IVGCVSW-4549 Add front end for new QLSTM layer
* Added new layer QLstm (Android R HAL 1.3)
* Made necessary updates to APIs
* Added unit tests
* This layer is functionally equivalent to the
original unquantized LSTM layer with some
additonal quantization features added. Due
to this, original LstmParams are used for
this layer.
Signed-off-by: James Conroy <james.conroy@arm.com>
Change-Id: I5b7f2d2fb6e17e81573b41a31bc55f49ae79608f
diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index c3c8421..c55f51d 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -447,6 +447,19 @@
return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
}
+bool LayerSupportBase::IsQLstmSupported(const TensorInfo& /*input*/,
+ const TensorInfo& /*previousOutputIn*/,
+ const TensorInfo& /*previousCellStateIn*/,
+ const TensorInfo& /*outputStateOut*/,
+ const TensorInfo& /*cellStateOut*/,
+ const TensorInfo& /*output*/,
+ const QLstmDescriptor& /*descriptor*/,
+ const LstmInputParamsInfo& /*paramsInfo*/,
+ Optional<std::string&> reasonIfUnsupported) const
+{
+ return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
bool LayerSupportBase::IsQuantizedLstmSupported(const TensorInfo& /*input*/,
const TensorInfo& /*previousCellStateIn*/,
const TensorInfo& /*previousOutputIn*/,
diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp
index 0639833..fcc3326 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -270,6 +270,16 @@
const TensorInfo& output,
Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+ bool IsQLstmSupported(const TensorInfo& input,
+ const TensorInfo& previousOutputIn,
+ const TensorInfo& previousCellStateIn,
+ const TensorInfo& outputStateOut,
+ const TensorInfo& cellStateOut,
+ const TensorInfo& output,
+ const QLstmDescriptor& descriptor,
+ const LstmInputParamsInfo& paramsInfo,
+ Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
bool IsQuantizedLstmSupported(const TensorInfo& input,
const TensorInfo& previousCellStateIn,
const TensorInfo& previousOutputIn,
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 85bda54..448de6a 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -519,6 +519,58 @@
void Validate(const WorkloadInfo& workloadInfo) const;
};
+struct QLstmQueueDescriptor : QueueDescriptorWithParameters<QLstmDescriptor>
+{
+ QLstmQueueDescriptor()
+ : 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_CellToInputWeights(nullptr)
+ , m_CellToForgetWeights(nullptr)
+ , m_CellToOutputWeights(nullptr)
+ , m_InputGateBias(nullptr)
+ , m_ForgetGateBias(nullptr)
+ , m_CellBias(nullptr)
+ , m_OutputGateBias(nullptr)
+ , m_ProjectionWeights(nullptr)
+ , m_ProjectionBias(nullptr)
+ , m_InputLayerNormWeights(nullptr)
+ , m_ForgetLayerNormWeights(nullptr)
+ , m_CellLayerNormWeights(nullptr)
+ , m_OutputLayerNormWeights(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_CellToInputWeights;
+ const ConstCpuTensorHandle* m_CellToForgetWeights;
+ const ConstCpuTensorHandle* m_CellToOutputWeights;
+ const ConstCpuTensorHandle* m_InputGateBias;
+ const ConstCpuTensorHandle* m_ForgetGateBias;
+ const ConstCpuTensorHandle* m_CellBias;
+ const ConstCpuTensorHandle* m_OutputGateBias;
+ const ConstCpuTensorHandle* m_ProjectionWeights;
+ const ConstCpuTensorHandle* m_ProjectionBias;
+ const ConstCpuTensorHandle* m_InputLayerNormWeights;
+ const ConstCpuTensorHandle* m_ForgetLayerNormWeights;
+ const ConstCpuTensorHandle* m_CellLayerNormWeights;
+ const ConstCpuTensorHandle* m_OutputLayerNormWeights;
+
+ void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
struct QuantizedLstmQueueDescriptor : QueueDescriptor
{
QuantizedLstmQueueDescriptor()
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index 5854bec..40ab798 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -749,6 +749,94 @@
result = layerSupportObject->IsQuantizeSupported(input, output, reason);
break;
}
+ case LayerType::QLstm:
