Add ConstTensorsAsInput support for Conv3d

 * Constant weights and biases are now stored as Constant layers.
 * Updated Serializer, Deserializer and unit tests to reflect this.
 * Updated TfLiteParser.
 * Updated Ref backend to handle constant weights and
   bias as inputs rather than reading from member variables.
 * Added Conv3d EndToEnd test.
 * Added NCDHW DataLayout and unit tests.

Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: I10cdd354ca5f1c748730f92ffdb36bf810f83c8e

35 files changed, 687 insertions, 350 deletions

diff --git a/include/armnn/Descriptors.hpp b/include/armnn/Descriptors.hpp
index b412bbdcc9..39ea824045 100644
--- a/include/armnn/Descriptors.hpp
+++ b/include/armnn/Descriptors.hpp
@@ -498,6 +498,9 @@ struct Convolution3dDescriptor : BaseDescriptor
                m_DataLayout  == rhs.m_DataLayout;
     }
 
+    /// Get the number of views/inputs.
+    uint32_t GetNumInputs() const;
+
     /// Padding left value in the width dimension.
     uint32_t             m_PadLeft;
     /// Padding right value in the width dimension.
@@ -524,7 +527,7 @@ struct Convolution3dDescriptor : BaseDescriptor
     uint32_t             m_DilationZ;
     /// Enable/disable bias.
     bool                 m_BiasEnabled;
-    /// The data layout to be used (NDHWC).
+    /// The data layout to be used (NDHWC, NCDHW).
     DataLayout           m_DataLayout;
 };
 
diff --git a/include/armnn/INetwork.hpp b/include/armnn/INetwork.hpp
index ab92f05112..707ae00bb3 100644
--- a/include/armnn/INetwork.hpp
+++ b/include/armnn/INetwork.hpp
@@ -258,13 +258,9 @@ public:
 
     /// Adds a 3D convolution layer to the network.
     /// @param convolution3dDescriptor - Description of the 3D convolution layer.
-    /// @param weights - Tensor for the weights data.
-    /// @param biases - Optional tensor for the bias data. If specified, must match the output tensor shape.
     /// @param name - Optional name for the layer.
     /// @return - Interface for configuring the layer.
     IConnectableLayer* AddConvolution3dLayer(const Convolution3dDescriptor& convolution3dDescriptor,
-                                             const ConstTensor& weights,
-                                             const Optional<ConstTensor>& biases,
                                              const char* name = nullptr);
 
     /// Adds a depth to space layer to the network.
diff --git a/include/armnn/Types.hpp b/include/armnn/Types.hpp
index 7f2e192102..4f39ebe16a 100644
--- a/include/armnn/Types.hpp
+++ b/include/armnn/Types.hpp
@@ -50,7 +50,8 @@ enum class DataLayout
 {
     NCHW = 1,
     NHWC = 2,
-    NDHWC = 3
+    NDHWC = 3,
+    NCDHW = 4
 };
 
 /// Define the behaviour of the internal profiler when outputting network details
diff --git a/include/armnn/TypesUtils.hpp b/include/armnn/TypesUtils.hpp
index d08f592d86..a1c11b74df 100644
--- a/include/armnn/TypesUtils.hpp
+++ b/include/armnn/TypesUtils.hpp
@@ -215,6 +215,7 @@ constexpr const char* GetDataLayoutName(DataLayout dataLayout)
         case DataLayout::NCHW:  return "NCHW";
         case DataLayout::NHWC:  return "NHWC";
         case DataLayout::NDHWC: return "NDHWC";
+        case DataLayout::NCDHW: return "NCDHW";
         default:                return "Unknown";
     }
 }
diff --git a/src/armnn/Descriptors.cpp b/src/armnn/Descriptors.cpp
index ab68097247..ef55ee7bb5 100644
--- a/src/armnn/Descriptors.cpp
+++ b/src/armnn/Descriptors.cpp
@@ -441,4 +441,15 @@ uint32_t FullyConnectedDescriptor::GetNumInputs() const
     return numInputs;
 }
 
+uint32_t Convolution3dDescriptor::GetNumInputs() const
+{
+    // Return 2 otherwise check if bias is enabled
+    unsigned int numInputs = 2;
+    if (m_BiasEnabled)
+    {
+        numInputs = 3;
+    }
+    return numInputs;
+}
+
 }
diff --git a/src/armnn/Graph.cpp b/src/armnn/Graph.cpp
index 30639b12e8..0591bea99a 100644
--- a/src/armnn/Graph.cpp
+++ b/src/armnn/Graph.cpp
@@ -588,7 +588,7 @@ void Graph::InferTensorInfos()
 }
 
 /// Throws exception due to a layer input not being connected to an output slot.
-/// Verifies weights and bias are set for FullyConnected layers on input slots 1
+/// Verifies weights and bias are set for layers on input slots 1
 /// and 2 respectively. Method checks if bias is enabled before ensuring it is set.
 ///
 /// @param layer constant pointer to a Layer object
@@ -600,7 +600,8 @@ void Graph::ConstructErrorMessageForUnconnectedInputs(Layer* const layer,
     std::ostringstream message;
     bool noWeightsAndBias = false;
 
-    if (layer->GetType() == armnn::LayerType::FullyConnected && slotIndex > 0)
+    if ((layer->GetType() == armnn::LayerType::FullyConnected ||
+         layer->GetType() == armnn::LayerType::Convolution3d) && slotIndex > 0)
     {
         // If weights are not set and is bias enabled, also check if bias is set
         if (slotIndex == 1 && layer->GetNumInputSlots() == 3)
@@ -608,7 +609,7 @@ void Graph::ConstructErrorMessageForUnconnectedInputs(Layer* const layer,
             const IOutputSlot* biasSource = layer->GetInputSlot(2).GetConnectedOutputSlot();
             if (biasSource == NULL)
             {
-                message << "FullyConnected layer weights and bias not set: ";
+                message << layer->GetName() << " layer weights and bias not set: ";
                 noWeightsAndBias = true;
             }
         }
@@ -618,11 +619,11 @@ void Graph::ConstructErrorMessageForUnconnectedInputs(Layer* const layer,
         {
             if (slotIndex == 1)
             {
-                message << "FullyConnected layer weights not set: ";
+                message << layer->GetName() << " layer weights not set: ";
             }
             else
             {
-                message << "FullyConnected layer bias not set: ";
+                message << layer->GetName() << " layer bias not set: ";
             }
         }
     }
diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp
index 99d7b96ec2..b516d519d5 100644
--- a/src/armnn/Network.cpp
+++ b/src/armnn/Network.cpp
@@ -114,11 +114,9 @@ IConnectableLayer* INetwork::AddConvolution2dLayer(const Convolution2dDescriptor
 
 
 IConnectableLayer* INetwork::AddConvolution3dLayer(const Convolution3dDescriptor& convolution3dDescriptor,
-                                                   const ConstTensor& weights,
-                                                   const Optional<ConstTensor>& biases,
                                                    const char* name)
 {
-    return pNetworkImpl->AddConvolution3dLayer(convolution3dDescriptor, weights, biases, name);
+    return pNetworkImpl->AddConvolution3dLayer(convolution3dDescriptor, name);
 }
 
 
@@ -1934,25 +1932,9 @@ IConnectableLayer* NetworkImpl::AddConvolution2dLayer(const Convolution2dDescrip
 }
 
