GitHub #667: Neon fold padding into average pool 2D quantization bug fix.

  * Originated from a GitHub issue: https://github.com/ARM-software/armnn/issues/667
  * Initially, Arm NN supports the pool 2D operation because there is no padding
    on the pool2d. Neon failure occurs when padding is followed by average pool 2D
    due to folding optimization.
  * Here we prevent the folding optimization from happening for the above special case
    and add it in as a backend specific optimization.

Signed-off-by: Cathal Corbett <cathal.corbett@arm.com>
Change-Id: Ia0fd90c3a6b4b9d29c81106f154617d2e893e26b

9 files changed, 391 insertions, 65 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 1d8ebe2952..7061c603a2 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -544,6 +544,7 @@ if(BUILD_UNIT_TESTS)
         src/armnn/test/optimizations/ConvertConstantsBFloatTests.cpp
         src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp
         src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp
+        src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp
         src/armnn/test/optimizations/FoldPadTests.cpp
         src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp
         src/armnn/test/optimizations/Fp32NetworkToFp16ConverterTests.cpp
diff --git a/src/armnn/optimizations/FoldPadIntoLayer2d.hpp b/src/armnn/optimizations/FoldPadIntoLayer2d.hpp
index eb6bc90afd..4c4bd80d41 100644
--- a/src/armnn/optimizations/FoldPadIntoLayer2d.hpp
+++ b/src/armnn/optimizations/FoldPadIntoLayer2d.hpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2017 Arm Ltd. All rights reserved.
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -73,6 +73,17 @@ inline bool IsNeutralElement(
         : tensorValue == GetZeroElement(tensorInfo);
 }
 
+inline bool IsPooling2dPadded(const Pooling2dDescriptor& poolDescriptor)
+{
+    const auto poolingPadValues = std::make_tuple(poolDescriptor.m_PadLeft, poolDescriptor.m_PadRight,
+                                                  poolDescriptor.m_PadTop, poolDescriptor.m_PadBottom);
+    if (poolingPadValues != std::make_tuple(0U, 0U, 0U, 0U))
+    {
+        return true;
+    }
+    return false;
+}
+
 template <typename Descriptor>
 bool TryFoldPadIntoLayer2d(
     const PadDescriptor& padDescriptor, Descriptor& layerDescriptor, const TensorInfo& tensorInfo)
@@ -101,25 +112,29 @@ bool TryFoldPadIntoLayer2d(
     return true;
 }
 
-inline bool TryFoldPadIntoLayer2d(
-    const PadDescriptor& padDescriptor, Pooling2dDescriptor& poolDescriptor, const TensorInfo& tensorInfo)
+inline bool TryFoldPadIntoLayer2d(const PadDescriptor& padDescriptor,
+                                  Pooling2dDescriptor& poolDescriptor,
+                                  const TensorInfo& tensorInfo,
+                                  bool isBackendOptimization = false)
 {
-    const auto poolingPadValues = std::make_tuple(poolDescriptor.m_PadLeft, poolDescriptor.m_PadRight,
-                                                  poolDescriptor.m_PadTop, poolDescriptor.m_PadBottom);
-    bool poolHasPadding = false;
-    if (poolingPadValues != std::make_tuple(0U, 0U, 0U, 0U))
+    // Cannot fold Average or L2 pooling if padding exists and the padding method is Exclude.
+    if (poolDescriptor.m_PoolType != PoolingAlgorithm::Max &&
+        IsPooling2dPadded(poolDescriptor) &&
+        poolDescriptor.m_PaddingMethod == PaddingMethod::Exclude)
     {
-        poolHasPadding = true;
+        return false;
     }
 
-    // We cannot fold Average or L2 pooling if there's is already padding and that padding method is Exclude.
-    if (poolDescriptor.m_PoolType != PoolingAlgorithm::Max) // PoolingAlgorithm::Average or PoolingAlgorithm::L2
+    // Cannot fold Average pooling if data type is quantized and layout is NHWC in Neon backend.
+    // Therefore, this specific case will become a backend specific optimization.
+    if  (!isBackendOptimization &&
+         tensorInfo.IsQuantized() &&
+         poolDescriptor.m_PoolType == PoolingAlgorithm::Average &&
+         poolDescriptor.m_DataLayout == DataLayout::NHWC)
     {
-        if ((poolHasPadding) && (poolDescriptor.m_PaddingMethod == PaddingMethod::Exclude))
-        {
-            return false;
-        }
+        return false;
     }
+
     poolDescriptor.m_PaddingMethod = PaddingMethod::IgnoreValue;
 