+ {
+ auto cLayer = boost::polymorphic_downcast<const QLstmLayer*>(&layer);
+ const QLstmDescriptor& descriptor = cLayer->GetParameters();
+
+ // Inputs
+ const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+ const TensorInfo& previousOutputIn = layer.GetInputSlot(1).GetConnection()->GetTensorInfo();
+ const TensorInfo& previousCellStateIn = layer.GetInputSlot(2).GetConnection()->GetTensorInfo();
+
+ // Outputs
+ const TensorInfo& outputStateOut = layer.GetOutputSlot(0).GetTensorInfo();
+ const TensorInfo& cellStateOut = layer.GetOutputSlot(1).GetTensorInfo();
+ const TensorInfo& output = layer.GetOutputSlot(2).GetTensorInfo();
+
+ // Lstm parameters
+ LstmInputParamsInfo paramsInfo;
+
+ // Basic parameters
+ paramsInfo.m_InputToForgetWeights = &cLayer->m_BasicParameters.m_InputToForgetWeights->GetTensorInfo();
+ paramsInfo.m_InputToCellWeights = &cLayer->m_BasicParameters.m_InputToCellWeights->GetTensorInfo();
+ paramsInfo.m_InputToOutputWeights = &cLayer->m_BasicParameters.m_InputToOutputWeights->GetTensorInfo();
+
+ paramsInfo.m_RecurrentToForgetWeights =
+ &cLayer->m_BasicParameters.m_RecurrentToForgetWeights->GetTensorInfo();
+ paramsInfo.m_RecurrentToCellWeights =
+ &cLayer->m_BasicParameters.m_RecurrentToCellWeights->GetTensorInfo();
+ paramsInfo.m_RecurrentToOutputWeights =
+ &cLayer->m_BasicParameters.m_RecurrentToOutputWeights->GetTensorInfo();
+
+ paramsInfo.m_ForgetGateBias = &cLayer->m_BasicParameters.m_ForgetGateBias->GetTensorInfo();
+ paramsInfo.m_CellBias = &cLayer->m_BasicParameters.m_CellBias->GetTensorInfo();
+ paramsInfo.m_OutputGateBias = &cLayer->m_BasicParameters.m_OutputGateBias->GetTensorInfo();
+
+ if(!descriptor.m_CifgEnabled)
+ {
+ paramsInfo.m_InputToInputWeights = &cLayer->m_CifgParameters.m_InputToInputWeights->GetTensorInfo();
+ paramsInfo.m_RecurrentToInputWeights =
+ &cLayer->m_CifgParameters.m_RecurrentToInputWeights->GetTensorInfo();
+ paramsInfo.m_InputGateBias = &cLayer->m_CifgParameters.m_InputGateBias->GetTensorInfo();
+ }
+
+ if(descriptor.m_ProjectionEnabled)
+ {
+ paramsInfo.m_ProjectionWeights = &cLayer->m_ProjectionParameters.m_ProjectionWeights->GetTensorInfo();
+ paramsInfo.m_ProjectionBias = &cLayer->m_ProjectionParameters.m_ProjectionBias->GetTensorInfo();
+ }
+
+ if(descriptor.m_PeepholeEnabled)
+ {
+ if (!descriptor.m_CifgEnabled)
+ {
+ paramsInfo.m_CellToInputWeights =
+ &cLayer->m_PeepholeParameters.m_CellToInputWeights->GetTensorInfo();
+ }
+
+ paramsInfo.m_CellToForgetWeights =
+ &cLayer->m_PeepholeParameters.m_CellToForgetWeights->GetTensorInfo();
+ paramsInfo.m_CellToOutputWeights = &cLayer->m_PeepholeParameters.m_CellToOutputWeights->GetTensorInfo();
+ }
+
+ if(descriptor.m_LayerNormEnabled)
+ {
+ if (!descriptor.m_CifgEnabled)
+ {
+ paramsInfo.m_InputLayerNormWeights =
+ &cLayer->m_LayerNormParameters.m_InputLayerNormWeights->GetTensorInfo();
+ }
+
+ paramsInfo.m_ForgetLayerNormWeights =
+ &cLayer->m_LayerNormParameters.m_ForgetLayerNormWeights->GetTensorInfo();
+ paramsInfo.m_CellLayerNormWeights =
+ &cLayer->m_LayerNormParameters.m_CellLayerNormWeights->GetTensorInfo();
+ paramsInfo.m_OutputLayerNormWeights =
+ &cLayer->m_LayerNormParameters.m_OutputLayerNormWeights->GetTensorInfo();
+ }
+
+ result = layerSupportObject->IsQLstmSupported(input,
+ previousOutputIn,
+ previousCellStateIn,
+ outputStateOut,
+ cellStateOut,
+ output,
+ descriptor,
+ paramsInfo,
+ reason);
+ break;
+ }
case LayerType::QuantizedLstm:
{
auto cLayer = boost::polymorphic_downcast<const QuantizedLstmLayer*>(&layer);
@@ -1387,6 +1475,12 @@
return std::unique_ptr<IWorkload>();
}
+std::unique_ptr<IWorkload> IWorkloadFactory::CreateQLstm(const QLstmQueueDescriptor& /*descriptor*/,
+ const WorkloadInfo& /*info*/) const
+{
+ return std::unique_ptr<IWorkload>();
+}
+
std::unique_ptr<IWorkload> IWorkloadFactory::CreateQuantizedLstm(const QuantizedLstmQueueDescriptor& /*descriptor*/,