 IConnectableLayer* NetworkImpl::AddConvolution3dLayer(const Convolution3dDescriptor& convolution3dDescriptor,
-                                                      const ConstTensor& weights,
-                                                      const Optional<ConstTensor>& biases,
                                                       const char* name)
 {
-    if (convolution3dDescriptor.m_BiasEnabled && !biases.has_value())
-    {
-        throw InvalidArgumentException("AddConvolution2dLayer: biases cannot be empty");
-    }
-
-    const auto layer = m_Graph->AddLayer<Convolution3dLayer>(convolution3dDescriptor, name);
-
-    layer->m_Weight = std::make_shared<ScopedTensorHandle>(weights);
-
-    if (convolution3dDescriptor.m_BiasEnabled)
-    {
-        layer->m_Bias = std::make_shared<ScopedTensorHandle>(biases.value());
-    }
-
-    return layer;
+    return m_Graph->AddLayer<Convolution3dLayer>(convolution3dDescriptor, name);
 }
 
 IConnectableLayer* NetworkImpl::AddDepthToSpaceLayer(const DepthToSpaceDescriptor& depthToSpaceDescriptor,
diff --git a/src/armnn/Network.hpp b/src/armnn/Network.hpp
index eb1d39d2f6..818a765296 100644
--- a/src/armnn/Network.hpp
+++ b/src/armnn/Network.hpp
@@ -87,8 +87,6 @@ public:
                                              const char* name = nullptr);
 
     IConnectableLayer* AddConvolution3dLayer(const Convolution3dDescriptor& convolution3dDescriptor,
-                                             const ConstTensor& weights,
-                                             const Optional<ConstTensor>& biases,
                                              const char* name = nullptr);
 
     IConnectableLayer* AddConstantLayer(const ConstTensor& input, const char* name = nullptr);
diff --git a/src/armnn/layers/Convolution3dLayer.cpp b/src/armnn/layers/Convolution3dLayer.cpp
index 0e38c0b129..1c2d1b9872 100644
--- a/src/armnn/layers/Convolution3dLayer.cpp
+++ b/src/armnn/layers/Convolution3dLayer.cpp
@@ -16,7 +16,7 @@ namespace armnn
 {
 
 Convolution3dLayer::Convolution3dLayer(const Convolution3dDescriptor& param, const char* name)
-    : LayerWithParameters(1, 1, LayerType::Convolution3d, param, name)
+    : LayerWithParameters(param.GetNumInputs(), 1, LayerType::Convolution3d, param, name)
 {
 }
 
@@ -25,12 +25,11 @@ void Convolution3dLayer::SerializeLayerParameters(ParameterStringifyFunction& fn
     const std::vector<TensorShape>& inputShapes =
     {
         GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape(),
-        m_Weight->GetTensorInfo().GetShape()
+        GetInputSlot(1).GetConnection()->GetTensorInfo().GetShape(),
     };
 
     // Conv3d Filter Layout: [D,H,W,I,O]
     const TensorShape filterShape = inputShapes[1];
-    DataLayoutIndexed dataLayoutIndex(m_Param.m_DataLayout);
     unsigned int filterDepth = filterShape[0];
     unsigned int filterHeight = filterShape[1];
     unsigned int filterWidth = filterShape[2];
@@ -48,18 +47,7 @@ void Convolution3dLayer::SerializeLayerParameters(ParameterStringifyFunction& fn
 
 std::unique_ptr<IWorkload> Convolution3dLayer::CreateWorkload(const IWorkloadFactory& factory) const
 {
-    // At this level constant data should not be released.
-    ARMNN_ASSERT_MSG(m_Weight != nullptr, "Convolution3dLayer: Weights data should not be null.");
-
     Convolution3dQueueDescriptor descriptor;
-    descriptor.m_Weight = m_Weight.get();
-
-    if (m_Param.m_BiasEnabled)
-    {
-        ARMNN_ASSERT_MSG(m_Bias != nullptr, "Convolution3dLayer: Bias data should not be null.");
-        descriptor.m_Bias = m_Bias.get();
-    }
-
     SetAdditionalInfo(descriptor);
 
     return factory.CreateConvolution3d(descriptor, PrepInfoAndDesc(descriptor));
@@ -68,14 +56,6 @@ std::unique_ptr<IWorkload> Convolution3dLayer::CreateWorkload(const IWorkloadFac
 Convolution3dLayer* Convolution3dLayer::Clone(Graph& graph) const
 {
     auto layer = CloneBase<Convolution3dLayer>(graph, m_Param, GetName());
-
-    layer->m_Weight = m_Weight ? m_Weight : nullptr;
-
-    if (layer->m_Param.m_BiasEnabled)
-    {
-        layer->m_Bias = m_Bias ? m_Bias : nullptr;
-    }
-
     return std::move(layer);
 }
 
@@ -117,36 +97,33 @@ std::vector<TensorShape> Convolution3dLayer::InferOutputShapes(const std::vector
     unsigned int outChannels = filterShape[4];
     unsigned int outBatchSize = inBatchSize;
 
-    TensorShape tensorShape = TensorShape( { outBatchSize, outDepth, outHeight, outWidth, outChannels } );
+    TensorShape tensorShape = m_Param.m_DataLayout == armnn::DataLayout::NDHWC ?
+            TensorShape( { outBatchSize, outDepth, outHeight, outWidth, outChannels } ) :
+            TensorShape( { outBatchSize, outChannels, outDepth, outHeight, outWidth });
 
     return std::vector<TensorShape>({ tensorShape });
 }
 
 void Convolution3dLayer::ValidateTensorShapesFromInputs()
 {
-    VerifyLayerConnections(1, CHECK_LOCATION());
+    VerifyLayerConnections(m_Param.GetNumInputs(), CHECK_LOCATION());
 
     const TensorShape& outputShape = GetOutputSlot(0).GetTensorInfo().GetShape();
 
     VerifyShapeInferenceType(outputShape, m_ShapeInferenceMethod);
 
-    // check if we m_Weight data is not nullptr
-    ARMNN_ASSERT_MSG(m_Weight != nullptr, "Convolution3dLayer: Weights data should not be null.");
+    ARMNN_ASSERT_MSG(GetInputSlot(1).GetConnection(),
+                     "Convolution3dLayer: Weights should be connected to input slot 1.");
 
     auto inferredShapes = InferOutputShapes({
         GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape(),
-        m_Weight->GetTensorInfo().GetShape() });
+        GetInputSlot(1).GetConnection()->GetTensorInfo().GetShape() });
 
     ARMNN_ASSERT(inferredShapes.size() == 1);
 
     ValidateAndCopyShape(outputShape, inferredShapes[0], m_ShapeInferenceMethod, "Convolution3dLayer");
 }
 
-Layer::ConstantTensors Convolution3dLayer::GetConstantTensorsByRef()
-{
-    return {m_Weight, m_Bias};
-}
-
 ARMNN_NO_DEPRECATE_WARN_BEGIN
 void Convolution3dLayer::Accept(ILayerVisitor& visitor) const
 {
@@ -157,16 +134,7 @@ ARMNN_NO_DEPRECATE_WARN_END
 
 void Convolution3dLayer::ExecuteStrategy(IStrategy& strategy) const
 {
-    ManagedConstTensorHandle managedWeight(m_Weight);
-    std::vector<armnn::ConstTensor> constTensors { { managedWeight.GetTensorInfo(), managedWeight.Map() } };
-
-    ManagedConstTensorHandle managedBias(m_Bias);
-    if (GetParameters().m_BiasEnabled)
-    {
-        constTensors.emplace_back(ConstTensor(managedBias.GetTensorInfo(), managedBias.Map()));
-    }
-
-    strategy.ExecuteStrategy(this, GetParameters(), constTensors, GetName());
+    strategy.ExecuteStrategy(this, GetParameters(), {}, GetName());
 }
 
