToCellWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+ m_Layer->m_QuantizedLstmParameters.m_InputToOutputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+
+ m_Layer->m_QuantizedLstmParameters.m_RecurrentToInputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+ m_Layer->m_QuantizedLstmParameters.m_RecurrentToForgetWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+ m_Layer->m_QuantizedLstmParameters.m_RecurrentToCellWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+ m_Layer->m_QuantizedLstmParameters.m_RecurrentToOutputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QuantisedAsymm8));
+
+ m_Layer->m_QuantizedLstmParameters.m_InputGateBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+ m_Layer->m_QuantizedLstmParameters.m_ForgetGateBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+ m_Layer->m_QuantizedLstmParameters.m_CellBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+ m_Layer->m_QuantizedLstmParameters.m_OutputGateBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+ }
+
+ ~DummyLayer()
+ {
+ dummyGraph.EraseLayer(m_Layer);
+ }
+
+ armnn::QuantizedLstmLayer* m_Layer;
+};
+
+template<>
struct DummyLayer<armnn::FullyConnectedLayer>
{
DummyLayer()
@@ -278,10 +341,12 @@ struct DummyLayer<armnn::FullyConnectedLayer>
m_Layer->m_Weight = std::make_unique<armnn::ScopedCpuTensorHandle>(
armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32));
}
+
~DummyLayer()
{
dummyGraph.EraseLayer(m_Layer);
}
+
armnn::FullyConnectedLayer* m_Layer;
};
@@ -392,6 +457,8 @@ DECLARE_LAYER_POLICY_2_PARAM(PreCompiled)
DECLARE_LAYER_POLICY_1_PARAM(Prelu)
+DECLARE_LAYER_POLICY_1_PARAM(QuantizedLstm)
+
DECLARE_LAYER_POLICY_1_PARAM(Division)
DECLARE_LAYER_POLICY_2_PARAM(Resize)
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-26 12:13:16 +0000