     return TryFoldPadIntoLayer2d<Pooling2dDescriptor>(padDescriptor, poolDescriptor, tensorInfo);
diff --git a/src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp b/src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp
new file mode 100644
index 0000000000..32627c62f7
--- /dev/null
+++ b/src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp
@@ -0,0 +1,114 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <GraphUtils.hpp>
+#include <TestUtils.hpp>
+
+#include <armnn/INetwork.hpp>
+
+#include <doctest/doctest.h>
+
+using namespace armnn;
+
+namespace
+{
+#if defined(ARMNNREF_ENABLED)||defined(ARMCOMPUTECL_ENABLED)
+void FoldPadIntoQuantizedAvgPoolTest(Compute backendId)
+{
+    // Create a network
+    INetworkPtr network = INetwork::Create();
+
+    const unsigned int inputShape[]  = {1, 2, 2, 3};
+    const unsigned int paddedShape[] = {1, 4, 4, 3};
+    const unsigned int outputShape[] = {1, 2, 2, 3};
+
+    TensorInfo inputInfo(4, inputShape, DataType::QAsymmU8, 1.0f, 0.0f);
+    TensorInfo paddedInfo(4, paddedShape, DataType::QAsymmU8, 1.0f, 0.0f);
+    TensorInfo outputInfo(4, outputShape, DataType::QAsymmU8, 1.0f, 0.0f);
+
+    IConnectableLayer* input = network->AddInputLayer(0, "input");
+    input->GetOutputSlot(0).SetTensorInfo(inputInfo);
+
+    PadDescriptor padDescriptor({{0, 0},
+                                 {1, 1},
+                                 {1, 1},
+                                 {0, 0}});
+
+    IConnectableLayer* padLayer = network->AddPadLayer(padDescriptor, "pad");
+    padLayer->GetOutputSlot(0).SetTensorInfo(paddedInfo);
+
+    Pooling2dDescriptor pooling2dDescriptor;
+    pooling2dDescriptor.m_PoolType   = PoolingAlgorithm::Average;
+    pooling2dDescriptor.m_PoolWidth  = 3;
+    pooling2dDescriptor.m_PoolHeight = 3;
+    pooling2dDescriptor.m_StrideX    = 1;
+    pooling2dDescriptor.m_StrideY    = 1;
+    pooling2dDescriptor.m_DataLayout = DataLayout::NHWC;
+
+    IConnectableLayer* pool2dLayer = network->AddPooling2dLayer(pooling2dDescriptor, "pool2d");
+    pool2dLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
+
+    IConnectableLayer* output = network->AddOutputLayer(0, "output");
+
+    // Connect up layers - input -> pad -> pool2d -> output
+    input->GetOutputSlot(0).Connect(padLayer->GetInputSlot(0));
+    padLayer->GetOutputSlot(0).Connect(pool2dLayer->GetInputSlot(0));
+    pool2dLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+
+    // Create ArmNN runtime
+    IRuntimePtr run = IRuntime::Create(IRuntime::CreationOptions());
+
+    // Optimise ArmNN network
+    IOptimizedNetworkPtr optNet = Optimize(*network, {backendId}, run->GetDeviceSpec());
+
+    auto checkPadFoldedIntoPool2d = [&](const Layer* const layer) {
+        if (!IsLayerOfType<Pooling2dLayer>(layer) || (layer->GetNameStr() != "folded-pad-into-pool2d"))
+        {
+            return false;
+        }
+
+        const auto                pool2dLayer       = static_cast<const Pooling2dLayer*>(layer);
+        const Pooling2dDescriptor pool2dLayerParams = pool2dLayer->GetParameters();
+
+        Pooling2dDescriptor pool2dLayerParamsNoPad = pool2dLayerParams;
+        pool2dLayerParamsNoPad.m_PadLeft       = 0;
+        pool2dLayerParamsNoPad.m_PadRight      = 0;
+        pool2dLayerParamsNoPad.m_PadTop        = 0;
+        pool2dLayerParamsNoPad.m_PadBottom     = 0;
+        // If we fold then PaddingMethod will be set to Ignore. The original will be Exclude.
+        pool2dLayerParamsNoPad.m_PaddingMethod = PaddingMethod::Exclude;
+
+        return (pool2dLayerParamsNoPad == pooling2dDescriptor) && (pool2dLayerParams.m_PadLeft == 1) &&
+            (pool2dLayerParams.m_PadRight == 1) && (pool2dLayerParams.m_PadTop == 1) &&
+            (pool2dLayerParams.m_PadBottom == 1) && (pool2dLayerParams.m_PaddingMethod == PaddingMethod::IgnoreValue);
+    };
+
+    Graph& graph = GetGraphForTesting(optNet.get());
+    CHECK(CheckSequence(graph.cbegin(), graph.cend(),
+                        &IsLayerOfType<InputLayer>,
+                        checkPadFoldedIntoPool2d,
+                        &IsLayerOfType<OutputLayer>));
+}
+#endif
+}
+
+
+TEST_SUITE("Optimizer_FoldPadIntoQuantizedAvgPoolCpuRef")
+{
+TEST_CASE("FoldPadIntoQuantizedAvgPoolCpuRefTest")
+{
+    FoldPadIntoQuantizedAvgPoolTest(Compute::CpuRef);
+}
+}
+
+#if defined(ARMCOMPUTECL_ENABLED)
+TEST_SUITE("Optimizer_FoldPadIntoQuantizedAvgPoolGpuAcc")
+{
+TEST_CASE("FoldPadIntoQuantizedAvgPoolGpuAccTest")
+{
+    FoldPadIntoQuantizedAvgPoolTest(Compute::GpuAcc);
+}
+}
+#endif
diff --git a/src/armnn/test/optimizations/FoldPadTests.cpp b/src/armnn/test/optimizations/FoldPadTests.cpp
index 4d7defcabe..b2672ea584 100644
--- a/src/armnn/test/optimizations/FoldPadTests.cpp
+++ b/src/armnn/test/optimizations/FoldPadTests.cpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -474,6 +474,68 @@ TEST_CASE("FoldPadLayerIntoPooling2dLayer_MaxPoolingLayerWithLargePadValueShould
                              &IsLayerOfType<OutputLayer>));
 }
 