 } // namespace armnn
diff --git a/src/armnn/layers/Convolution3dLayer.hpp b/src/armnn/layers/Convolution3dLayer.hpp
index bef5715098..7cbd6428dc 100644
--- a/src/armnn/layers/Convolution3dLayer.hpp
+++ b/src/armnn/layers/Convolution3dLayer.hpp
@@ -16,12 +16,6 @@ class ScopedTensorHandle;
 class Convolution3dLayer : public LayerWithParameters<Convolution3dDescriptor>
 {
 public:
-
-    /// A unique pointer to store Weight values.
-    std::shared_ptr<ConstTensorHandle> m_Weight;
-    /// A unique pointer to store Bias values.
-    std::shared_ptr<ConstTensorHandle> m_Bias;
-
     /// Makes a workload for the Convolution3d type.
     /// @param [in] graph The graph where this layer can be found.
     /// @param [in] factory The workload factory which will create the workload.
@@ -59,10 +53,6 @@ protected:
 
     /// Default destructor
     ~Convolution3dLayer() = default;
-
-    /// Retrieve the handles to the constant values stored by the layer.
-    /// @return A vector of the constant tensors stored by this layer.
-    ConstantTensors GetConstantTensorsByRef() override;
 };
 
 } // namespace
diff --git a/src/armnnDeserializer/Deserializer.cpp b/src/armnnDeserializer/Deserializer.cpp
index 6b73946af2..c088ef7b54 100644
--- a/src/armnnDeserializer/Deserializer.cpp
+++ b/src/armnnDeserializer/Deserializer.cpp
@@ -449,6 +449,8 @@ armnn::DataLayout ToDataLayout(armnnSerializer::DataLayout dataLayout)
             return armnn::DataLayout::NHWC;
         case armnnSerializer::DataLayout::DataLayout_NDHWC:
             return armnn::DataLayout::NDHWC;
+        case armnnSerializer::DataLayout::DataLayout_NCDHW:
+            return armnn::DataLayout::NCDHW;
         case armnnSerializer::DataLayout::DataLayout_NCHW:
         default:
             return armnn::DataLayout::NCHW;
@@ -1402,7 +1404,6 @@ void IDeserializer::DeserializerImpl::ParseConvolution3d(GraphPtr graph, unsigne
     CHECK_LAYERS(graph, 0, layerIndex);
     auto inputs = GetInputs(graph, layerIndex);
     CHECK_LOCATION();
-    CHECK_VALID_SIZE(inputs.size(), 1);
 
     auto outputs = GetOutputs(graph, layerIndex);
     CHECK_VALID_SIZE(outputs.size(), 1);
@@ -1424,22 +1425,14 @@ void IDeserializer::DeserializerImpl::ParseConvolution3d(GraphPtr graph, unsigne
     descriptor.m_DilationX = serializerDescriptor->dilationX();
     descriptor.m_DilationY = serializerDescriptor->dilationY();
     descriptor.m_DilationZ = serializerDescriptor->dilationZ();
-    descriptor.m_BiasEnabled = serializerDescriptor->biasEnabled();;
+    descriptor.m_BiasEnabled = serializerDescriptor->biasEnabled();
     descriptor.m_DataLayout = ToDataLayout(serializerDescriptor->dataLayout());
 
-    armnn::ConstTensor weights = ToConstTensor(serializerLayer->weights());
-    armnn::ConstTensor biases;
+    uint32_t numInputs = descriptor.GetNumInputs();
+    CHECK_VALID_SIZE(inputs.size(), numInputs);
+
+    IConnectableLayer* layer = m_Network->AddConvolution3dLayer(descriptor, layerName.c_str());
 
-    armnn::Optional<armnn::ConstTensor> optionalBiases = armnn::EmptyOptional();
-    if (descriptor.m_BiasEnabled)
-    {
-        biases = ToConstTensor(serializerLayer->biases());
-        optionalBiases = armnn::Optional<armnn::ConstTensor>(biases);
-    }
-    IConnectableLayer* layer = m_Network->AddConvolution3dLayer(descriptor,
-                                                                weights,
-                                                                optionalBiases,
-                                                                layerName.c_str());
     armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]);
     layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
 
diff --git a/src/armnnDeserializer/test/DeserializeConvolution3d.cpp b/src/armnnDeserializer/test/DeserializeConvolution3d.cpp
index 057ab6fbda..23fd811cdb 100644
--- a/src/armnnDeserializer/test/DeserializeConvolution3d.cpp
+++ b/src/armnnDeserializer/test/DeserializeConvolution3d.cpp
@@ -30,13 +30,11 @@ struct Convolution3dFixture : public ParserFlatbuffersSerializeFixture
                   base: {
                     layerName: "InputLayer",
                     layerType: "Input",
-                    inputSlots: [{
-                        index: 0,
-                        connection: {sourceLayerIndex:0, outputSlotIndex:0 },
-                    }],
+                    inputSlots: [
+
+                    ],
                     outputSlots: [
                       {
-                        index: 0,
                         tensorInfo: {
                           dimensions: )" + inputShape + R"(,
                           dataType: )" + dataType + R"(,
@@ -56,26 +54,19 @@ struct Convolution3dFixture : public ParserFlatbuffersSerializeFixture
               }
             },
             {
-              layer_type: "Convolution3dLayer",
+              layer_type: "ConstantLayer",
               layer: {
                 base: {
                   index: 1,
-                  layerName: "convolution3d",
-                  layerType: "Convolution2d",
+                  layerName: "Weights",
+                  layerType: "Constant",
                   inputSlots: [
-                    {
-                      index: 0,
-                      connection: {
-                        sourceLayerIndex: 0,
-                        outputSlotIndex: 0
-                      }
-                    }
+
                   ],
                   outputSlots: [
                     {
-                      index: 0,
                       tensorInfo: {
-                        dimensions: )" + outputShape + R"(,
+                        dimensions: )" + weightsShape + R"(,
                         dataType: )" + dataType + R"(,
                         quantizationScale: 0.1,
                         dimensionSpecificity: [
@@ -89,12 +80,7 @@ struct Convolution3dFixture : public ParserFlatbuffersSerializeFixture
                     }
                   ]
                 },
-                descriptor: {
-                  strideX: 2,
-                  strideY: 2,
-                  strideZ: 2
-                },
-                weights: {
+                input: {
                   info: {
                     dimensions: )" + weightsShape + R"(,
                     dataType: )" + dataType + R"(,
@@ -127,29 +113,71 @@ struct Convolution3dFixture : public ParserFlatbuffersSerializeFixture
               }
             },
             {
+              layer_type: "Convolution3dLayer",
+              layer: {
+                base: {
+                  index: 2,
+                  layerName: "convolution3d",
+                  layerType: "Convolution3d",
+                  inputSlots: [
+                    {
+                      connection: {
+                        sourceLayerIndex: 0,
+                        outputSlotIndex: 0
+                      }
+                    },
+                    {
+                      index: 1,
+                      connection: {
+                        sourceLayerIndex: 1,
+                        outputSlotIndex: 0
+                      }
+                    }
+                  ],
+                  outputSlots: [
+                    {
+                      tensorInfo: {
+                        dimensions: )" + outputShape + R"(,
+                        dataType: )" + dataType + R"(,
+                        quantizationScale: 0.1,
+                        dimensionSpecificity: [
+                          true,
+                          true,
+                          true,
+                          true,
+                          true
+                        ]
+                      }
+                    }
+                  ]
+                },
+                descriptor: {
+                  strideX: 2,
+                  strideY: 2,
+                  strideZ: 2
+                }
+              }
+            },
+            {
               layer_type: "OutputLayer",
               layer: {
                 base: {
                   layerBindingId: 2,
                   base: {
-                    index: 2,
+                    index: 3,
                     layerName: "OutputLayer",
                     layerType: "Output",
                     inputSlots: [
                       {
                         connection: {
-                          sourceLayerIndex: 1,
+                          sourceLayerIndex: 2,
                           outputSlotIndex: 0
                         }
                       }
                     ],
-                    outputSlots: [{
-                        index: 0,
-                        tensorInfo: {
-                            dimensions: )" + outputShape + R"(,
-                            dataType: )" + dataType + R"(
-                        },
-                    }]
+                    outputSlots: [
+
+                    ]
                   }
                 }
               }
diff --git a/src/armnnSerializer/ArmnnSchema.fbs b/src/armnnSerializer/ArmnnSchema.fbs
index 77982888c8..c577a11a52 100644
--- a/src/armnnSerializer/ArmnnSchema.fbs
+++ b/src/armnnSerializer/ArmnnSchema.fbs
@@ -46,7 +46,8 @@ enum DataType : byte {
 enum DataLayout : byte {
     NHWC = 0,
     NCHW = 1,
-    NDHWC = 2
+    NDHWC = 2,
+    NCDHW = 3
 }
 