const WorkloadInfo& /*info*/) const
{
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index 0fc7ab9..98a6c36 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -197,6 +197,9 @@
virtual std::unique_ptr<IWorkload> CreateQuantize(const QuantizeQueueDescriptor& descriptor,
const WorkloadInfo& Info) const;
+ virtual std::unique_ptr<IWorkload> CreateQLstm(const QLstmQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const;
+
virtual std::unique_ptr<IWorkload> CreateQuantizedLstm(const QuantizedLstmQueueDescriptor& descriptor,
const WorkloadInfo& info) const;
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index d646847..7534c8a 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -308,6 +308,82 @@
{
};
+template <typename QLstmLayerType>
+struct DummyQLstmLayer
+{
+ DummyQLstmLayer()
+ {
+ typename QLstmLayerType::DescriptorType desc;
+ desc.m_CifgEnabled = false;
+ desc.m_PeepholeEnabled = true;
+ desc.m_ProjectionEnabled = true;
+ desc.m_LayerNormEnabled = true;
+
+ m_Layer = dummyGraph.AddLayer<QLstmLayerType>(armnn::QLstmDescriptor(), "qLstm");
+
+ // Basic params
+ m_Layer->m_BasicParameters.m_InputToForgetWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+ m_Layer->m_BasicParameters.m_InputToCellWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+ m_Layer->m_BasicParameters.m_InputToOutputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+
+ m_Layer->m_BasicParameters.m_RecurrentToForgetWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+ m_Layer->m_BasicParameters.m_RecurrentToCellWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+ m_Layer->m_BasicParameters.m_RecurrentToOutputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+
+ m_Layer->m_BasicParameters.m_ForgetGateBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+ m_Layer->m_BasicParameters.m_CellBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+ m_Layer->m_BasicParameters.m_OutputGateBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+
+ // CIFG optional params
+ m_Layer->m_CifgParameters.m_InputToInputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+ m_Layer->m_CifgParameters.m_RecurrentToInputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+ m_Layer->m_CifgParameters.m_InputGateBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+
+ // Projection optional params
+ m_Layer->m_ProjectionParameters.m_ProjectionWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS8));
+ m_Layer->m_ProjectionParameters.m_ProjectionBias = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Signed32));
+
+ // Peephole optional params
+ m_Layer->m_PeepholeParameters.m_CellToInputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS16));
+ m_Layer->m_PeepholeParameters.m_CellToForgetWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS16));
+ m_Layer->m_PeepholeParameters.m_CellToOutputWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS16));
+
+ // Layer normalization optional params
+ m_Layer->m_LayerNormParameters.m_InputLayerNormWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS16));
+ m_Layer->m_LayerNormParameters.m_ForgetLayerNormWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS16));
+ m_Layer->m_LayerNormParameters.m_CellLayerNormWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS16));
+ m_Layer->m_LayerNormParameters.m_OutputLayerNormWeights = std::make_unique<armnn::ScopedCpuTensorHandle>(
+ armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::QSymmS16));
+ }
+
+ ~DummyQLstmLayer()
+ {
+ dummyGraph.EraseLayer(m_Layer);
+ }
+
+ armnn::QLstmLayer* m_Layer;
+};
+
template<>
struct DummyLayer<armnn::QuantizedLstmLayer, void>
{
@@ -513,6 +589,8 @@
DECLARE_LAYER_POLICY_1_PARAM(Prelu)
+DECLARE_LAYER_POLICY_2_PARAM(QLstm)
+
DECLARE_LAYER_POLICY_1_PARAM(QuantizedLstm)
DECLARE_LAYER_POLICY_1_PARAM(Division)