+TEST_CASE("FoldPadLayerIntoPooling2dLayer_QuantizedAveragePoolingShouldNotBeFolded")
+{
+    Graph              graph;
+    const unsigned int inputShape[]  = {1, 2, 2, 3};
+    const unsigned int paddedShape[] = {1, 4, 4, 3};
+    const unsigned int outputShape[] = {1, 2, 2, 3};
+
+    TensorInfo inputInfo(4, inputShape, DataType::QAsymmU8);
+    TensorInfo paddedInfo(4, paddedShape, DataType::QAsymmU8);
+    TensorInfo outputInfo(4, outputShape, DataType::QAsymmU8);
+
+    Layer* input = graph.AddLayer<InputLayer>(0, "input");
+    input->GetOutputSlot().SetTensorInfo(inputInfo);
+
+    PadDescriptor padDescriptor({{0, 0},
+                                 {1, 1},
+                                 {1, 1},
+                                 {0, 0}});
+
+    PadLayer* padLayer = graph.AddLayer<PadLayer>(padDescriptor, "pad");
+    padLayer->GetOutputSlot().SetTensorInfo(paddedInfo);
+
+    Pooling2dDescriptor pooling2dDescriptor;
+    pooling2dDescriptor.m_PoolType   = PoolingAlgorithm::Average;
+    pooling2dDescriptor.m_PoolWidth  = 3;
+    pooling2dDescriptor.m_PoolHeight = 3;
+    pooling2dDescriptor.m_StrideX    = 1;
+    pooling2dDescriptor.m_StrideY    = 1;
+    pooling2dDescriptor.m_DataLayout = DataLayout::NHWC;
+
+    Pooling2dLayer* pool2dLayer = graph.AddLayer<Pooling2dLayer>(pooling2dDescriptor, "pool2d");
+    pool2dLayer->GetOutputSlot().SetTensorInfo(outputInfo);
+
+    Layer* output = graph.AddLayer<OutputLayer>(0, "output");
+
+    // Connect up layers - input -> pad -> pool2d -> output
+    input->GetOutputSlot().Connect(padLayer->GetInputSlot(0));
+    padLayer->GetOutputSlot().Connect(pool2dLayer->GetInputSlot(0));
+    pool2dLayer->GetOutputSlot().Connect(output->GetInputSlot(0));
+
+    auto checkSimplePool2d = [&](const Layer* const layer) {
+        const auto pool2dLayer = static_cast<const Pooling2dLayer*>(layer);
+        return IsLayerOfType<Pooling2dLayer>(layer) && (layer->GetNameStr() == "pool2d") &&
+            (pool2dLayer->GetParameters() == pooling2dDescriptor);
+    };
+
+    CHECK(CheckSequence(graph.cbegin(), graph.cend(),
+                        &IsLayerOfType<InputLayer>,
+                        &IsLayerOfType<PadLayer>,
+                        checkSimplePool2d,
+                        &IsLayerOfType<OutputLayer>));
+
+    armnn::Optimizer::Pass(graph, MakeOptimizations(FoldPadIntoPooling2d()));
+
+    // The optimization should not have modified the graph.
+    CHECK(CheckSequence(graph.cbegin(), graph.cend(),
+                        &IsLayerOfType<InputLayer>,
+                        &IsLayerOfType<PadLayer>,
+                        checkSimplePool2d,
+                        &IsLayerOfType<OutputLayer>));
+}
+
 #if defined(ARMNNREF_ENABLED)
 TEST_CASE("FoldPadLayerIntoPooling2dLayer_ExecuteInferenceWithAndWithoutOptimization")
 {
diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
index a26442cb86..766bf2d2cc 100644
--- a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -9,6 +9,7 @@
 #include <armnn/utility/Assert.hpp>
 