 enum ReduceOperation: byte {
@@ -287,8 +288,6 @@ table Convolution2dDescriptor {
 table Convolution3dLayer {
     base:LayerBase;
     descriptor:Convolution3dDescriptor;
-    weights:ConstTensor;
-    biases:ConstTensor;
 }
 
 table Convolution3dDescriptor {
diff --git a/src/armnnSerializer/ArmnnSchema_generated.h b/src/armnnSerializer/ArmnnSchema_generated.h
index 8234aa9c47..712ad28574 100644
--- a/src/armnnSerializer/ArmnnSchema_generated.h
+++ b/src/armnnSerializer/ArmnnSchema_generated.h
@@ -540,31 +540,34 @@ enum DataLayout {
   DataLayout_NHWC = 0,
   DataLayout_NCHW = 1,
   DataLayout_NDHWC = 2,
+  DataLayout_NCDHW = 3,
   DataLayout_MIN = DataLayout_NHWC,
-  DataLayout_MAX = DataLayout_NDHWC
+  DataLayout_MAX = DataLayout_NCDHW
 };
 
-inline const DataLayout (&EnumValuesDataLayout())[3] {
+inline const DataLayout (&EnumValuesDataLayout())[4] {
   static const DataLayout values[] = {
     DataLayout_NHWC,
     DataLayout_NCHW,
-    DataLayout_NDHWC
+    DataLayout_NDHWC,
+    DataLayout_NCDHW
   };
   return values;
 }
 
 inline const char * const *EnumNamesDataLayout() {
-  static const char * const names[4] = {
+  static const char * const names[5] = {
     "NHWC",
     "NCHW",
     "NDHWC",
+    "NCDHW",
     nullptr
   };
   return names;
 }
 
 inline const char *EnumNameDataLayout(DataLayout e) {
-  if (flatbuffers::IsOutRange(e, DataLayout_NHWC, DataLayout_NDHWC)) return "";
+  if (flatbuffers::IsOutRange(e, DataLayout_NHWC, DataLayout_NCDHW)) return "";
   const size_t index = static_cast<size_t>(e);
   return EnumNamesDataLayout()[index];
 }
@@ -3250,9 +3253,7 @@ struct Convolution3dLayer FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
   typedef Convolution3dLayerBuilder Builder;
   enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE {
     VT_BASE = 4,
-    VT_DESCRIPTOR = 6,
-    VT_WEIGHTS = 8,
-    VT_BIASES = 10
+    VT_DESCRIPTOR = 6
   };
   const armnnSerializer::LayerBase *base() const {
     return GetPointer<const armnnSerializer::LayerBase *>(VT_BASE);
@@ -3260,22 +3261,12 @@ struct Convolution3dLayer FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
   const armnnSerializer::Convolution3dDescriptor *descriptor() const {
     return GetPointer<const armnnSerializer::Convolution3dDescriptor *>(VT_DESCRIPTOR);
   }
-  const armnnSerializer::ConstTensor *weights() const {
-    return GetPointer<const armnnSerializer::ConstTensor *>(VT_WEIGHTS);
-  }
-  const armnnSerializer::ConstTensor *biases() const {
-    return GetPointer<const armnnSerializer::ConstTensor *>(VT_BIASES);
-  }
   bool Verify(flatbuffers::Verifier &verifier) const {
     return VerifyTableStart(verifier) &&
            VerifyOffset(verifier, VT_BASE) &&
            verifier.VerifyTable(base()) &&
            VerifyOffset(verifier, VT_DESCRIPTOR) &&
            verifier.VerifyTable(descriptor()) &&
-           VerifyOffset(verifier, VT_WEIGHTS) &&
-           verifier.VerifyTable(weights()) &&
-           VerifyOffset(verifier, VT_BIASES) &&
-           verifier.VerifyTable(biases()) &&
            verifier.EndTable();
   }
 };
@@ -3290,12 +3281,6 @@ struct Convolution3dLayerBuilder {
   void add_descriptor(flatbuffers::Offset<armnnSerializer::Convolution3dDescriptor> descriptor) {
     fbb_.AddOffset(Convolution3dLayer::VT_DESCRIPTOR, descriptor);
   }
-  void add_weights(flatbuffers::Offset<armnnSerializer::ConstTensor> weights) {
-    fbb_.AddOffset(Convolution3dLayer::VT_WEIGHTS, weights);
-  }
-  void add_biases(flatbuffers::Offset<armnnSerializer::ConstTensor> biases) {
-    fbb_.AddOffset(Convolution3dLayer::VT_BIASES, biases);
-  }
   explicit Convolution3dLayerBuilder(flatbuffers::FlatBufferBuilder &_fbb)
         : fbb_(_fbb) {
     start_ = fbb_.StartTable();
@@ -3311,12 +3296,8 @@ struct Convolution3dLayerBuilder {
 inline flatbuffers::Offset<Convolution3dLayer> CreateConvolution3dLayer(
     flatbuffers::FlatBufferBuilder &_fbb,
     flatbuffers::Offset<armnnSerializer::LayerBase> base = 0,
-    flatbuffers::Offset<armnnSerializer::Convolution3dDescriptor> descriptor = 0,
-    flatbuffers::Offset<armnnSerializer::ConstTensor> weights = 0,
-    flatbuffers::Offset<armnnSerializer::ConstTensor> biases = 0) {
+    flatbuffers::Offset<armnnSerializer::Convolution3dDescriptor> descriptor = 0) {
   Convolution3dLayerBuilder builder_(_fbb);
-  builder_.add_biases(biases);
-  builder_.add_weights(weights);
   builder_.add_descriptor(descriptor);
   builder_.add_base(base);
   return builder_.Finish();
diff --git a/src/armnnSerializer/Serializer.cpp b/src/armnnSerializer/Serializer.cpp
index 7e1b74e10d..84a9d53b69 100644
--- a/src/armnnSerializer/Serializer.cpp
+++ b/src/armnnSerializer/Serializer.cpp
@@ -388,18 +388,15 @@ void SerializerStrategy::SerializeConvolution2dLayer(const armnn::IConnectableLa
     CreateAnyLayer(flatBufferLayer.o, serializer::Layer::Layer_Convolution2dLayer);
 }
 
