IVGCVSW-6954 'Arm NN SL Improvements'

* Move the Conv2D and DepthwiseConv2D validation to Optimization level
  when the weights and tensors are as constant inputs
* Take into account offset and scales values when doing INT8 to FP32 dequantization

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I1f81f15640395ac041923b10dbe9151159715117

3 files changed, 92 insertions, 35 deletions

diff --git a/shim/sl/canonical/ConversionUtils.cpp b/shim/sl/canonical/ConversionUtils.cpp
index 96a8ddca6a..f48af32e21 100644
--- a/shim/sl/canonical/ConversionUtils.cpp
+++ b/shim/sl/canonical/ConversionUtils.cpp
@@ -67,6 +67,11 @@ void LayerInputHandle::SanitizeQuantizationScale(LayerInputHandle& weight, Layer
     }
 }
 
+armnn::IOutputSlot* LayerInputHandle::GetOutputSlot() const
+{
+    return m_OutputSlot;
+}
+
 ConstTensorPin::ConstTensorPin(bool optional)
     : m_Optional(optional)
 {}
@@ -276,17 +281,6 @@ LayerInputHandle ConvertToLayerInputHandle(const Operation& operation,
             case OperandLifeTime::CONSTANT_REFERENCE:
             {
                 auto constantTensorDataType = operandTensorInfo.GetDataType();
-                if (inputHandle)
-                {
-                    if ((inputHandle->GetTensorInfo().GetDataType() == armnn::DataType::Float32
-                         || inputHandle->GetTensorInfo().GetDataType() == armnn::DataType::Float16)
-                        && (operandTensorInfo.GetDataType() == armnn::DataType::QAsymmU8
-                            || operandTensorInfo.GetDataType() == armnn::DataType::QAsymmS8))
-                    {
-                        constantTensorDataType = inputHandle->GetTensorInfo().GetDataType();
-                    }
-                }
-
                 // The tensor has an already known constant value, and can be converted into an ArmNN Constant layer.
                 ConstTensorPin tensorPin = ConvertOperandToConstTensorPin(*operand,
                                                                           model,
@@ -1029,4 +1023,34 @@ bool SetupAndTrackLayerOutputSlot(const Operation& operation,
     return true;
 }
 
+bool IsConnectedToDequantize(armnn::IOutputSlot* ioutputSlot)
+{
+    VLOG(DRIVER) << "ConversionUtils::IsConnectedToDequantize()";
+    if (!ioutputSlot)
+    {
+        return false;
+    }
+    VLOG(DRIVER) << "ConversionUtils::IsConnectedToDequantize() ioutputSlot is valid.";
+    // Find the connections and layers..
+    armnn::IConnectableLayer& owningLayer = ioutputSlot->GetOwningIConnectableLayer();
+    if (owningLayer.GetType() == armnn::LayerType::Dequantize)
+    {
+        VLOG(DRIVER) << "ConversionUtils::IsConnectedToDequantize() connected to Dequantize Layer.";
+        armnn::IInputSlot& inputSlot = owningLayer.GetInputSlot(0);
+        armnn::IOutputSlot* connection = inputSlot.GetConnection();
+        if (connection)
+        {
+            VLOG(DRIVER) << "ConversionUtils::IsConnectedToDequantize() Dequantize Layer has a connection.";
+            armnn::IConnectableLayer& connectedLayer =
+                    connection->GetOwningIConnectableLayer();
+            if (connectedLayer.GetType() == armnn::LayerType::Constant)
+            {
+                VLOG(DRIVER) << "ConversionUtils::IsConnectedToDequantize() Dequantize Layer connected to Constant";
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
 } // namespace armnn_driver
diff --git a/shim/sl/canonical/ConversionUtils.hpp b/shim/sl/canonical/ConversionUtils.hpp
index 8058bcb379..beee00d11a 100644
--- a/shim/sl/canonical/ConversionUtils.hpp
+++ b/shim/sl/canonical/ConversionUtils.hpp
@@ -79,6 +79,8 @@ public:
 
     void SanitizeQuantizationScale(LayerInputHandle& weight, LayerInputHandle& input);
 
+    armnn::IOutputSlot* GetOutputSlot() const;
+
 private:
     armnn::IOutputSlot* m_OutputSlot;
     bool                m_Valid;
@@ -1012,4 +1014,6 @@ ConstTensorPin DequantizeAndMakeConstTensorPin(const Operation& operation,
                                                size_t operandIndex,
                                                bool optional = false);
 