 #include <aclCommon/ArmComputeUtils.hpp>
+#include <backendsCommon/SubgraphUtils.hpp>
 
 namespace armnn
 {
@@ -20,36 +21,6 @@ namespace
 // this helper only works if all layers where the inputs connect to are not selected
 //
 
-SubgraphView::IInputSlots CreateIInputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
-{
-    SubgraphView::IInputSlots result;
-    for (auto&& layer : layers)
-    {
-        for (unsigned int i = 0 ; i < layer->GetNumInputSlots(); ++i)
-        {
-            result.push_back(&(layer->GetInputSlot(i)));
-        }
-    }
-    return result;
-}
-
-//
-// this helper only works if all layers where the outputs connect to are not selected
-//
-
-SubgraphView::IOutputSlots CreateIOutputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
-{
-    SubgraphView::IOutputSlots result;
-    for (auto &&layer: layers)
-    {
-        for (unsigned int i = 0; i < layer->GetNumOutputSlots(); ++i)
-        {
-            result.push_back(&(layer->GetOutputSlot(i)));
-        }
-    }
-    return result;
-}
-
 bool checkDataTypeInputandOutput(const Layer& layer)
 {
     auto inputInfo = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
@@ -79,19 +50,6 @@ bool checkDataTypeInputandOutput(const Layer& layer)
 