-// Build FlatBuffer for Convolution2dLayer
+// Build FlatBuffer for Convolution3dLayer
 void SerializerStrategy::SerializeConvolution3dLayer(const armnn::IConnectableLayer* layer,
                                                      const armnn::Convolution3dDescriptor& descriptor,
-                                                     const std::vector<armnn::ConstTensor>& constants,
                                                      const char* name)
 {
     IgnoreUnused(name);
 
-    const armnn::ConstTensor weights = constants[0];
-
     // Create FlatBuffer BaseLayer
-    auto flatBufferBaseLayer = CreateLayerBase(layer, serializer::LayerType::LayerType_Convolution2d);
+    auto flatBufferBaseLayer = CreateLayerBase(layer, serializer::LayerType::LayerType_Convolution3d);
 
     auto flatBufferDescriptor = CreateConvolution3dDescriptor(m_flatBufferBuilder,
                                                               descriptor.m_PadLeft,
@@ -416,21 +413,11 @@ void SerializerStrategy::SerializeConvolution3dLayer(const armnn::IConnectableLa
                                                               descriptor.m_DilationZ,
                                                               descriptor.m_BiasEnabled,
                                                               GetFlatBufferDataLayout(descriptor.m_DataLayout));
-    auto flatBufferWeightsConstTensorInfo = CreateConstTensorInfo(weights);
-    flatbuffers::Offset<serializer::ConstTensor> flatBufferBiasesConstTensorInfo;
-
-    if (constants.size() > 1)
-    {
-        const armnn::ConstTensor biases = constants[1];
-        flatBufferBiasesConstTensorInfo = CreateConstTensorInfo(biases);
-    }
 
-    // Create the FlatBuffer Convolution2dLayer
+    // Create the FlatBuffer Convolution3dLayer
     auto flatBufferLayer = CreateConvolution3dLayer(m_flatBufferBuilder,
                                                     flatBufferBaseLayer,
-                                                    flatBufferDescriptor,
-                                                    flatBufferWeightsConstTensorInfo,
-                                                    flatBufferBiasesConstTensorInfo);
+                                                    flatBufferDescriptor);
 
     // Add the AnyLayer to the FlatBufferLayers
     CreateAnyLayer(flatBufferLayer.o, serializer::Layer::Layer_Convolution3dLayer);
@@ -2038,7 +2025,6 @@ void SerializerStrategy::ExecuteStrategy(const armnn::IConnectableLayer* layer,
                     static_cast<const armnn::Convolution3dDescriptor&>(descriptor);
             SerializeConvolution3dLayer(layer,
                                         layerDescriptor,
-                                        constants,
                                         name);
             break;
         }
diff --git a/src/armnnSerializer/Serializer.hpp b/src/armnnSerializer/Serializer.hpp
index 2f827ac059..1c0a9a619f 100644
--- a/src/armnnSerializer/Serializer.hpp
+++ b/src/armnnSerializer/Serializer.hpp
@@ -150,7 +150,6 @@ private:
 
     void SerializeConvolution3dLayer(const armnn::IConnectableLayer* layer,
                                      const armnn::Convolution3dDescriptor& descriptor,
-                                     const std::vector<armnn::ConstTensor>& constants,
                                      const char* name = nullptr);
 
     void SerializeDepthToSpaceLayer(const armnn::IConnectableLayer* layer,
diff --git a/src/armnnSerializer/SerializerUtils.cpp b/src/armnnSerializer/SerializerUtils.cpp
index fca6db8449..5ad27715c4 100644
--- a/src/armnnSerializer/SerializerUtils.cpp
+++ b/src/armnnSerializer/SerializerUtils.cpp
@@ -99,6 +99,8 @@ armnnSerializer::DataLayout GetFlatBufferDataLayout(armnn::DataLayout dataLayout
             return armnnSerializer::DataLayout::DataLayout_NHWC;
         case armnn::DataLayout::NDHWC:
             return armnnSerializer::DataLayout::DataLayout_NDHWC;
+        case armnn::DataLayout::NCDHW:
+            return armnnSerializer::DataLayout::DataLayout_NCDHW;
         case armnn::DataLayout::NCHW:
         default:
             return armnnSerializer::DataLayout::DataLayout_NCHW;
diff --git a/src/armnnSerializer/test/SerializerTestUtils.hpp b/src/armnnSerializer/test/SerializerTestUtils.hpp
index c6f148b1a1..ce4d2cc330 100644
--- a/src/armnnSerializer/test/SerializerTestUtils.hpp
+++ b/src/armnnSerializer/test/SerializerTestUtils.hpp
@@ -69,6 +69,7 @@ public:
         {
             case armnn::LayerType::Input: break;
             case armnn::LayerType::Output: break;
+            case armnn::LayerType::Constant: break;
             default:
             {
                 VerifyNameAndConnections(layer, name);
diff --git a/src/armnnSerializer/test/SerializerTests.cpp b/src/armnnSerializer/test/SerializerTests.cpp
index f2c9852607..2bffe0b9fd 100644
--- a/src/armnnSerializer/test/SerializerTests.cpp
+++ b/src/armnnSerializer/test/SerializerTests.cpp
@@ -472,25 +472,26 @@ TEST_CASE("SerializeConvolution3d")
 
     armnn::INetworkPtr network = armnn::INetwork::Create();
     armnn::IConnectableLayer* const inputLayer  = network->AddInputLayer(0);
-    armnn::IConnectableLayer* const convLayer   =
-            network->AddConvolution3dLayer(descriptor,
-                                           weights,
-                                           armnn::Optional<armnn::ConstTensor>(biases),
-                                           layerName.c_str());
+    armnn::IConnectableLayer* const weightsLayer = network->AddConstantLayer(weights, "Weights");
+    armnn::IConnectableLayer* const biasesLayer = network->AddConstantLayer(biases, "Biases");
+    armnn::IConnectableLayer* const convLayer   = network->AddConvolution3dLayer(descriptor, layerName.c_str());
     armnn::IConnectableLayer* const outputLayer = network->AddOutputLayer(0);
 
     inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0));
+    weightsLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(1));
+    biasesLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(2));
     convLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
 
     inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
+    weightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo);
+    biasesLayer->GetOutputSlot(0).SetTensorInfo(biasesInfo);
     convLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
 
     armnn::INetworkPtr deserializedNetwork = DeserializeNetwork(SerializeNetwork(*network));
     CHECK(deserializedNetwork);
 
-    const std::vector<armnn::ConstTensor>& constants {weights, biases};
-    LayerVerifierBaseWithDescriptorAndConstants<armnn::Convolution3dDescriptor> verifier(
-            layerName, {inputInfo}, {outputInfo}, descriptor, constants);
+    LayerVerifierBaseWithDescriptor<armnn::Convolution3dDescriptor> verifier(
+            layerName, {inputInfo, weightsInfo, biasesInfo}, {outputInfo}, descriptor);
     deserializedNetwork->ExecuteStrategy(verifier);
 }
 
diff --git a/src/armnnTfLiteParser/TfLiteParser.cpp b/src/armnnTfLiteParser/TfLiteParser.cpp
index 81d491a1a1..7db5d85b13 100644
--- a/src/armnnTfLiteParser/TfLiteParser.cpp
+++ b/src/armnnTfLiteParser/TfLiteParser.cpp
@@ -1099,36 +1099,29 @@ void TfLiteParserImpl::ParseConv3D(size_t subgraphIndex, size_t operatorIndex)
 
     auto filterTensorAndData = CreateConstTensorNonPermuted(inputs[1], filterTensorInfo);
 