+bool IsConnectedToDequantize(armnn::IOutputSlot* ioutputSlot);
+
 } // namespace armnn_driver
diff --git a/shim/sl/canonical/Converter.cpp b/shim/sl/canonical/Converter.cpp
index fc983dc081..b50b0a9397 100644
--- a/shim/sl/canonical/Converter.cpp
+++ b/shim/sl/canonical/Converter.cpp
@@ -998,9 +998,20 @@ bool Converter::ConvertConv2d(const Operation& operation, const Model& model, Co
     desc.m_BiasEnabled = true;
     Optional<TensorInfo> biases(biasInfo);
 
-    bool isSupported = false;
-    auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported)
+    bool requiresValidation = true;
+    const Operand* weightsOperand = GetInputOperand(operation, 1, model);
+    const Operand* biasOperand = GetInputOperand(operation, 2, model);
+    if (IsConnectedToDequantize(weightsInput.GetOutputSlot())
+                || IsConnectedToDequantize(biasInput.GetOutputSlot()))
     {
+        // Do not require validation for now. There will be an optimization step
+        // [ConvertConstDequantisationLayersToConstLayers] will convert layers to Constant layers
+        // then at the end of the optimization there will be layer supported validation.
+        requiresValidation = false;
+        VLOG(DRIVER) << "Converter::ConvertConv2d(): Weights and Biases are as INPUTS.";
+    }
+
+    auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) {
         FORWARD_LAYER_SUPPORT_FUNC(__func__,
                                    IsConvolution2dSupported,
                                    data.m_Backends,
@@ -1012,18 +1023,23 @@ bool Converter::ConvertConv2d(const Operation& operation, const Model& model, Co
                                    biases);
     };
 
-    if(!IsDynamicTensor(outputInfo))
+    if (requiresValidation)
     {
-        validateFunc(outputInfo, isSupported);
-    }
-    else
-    {
-        isSupported = AreDynamicTensorsSupported();
-    }
+        VLOG(DRIVER) << "Converter::ConvertConv2d(): Requires Validation!";
+        bool isSupported = false;
+        if (!IsDynamicTensor(outputInfo))
+        {
+            validateFunc(outputInfo, isSupported);
+        }
+        else
+        {
+            isSupported = AreDynamicTensorsSupported();
+        }
 
-    if (!isSupported)
-    {
-        return false;
+        if (!isSupported)
+        {
+            return false;
+        }
     }
 
     armnn::IConnectableLayer* startLayer = data.m_Network->AddConvolution2dLayer(desc);
@@ -1231,9 +1247,17 @@ bool Converter::ConvertDepthwiseConv2d(const Operation& operation, const Model&
     desc.m_BiasEnabled = true;
     Optional<TensorInfo> biases(biasInfo);
 
-    bool isSupported = false;
-    auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported)
+    bool requiresValidation = true;
+    if (IsConnectedToDequantize(weightsInput.GetOutputSlot()) || IsConnectedToDequantize(biasInput.GetOutputSlot()))
     {
+        // Do not require validation for now. There will be an optimization step
+        // [ConvertConstDequantisationLayersToConstLayers] will convert layers to Constant layers
+        // then at the end of the optimization there will be layer supported validation.
+        requiresValidation = false;
+        VLOG(DRIVER) << "Converter::ConvertDepthwiseConv2d(): Weights and Biases are as INPUTS.";
+    }
+
+    auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) {
         FORWARD_LAYER_SUPPORT_FUNC(__func__,
                                    IsDepthwiseConvolutionSupported,
                                    data.m_Backends,
@@ -1245,18 +1269,23 @@ bool Converter::ConvertDepthwiseConv2d(const Operation& operation, const Model&
                                    biases);
     };
 
-    if(!IsDynamicTensor(outputInfo))
+    if (requiresValidation)
     {
-        validateFunc(outputInfo, isSupported);
-    }
-    else
-    {
-        isSupported = AreDynamicTensorsSupported();
-    }
+        VLOG(DRIVER) << "Converter::ConvertDepthwiseConv2d(): Requires Validation!";
+        bool isSupported = false;
+        if (!IsDynamicTensor(outputInfo))
+        {
+            validateFunc(outputInfo, isSupported);
+        }
+        else
+        {
+            isSupported = AreDynamicTensorsSupported();
+        }
 
-    if (!isSupported)
-    {
-        return false;
+        if (!isSupported)
+        {
+            return false;
+        }
     }
 
     armnn::IConnectableLayer* startLayer = data.m_Network->AddDepthwiseConvolution2dLayer(desc);