 } // namespace
 
-inline void ReportUntouchedLayers(OptimizationViews& optimizationViews, std::map<LayerGuid, Layer*> untouched)
-{
-    std::vector<Layer*> untouchedVector;
-    for (const auto& pair : untouched)
-    {
-        Layer* layer = pair.second;
-        SubgraphView subgraphView({layer},
-                                  CreateIInputsFrom({layer}),
-                                  CreateIOutputsFrom({layer}));
-        optimizationViews.AddUntouchedSubgraph(std::move(subgraphView));
-    }
-}
-
 template<typename LayerType>
 LayerType* FuseLayer(OptimizationViews& optimizationViews,
                      LayerType* baseLayer,
diff --git a/src/backends/backendsCommon/SubgraphUtils.hpp b/src/backends/backendsCommon/SubgraphUtils.hpp
new file mode 100644
index 0000000000..bd3d698a98
--- /dev/null
+++ b/src/backends/backendsCommon/SubgraphUtils.hpp
@@ -0,0 +1,99 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <optimizations/FoldPadIntoLayer2d.hpp>
+
+namespace armnn
+{
+
+namespace
+{
+
+//
+// this helper only works if all layers where the inputs connect to are not selected
+//
+
+SubgraphView::IInputSlots CreateIInputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
+{
+    SubgraphView::IInputSlots result;
+    for (auto&& layer : layers)
+    {
+        for (unsigned int i = 0 ; i < layer->GetNumInputSlots(); ++i)
+        {
+            result.push_back(&(layer->GetInputSlot(i)));
+        }
+    }
+    return result;
+}
+
+//
+// this helper only works if all layers where the outputs connect to are not selected
+//
+
+SubgraphView::IOutputSlots CreateIOutputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
+{
+    SubgraphView::IOutputSlots result;
+    for (auto &&layer: layers)
+    {
+        for (unsigned int i = 0; i < layer->GetNumOutputSlots(); ++i)
+        {
+            result.push_back(&(layer->GetOutputSlot(i)));
+        }
+    }
+    return result;
+}
+
+}
+
+inline void ReportUntouchedLayers(OptimizationViews& optimizationViews, std::map<LayerGuid, Layer*> untouched)
+{
+    std::vector<Layer*> untouchedVector;
+    for (const auto& pair : untouched)
+    {
+        Layer* layer = pair.second;
+        SubgraphView subgraphView({layer},
+                                  CreateIInputsFrom({layer}),
+                                  CreateIOutputsFrom({layer}));
+        optimizationViews.AddUntouchedSubgraph(std::move(subgraphView));
+    }
+}
+
+template<typename LayerType>
+LayerType* FoldPadLayer(OptimizationViews& optimizationViews,
+                        LayerType* baseLayer,
+                        LayerType* replacementLayer,
+                        PadLayer* padLayer)
+{
+    SubgraphView substitutionSubgraph({padLayer, baseLayer},
+                                      CreateIInputsFrom({padLayer}),
+                                      CreateIOutputsFrom({baseLayer}));
+    SubgraphView replacementSubgraph(replacementLayer);
+
+    optimizationViews.AddSubstitution({substitutionSubgraph, replacementSubgraph});
+
+    return replacementLayer;
+}
+
+template<typename LayerType>
+LayerType* FoldPadIntoAveragePool2d(OptimizationViews& optimizationViews,
+                                    Pooling2dLayer* baseLayer,
+                                    Pooling2dDescriptor& poolDescriptor,
+                                    PadLayer* padLayer)
+{
+        IConnectableLayer* replacement =
+            optimizationViews.GetINetwork()->AddPooling2dLayer(poolDescriptor, "folded-pad-into-pool2d");
+        LayerType* replacementLayer = PolymorphicDowncast<LayerType*>(replacement);
+
+        FoldPadLayer(optimizationViews,
+                     baseLayer,
+                     replacementLayer,
+                     padLayer);
+
+        return replacementLayer;
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp
index 1fe53de62a..d2e8fbfe32 100644
--- a/src/backends/cl/ClBackend.cpp
+++ b/src/backends/cl/ClBackend.cpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2017 Arm Ltd. All rights reserved.
+// Copyright © 2022 Arm Ltd. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -566,6 +566,31 @@ OptimizationViews ClBackend::OptimizeSubgraphView(const SubgraphView& subgraph,
                 untouched.erase(baseLayer->GetGuid());
             }
         }
+
+        // Special case to fuse padding into average pooling 2d for quantized datatype.
+        // Required to be done as a backend specific optimization as Neon does not support this special case.
+        if (base.GetType() == LayerType::Pooling2d)
+        {
+            Pooling2dLayer* baseLayer = PolymorphicDowncast<Pooling2dLayer*>(&base);
+            Pooling2dDescriptor poolingDescriptor = baseLayer->GetParameters();
+
+            if (baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer().GetType() == LayerType::Pad)
+            {
+                PadLayer* padLayer = PolymorphicDowncast<PadLayer*>(
+                    &baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer());
+                if (padLayer->GetOutputSlot(0).GetNumConnections() == 1 &&
+                    optimizations::pad_fold::TryFoldPadIntoLayer2d(padLayer->GetParameters(),
+                                                                   poolingDescriptor,
+                                                                   padLayer->GetOutputSlot().GetTensorInfo(),
+                                                                   true))
+                {
+                    FoldPadIntoAveragePool2d<Pooling2dLayer>(optimizationViews, baseLayer,
+                                                             poolingDescriptor, padLayer);
+                    untouched.erase(baseLayer->GetGuid());
+                    untouched.erase(padLayer->GetGuid());
+                }
+            }
+        }
     }
 