-    armnn::IConnectableLayer* layer = nullptr;
     auto layerName = fmt::format("Conv3D:{}:{}", subgraphIndex, operatorIndex);
 
+    auto inputTensorIndexes = AsUnsignedVector(GetInputTensorIds(m_Model, subgraphIndex, operatorIndex));
+    // Add the first input and weights tensor to the registration list.
+    // The constant weights will be added by SetupConstantLayers.
+    std::vector<unsigned int> tensorIndexesToRegister = {inputTensorIndexes[0], inputTensorIndexes[1]};
+
     if (inputs.size() == 3)
     {
         desc.m_BiasEnabled = true;
-        armnn::TensorInfo biasTensorInfo = ToTensorInfo(inputs[2]);
-        auto biasTensorAndData = CreateConstTensorNonPermuted(inputs[2], biasTensorInfo);
-        layer = m_Network->AddConvolution3dLayer(desc,
-                                                 filterTensorAndData,
-                                                 Optional<ConstTensor>(biasTensorAndData),
-                                                 layerName.c_str());
-    }
-    else
-    {
-        layer = m_Network->AddConvolution3dLayer(desc,
-                                                 filterTensorAndData,
-                                                 EmptyOptional(),
-                                                 layerName.c_str());
+
+        // Add the biases input to the registration list, a constant layer will be added by SetupConstantLayers.
+        tensorIndexesToRegister.emplace_back(inputTensorIndexes[2]);
     }
 
+    armnn::IConnectableLayer* layer = m_Network->AddConvolution3dLayer(desc, layerName.c_str());
     ARMNN_ASSERT(layer != nullptr);
 
     armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0], true);
     layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
 
     // Register the input connection slots for the layer, connections are made after all layers have been created
-    // only the tensors for the inputs are relevant, exclude the const tensors
-    auto inputTensorIndexes = AsUnsignedVector(GetInputTensorIds(m_Model, subgraphIndex, operatorIndex));
-    RegisterInputSlots(subgraphIndex, operatorIndex, layer, {inputTensorIndexes[0]});
+    RegisterInputSlots(subgraphIndex, operatorIndex, layer, tensorIndexesToRegister);
 
     layer = AddFusedActivationLayer(layer, 0, options->fused_activation_function);
     // Register the output connection slots for the layer, connections are made after all layers have been created
diff --git a/src/armnnUtils/DataLayoutIndexed.cpp b/src/armnnUtils/DataLayoutIndexed.cpp
index c1c98fc0fd..01505a0a31 100644
--- a/src/armnnUtils/DataLayoutIndexed.cpp
+++ b/src/armnnUtils/DataLayoutIndexed.cpp
@@ -31,6 +31,12 @@ DataLayoutIndexed::DataLayoutIndexed(armnn::DataLayout dataLayout)
             m_WidthIndex    = 3;
             m_ChannelsIndex = 4;
             break;
+        case armnn::DataLayout::NCDHW:
+            m_ChannelsIndex = 1;
+            m_DepthIndex    = 2;
+            m_HeightIndex   = 3;
+            m_WidthIndex    = 4;
+            break;
         default:
             throw armnn::InvalidArgumentException("Unknown DataLayout value: " +
                                                   std::to_string(static_cast<int>(dataLayout)));
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index 27b59ea3a6..2716c827af 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -1320,7 +1320,12 @@ void Convolution3dQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) co
 {
     const std::string descriptorName{"Convolution3dQueueDescriptor"};
 
-    ValidateNumInputs(workloadInfo,  descriptorName, 1);
+    uint32_t numInputs = 2;
+    if (m_Parameters.m_BiasEnabled)
+    {
+        numInputs = 3;
+    }
+    ValidateNumInputs(workloadInfo,  descriptorName, numInputs);
     ValidateNumOutputs(workloadInfo, descriptorName, 1);
 
     const TensorInfo& inputTensorInfo  = workloadInfo.m_InputTensorInfos[0];
@@ -1329,9 +1334,7 @@ void Convolution3dQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) co
     ValidateTensorNumDimensions(inputTensorInfo,  descriptorName, 5, "input");
     ValidateTensorNumDimensions(outputTensorInfo, descriptorName, 5, "output");
 
-    ValidatePointer(m_Weight, descriptorName, "weight");
-
-    const TensorInfo& weightTensorInfo = m_Weight->GetTensorInfo();
+    const TensorInfo& weightTensorInfo = workloadInfo.m_InputTensorInfos[1];
     ValidateTensorNumDimensions(weightTensorInfo, descriptorName, 5, "weight");
 
     ValidateWeightDataType(inputTensorInfo, weightTensorInfo, descriptorName);
@@ -1339,9 +1342,7 @@ void Convolution3dQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) co
     Optional<TensorInfo> optionalBiasTensorInfo;
     if (m_Parameters.m_BiasEnabled)
     {
-        ValidatePointer(m_Bias, descriptorName, "bias");
-
-        optionalBiasTensorInfo = MakeOptional<TensorInfo>(m_Bias->GetTensorInfo());
+        optionalBiasTensorInfo = MakeOptional<TensorInfo>(workloadInfo.m_InputTensorInfos[2]);
         const TensorInfo& biasTensorInfo = optionalBiasTensorInfo.value();
 
         ValidateTensorDataType(biasTensorInfo, GetBiasDataType(inputTensorInfo.GetDataType()), descriptorName, "bias");
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 29d39d14a9..4e56aaf823 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -208,18 +208,9 @@ struct Convolution2dQueueDescriptor : QueueDescriptorWithParameters<Convolution2
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
-// Convolution 2D layer workload data.
+// Convolution 3D layer workload data.
 struct Convolution3dQueueDescriptor : QueueDescriptorWithParameters<Convolution3dDescriptor>
 {
-    Convolution3dQueueDescriptor()
-        : m_Weight(nullptr)
-        , m_Bias(nullptr)
-    {
-    }
-
-    const ConstTensorHandle* m_Weight;
-    const ConstTensorHandle* m_Bias;
-
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index 3b7f3a0f1f..55ce3554f9 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -250,7 +250,11 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId,
             const TensorInfo input  = OverrideDataType(layer.GetInputSlot(0).GetConnection()->GetTensorInfo(),
                                                        dataType);
             const TensorInfo output = OverrideDataType(layer.GetOutputSlot(0).GetTensorInfo(), dataType);
-            ARMNN_ASSERT(cLayer->m_Weight.get() != nullptr);
+
+            ARMNN_ASSERT_MSG(layer.GetInputSlot(1).GetConnection(),
+                             "Convolution3dLayer: Weights should be connected as a Constant Layer.");
+            const TensorInfo weights = OverrideDataType(layer.GetInputSlot(1).GetConnection()->GetTensorInfo(),
+                                                        dataType);
 
             const Convolution3dDescriptor& descriptor = cLayer->GetParameters();
 
@@ -258,14 +262,15 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId,
             Optional<TensorInfo> biases;
             if (descriptor.m_BiasEnabled)
             {
-                biases = OverrideDataType(cLayer->m_Bias->GetTensorInfo(), GetBiasTypeFromWeightsType(dataType));
+                biases = OverrideDataType(layer.GetInputSlot(2).GetConnection()->GetTensorInfo(),
+                                          GetBiasTypeFromWeightsType(dataType));
             }
 
             result = layerSupportObject.IsConvolution3dSupported(
                                               input,
                                               output,
                                               descriptor,
-                                              OverrideDataType(cLayer->m_Weight->GetTensorInfo(), dataType),
+                                              weights,
                                               biases,
                                               reason);
             break;
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index e3221c5ae4..b90407fd7c 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -13,6 +13,7 @@ list(APPEND armnnBackendsCommonUnitTests_sources
     ChannelShuffleEndToEndTestImpl.hpp
     ComparisonEndToEndTestImpl.hpp
     CompatibilityTests.cpp
+    Convolution3dEndToEndTestImpl.hpp
     CustomMemoryOptimizerStrategyTests.cpp
     DefaultAsyncExecuteTest.cpp
     DepthToSpaceEndToEndTestImpl.hpp
diff --git a/src/backends/backendsCommon/test/Convolution3dEndToEndTestImpl.hpp b/src/backends/backendsCommon/test/Convolution3dEndToEndTestImpl.hpp
new file mode 100644
index 0000000000..33bf9a180b
--- /dev/null
+++ b/src/backends/backendsCommon/test/Convolution3dEndToEndTestImpl.hpp
@@ -0,0 +1,167 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+#pragma once
+
+#include "EndToEndTestImpl.hpp"
+#include "QuantizeHelper.hpp"
+
+#include <ResolveType.hpp>
+
+#include <backendsCommon/test/CommonTestUtils.hpp>
+#include <backendsCommon/test/DataLayoutUtils.hpp>
+
+#include <map>
+#include <vector>
+
+namespace
+{
+
+armnn::INetworkPtr CreateConvolution3dNetwork(const armnn::Convolution3dDescriptor& descriptor,
+                                              const armnn::TensorInfo& inputInfo,
+                                              const armnn::TensorInfo& weightsInfo,
+                                              const armnn::TensorInfo& biasInfo,
+                                              const armnn::TensorInfo& outputInfo,
+                                              const armnn::ConstTensor& weights,
+                                              const armnn::ConstTensor& biases)
+{
+    using namespace armnn;
+
+    INetworkPtr network(INetwork::Create());
+    IConnectableLayer* input = network->AddInputLayer(0, "input");
+    armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weights, "Weights");
+    armnn::IConnectableLayer* biasLayer = network->AddConstantLayer(biases, "Bias");
+    IConnectableLayer* convolution3d = network->AddConvolution3dLayer(descriptor, "convolution3d");
+    IConnectableLayer* output = network->AddOutputLayer(0, "output");
+
+    Connect(input, convolution3d, inputInfo, 0, 0);
+    Connect(weightsLayer, convolution3d, weightsInfo, 0, 1);
+    Connect(biasLayer, convolution3d, biasInfo, 0, 2);
+    Connect(convolution3d, output, outputInfo, 0, 0);
+
+    return network;
+}
+
+} // anonymous namespace
+
+template<armnn::DataType ArmnnType, armnn::DataType ArmnnBType>
+void Convolution3dEndToEnd(const std::vector<armnn::BackendId>& backends,
+                           armnn::DataLayout dataLayout)
+{
+    using namespace armnn;
+    using T  = ResolveType<ArmnnType>;
+    using BT = ResolveType<ArmnnBType>;
+
+    const float   qScale  = IsQuantizedType<T>() ? 0.25f : 1.0f;
+    const int32_t qOffset = IsQuantizedType<T>() ? 50    : 0;
+
+    TensorInfo inputInfo({ 1, 5, 5, 5, 1 }, ArmnnType, qScale, qOffset);
+    TensorInfo outputInfo({ 1, 2, 2, 2, 1 }, ArmnnType, qScale, qOffset);
+    TensorInfo weightsInfo({ 3, 3, 3, 1, 1 }, ArmnnType, qScale, qOffset, true);
+    TensorInfo biasesInfo({ 1 }, ArmnnBType, qScale * qScale, 0, true);
+
+    std::vector<float> inputData =
+    {
+        0.0f,  1.0f,  2.0f,  3.0f,  4.0f,
+        5.0f,  6.0f,  7.0f,  8.0f,  9.0f,
+        10.0f, 11.0f, 12.0f, 13.0f, 14.0f,
+        15.0f, 16.0f, 17.0f, 18.0f, 19.0f,
+
+        20.0f, 21.0f, 22.0f, 23.0f, 24.0f,
+        25.0f, 26.0f, 27.0f, 28.0f, 29.0f,
+        30.0f, 31.0f, 32.0f, 33.0f, 34.0f,
+        35.0f, 36.0f, 37.0f, 38.0f, 39.0f,
+        40.0f, 41.0f, 42.0f, 43.0f, 44.0f,
+
+        45.0f, 46.0f, 47.0f, 48.0f, 49.0f,
+        50.0f, 51.0f, 52.0f, 53.0f, 54.0f,
+        55.0f, 56.0f, 57.0f, 58.0f, 59.0f,
+        60.0f, 61.0f, 62.0f, 63.0f, 64.0f,
+        65.0f, 66.0f, 67.0f, 68.0f, 69.0f,
+
+        70.0f, 71.0f, 72.0f, 73.0f, 74.0f,
+        75.0f, 76.0f, 77.0f, 78.0f, 79.0f,
+        80.0f, 81.0f, 82.0f, 83.0f, 84.0f,
+        85.0f, 86.0f, 87.0f, 88.0f, 89.0f,
+        90.0f, 91.0f, 92.0f, 93.0f, 94.0f,
+        95.0f, 96.0f, 97.0f, 98.0f, 99.0f,
+
+        100.0f, 101.0f, 102.0f, 103.0f, 104.0f,
+        105.0f, 106.0f, 107.0f, 108.0f, 109.0f,
+        110.0f, 111.0f, 112.0f, 113.0f, 114.0f,
+        115.0f, 116.0f, 117.0f, 118.0f, 119.0f,
+        120.0f, 121.0f, 122.0f, 123.0f, 124.0f
+    };
+
+    std::vector<float> weightsData =
+    {
+        1.0f, 1.0f, 1.0f,
+        1.0f, 1.0f, 1.0f,
+        1.0f, 1.0f, 1.0f,
+
+        0.0f, 0.0f, 0.0f,
+        0.0f, 0.0f, 0.0f,
+        0.0f, 0.0f, 0.0f,
+
+        1.0f, 1.0f, 1.0f,
+        1.0f, 1.0f, 1.0f,
+        1.0f, 1.0f, 1.0f,
+    };
+
+    std::vector<float> biasesData = { 1.f };
+
+    std::vector<float> expectedOutputData =
+    {
+        559.0f, 595.0f,
+
+        739.0f, 775.0f,
+
+        1459.0f, 1495.0f,
+
+        1639.0f, 1675.0f,
+    };
+
+    Convolution3dDescriptor descriptor;
+    descriptor.m_PadLeft     = 0;
+    descriptor.m_PadRight    = 0;
+    descriptor.m_PadTop      = 0;
+    descriptor.m_PadBottom   = 0;
+    descriptor.m_PadFront    = 0;
+    descriptor.m_PadBack     = 0;
+    descriptor.m_StrideX     = 2;
+    descriptor.m_StrideY     = 2;
+    descriptor.m_StrideZ     = 2;
+    descriptor.m_BiasEnabled = true;
+    descriptor.m_DataLayout  = dataLayout;
+
+    // Permute input and output if NCDHW.
+    if (dataLayout == DataLayout::NCDHW)
+    {
+        PermuteTensorNdhwcToNcdhw(inputInfo, inputData);
+        PermuteTensorNdhwcToNcdhw(outputInfo, expectedOutputData);
+    }
+
+    // Quantize data
+    std::vector<T> qInputData          = armnnUtils::QuantizedVector<T>(inputData, qScale, qOffset);
+    std::vector<T> qWeightsData        = armnnUtils::QuantizedVector<T>(weightsData, qScale, qOffset);
+    std::vector<T> qExpectedOutputData = armnnUtils::QuantizedVector<T>(expectedOutputData, qScale, qOffset);
+
+    std::vector<BT> qBiasesData = armnnUtils::QuantizedVector<BT>(biasesData, qScale * qScale, 0);
+
+    ConstTensor weights(weightsInfo, qWeightsData);
+    ConstTensor biases(biasesInfo, qBiasesData);
+
+    INetworkPtr network = CreateConvolution3dNetwork(descriptor,
+                                                     inputInfo,
+                                                     weightsInfo,
+                                                     biasesInfo,
+                                                     outputInfo,
+                                                     weights,
+                                                     biases);
+
+    EndToEndLayerTestImpl<ArmnnType, ArmnnType>(std::move(network),
+                                                { { 0, qInputData } },
+                                                { { 0, qExpectedOutputData } },
+                                                backends);
+}
diff --git a/src/backends/backendsCommon/test/DataLayoutUtils.hpp b/src/backends/backendsCommon/test/DataLayoutUtils.hpp
index 9411212f4f..89b3900979 100644
--- a/src/backends/backendsCommon/test/DataLayoutUtils.hpp
+++ b/src/backends/backendsCommon/test/DataLayoutUtils.hpp
@@ -34,3 +34,27 @@ void PermuteTensorNhwcToNchw(armnn::TensorInfo& tensorInfo, std::vector<T>& tens
 