     if (optimizationViews.GetSubstitutions().empty())
diff --git a/src/backends/reference/RefBackend.cpp b/src/backends/reference/RefBackend.cpp
index a33a7756a0..8c8879c8be 100644
--- a/src/backends/reference/RefBackend.cpp
+++ b/src/backends/reference/RefBackend.cpp
@@ -14,6 +14,7 @@
 #include <armnn/backends/IMemoryManager.hpp>
 #include <armnn/utility/PolymorphicDowncast.hpp>
 #include <backendsCommon/DefaultAllocator.hpp>
+#include <backendsCommon/SubgraphUtils.hpp>
 
 #include <Optimizer.hpp>
 
@@ -70,11 +71,61 @@ IBackendInternal::ILayerSupportSharedPtr RefBackend::GetLayerSupport() const
     return layerSupport;
 }
 
-OptimizationViews RefBackend::OptimizeSubgraphView(const SubgraphView& subgraph) const
+OptimizationViews RefBackend::OptimizeSubgraphView(const SubgraphView& subgraph,
+                                                   const ModelOptions& modelOptions) const
 {
-    OptimizationViews optimizationViews;
-
-    optimizationViews.AddUntouchedSubgraph(SubgraphView(subgraph));
+    OptimizationViews optimizationViews(modelOptions);
+
+    auto it = subgraph.endIConnectable();
+    std::map<LayerGuid, Layer*> untouched;
+
+    while (it != subgraph.beginIConnectable())
+    {
+        --it;
+        Layer& base = *(PolymorphicDowncast<Layer*>(*it));
+        untouched.insert({base.GetGuid(), &base});
+    }
+
+    it = subgraph.endIConnectable();
+    while (it != subgraph.beginIConnectable())
+    {
+        --it;
+        Layer& base = *(PolymorphicDowncast<Layer*>(*it));
+
+        // Special case to fuse padding into average pooling 2d for quantized datatype.
+        // Required to be done as a backend specific optimization as Neon does not support this special case.
+        if (base.GetType() == LayerType::Pooling2d)
+        {
+            Pooling2dLayer* baseLayer = PolymorphicDowncast<Pooling2dLayer*>(&base);
+            Pooling2dDescriptor poolingDescriptor = baseLayer->GetParameters();
+
+            if (baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer().GetType() == LayerType::Pad)
+            {
+                PadLayer* padLayer = PolymorphicDowncast<PadLayer*>(
+                    &baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer());
+                if (padLayer->GetOutputSlot(0).GetNumConnections() == 1 &&
+                    optimizations::pad_fold::TryFoldPadIntoLayer2d(padLayer->GetParameters(),
+                                                                   poolingDescriptor,
+                                                                   padLayer->GetOutputSlot().GetTensorInfo(),
+                                                                   true))
+                {
+                    FoldPadIntoAveragePool2d<Pooling2dLayer>(optimizationViews, baseLayer,
+                                                             poolingDescriptor, padLayer);
+                    untouched.erase(baseLayer->GetGuid());
+                    untouched.erase(padLayer->GetGuid());
+                }
+            }
+        }
+    }
+
+    if (optimizationViews.GetSubstitutions().empty())
+    {
+        optimizationViews.AddUntouchedSubgraph(SubgraphView(subgraph));
+    }
+    else
+    {
+        ReportUntouchedLayers(optimizationViews, untouched);
+    }
 
     return optimizationViews;
 }
diff --git a/src/backends/reference/RefBackend.hpp b/src/backends/reference/RefBackend.hpp
index 9828d09f51..ecbe4d5ba9 100644
--- a/src/backends/reference/RefBackend.hpp
+++ b/src/backends/reference/RefBackend.hpp
@@ -50,7 +50,8 @@ public:
 
     IBackendInternal::ILayerSupportSharedPtr GetLayerSupport() const override;
 
-    OptimizationViews OptimizeSubgraphView(const SubgraphView& subgraph) const override;
+    OptimizationViews OptimizeSubgraphView(const SubgraphView& subgraph,
+                                           const ModelOptions& modelOptions) const override;
 
     std::vector<ITensorHandleFactory::FactoryId> GetHandleFactoryPreferences() const override;