     tensorData = tmp;
 }
+
+template<typename T>
+void PermuteTensorNdhwcToNcdhw(armnn::TensorInfo& tensorInfo, std::vector<T>& tensorData)
+{
+    const armnn::PermutationVector ndhwcToNcdhw = { 0, 2, 3, 4, 1 };
+
+    tensorInfo = armnnUtils::Permuted(tensorInfo, ndhwcToNcdhw);
+
+    std::vector<T> tmp(tensorData.size());
+    armnnUtils::Permute(tensorInfo.GetShape(), ndhwcToNcdhw, tensorData.data(), tmp.data(), sizeof(T));
+    tensorData = tmp;
+}
+
+template<typename T>
+void PermuteTensorNcdhwToNdhwc(armnn::TensorInfo& tensorInfo, std::vector<T>& tensorData)
+{
+    const armnn::PermutationVector ncdhwToNdhwc = { 0, 4, 1, 2, 3 };
+
+    tensorInfo = armnnUtils::Permuted(tensorInfo, ncdhwToNdhwc);
+
+    std::vector<T> tmp(tensorData.size());
+    armnnUtils::Permute(tensorInfo.GetShape(), ncdhwToNdhwc, tensorData.data(), tmp.data(), sizeof(T));
+    tensorData = tmp;
+}
diff --git a/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp b/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp
index c3a6aa1a3c..f9bdfde622 100644
--- a/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp
+++ b/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp
@@ -407,7 +407,7 @@ void FullyConnectedErrorChecking(const std::vector<armnn::BackendId>& backends,
             }
             catch (const LayerValidationException& exc)
             {
-                CHECK(strcmp(exc.what(), "FullyConnected layer weights not set: Input slot(s) 1 not connected "
+                CHECK(strcmp(exc.what(), "Fully_Connected layer weights not set: Input slot(s) 1 not connected "
                                          "to an output slot on FullyConnected layer \"Fully_Connected\"") == 0);
             }
         }
@@ -434,7 +434,7 @@ void FullyConnectedErrorChecking(const std::vector<armnn::BackendId>& backends,
             }
             catch (const LayerValidationException& exc)
             {
-                CHECK(strcmp(exc.what(), "FullyConnected layer bias not set: Input slot(s) 2 not connected "
+                CHECK(strcmp(exc.what(), "Fully_Connected layer bias not set: Input slot(s) 2 not connected "
                                          "to an output slot on FullyConnected layer \"Fully_Connected\"") == 0);
             }
         }
@@ -457,7 +457,7 @@ void FullyConnectedErrorChecking(const std::vector<armnn::BackendId>& backends,
         }
         catch (const LayerValidationException& exc)
         {
-            CHECK(strcmp(exc.what(), "FullyConnected layer weights and bias not set: Input slot(s) 1 & 2 not "
+            CHECK(strcmp(exc.what(), "Fully_Connected layer weights and bias not set: Input slot(s) 1 & 2 not "
                                      "connected to an output slot on FullyConnected layer \"Fully_Connected\"") == 0);
         }
 
diff --git a/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.cpp
index 259272d996..1406ab039b 100644
--- a/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.cpp
+++ b/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.cpp
@@ -11,6 +11,7 @@
 
 #include <backendsCommon/TensorHandle.hpp>
 
+#include <backendsCommon/test/DataLayoutUtils.hpp>
 #include <backendsCommon/test/TensorCopyUtils.hpp>
 #include <backendsCommon/test/WorkloadTestUtils.hpp>
 
@@ -228,23 +229,20 @@ LayerTestResult<T, 5> SimpleConvolution3dTestImpl(
                         biasDesc.GetQuantizationScale(), biasDesc.GetQuantizationOffset());
     }
 
+    // Permute input and output if data layout is NCDHW.
+    if (dataLayout == armnn::DataLayout::NCDHW)
+    {
+        PermuteTensorNdhwcToNcdhw(inputTensorInfo, inputData);
+        PermuteTensorNdhwcToNcdhw(outputTensorInfo, outputData);
+    }
+
     std::vector<T> actualOutput(outputTensorInfo.GetNumElements());
 
-    std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> input0Handle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> input1Handle = tensorHandleFactory.CreateTensorHandle(kernelDesc);
     std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
 
-    armnn::ScopedTensorHandle weightsTensor(kernelDesc);
-    AllocateAndCopyDataToITensorHandle(&weightsTensor, kernel.data());
-
-    armnn::ScopedTensorHandle biasTensor(biasDesc);
-    if (biasEnabled)
-    {
-        AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data());
-    }
-
     armnn::Convolution3dQueueDescriptor data;
-    data.m_Weight = &weightsTensor;
-    data.m_Bias = &biasTensor; // Still set this whether or not bias is enabled - it can be a source of bugs.
     data.m_Parameters.m_StrideX = strideX;
     data.m_Parameters.m_StrideY = strideY;
     data.m_Parameters.m_StrideZ = strideZ;
@@ -261,14 +259,29 @@ LayerTestResult<T, 5> SimpleConvolution3dTestImpl(
     data.m_Parameters.m_BiasEnabled = biasEnabled;
 
     armnn::WorkloadInfo info;
-    AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
+    AddInputToWorkload(data, info, inputTensorInfo, input0Handle.get());
+    AddInputToWorkload(data, info, kernelDesc, input1Handle.get());
     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 
+    std::unique_ptr<armnn::ITensorHandle> input2Handle = nullptr;
+    if (biasEnabled)
+    {
+        input2Handle = tensorHandleFactory.CreateTensorHandle(biasDesc);
+        AddInputToWorkload(data, info, biasDesc, input2Handle.get());
+    }
+
     std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConvolution3d(data, info);
-    inputHandle->Allocate();
+    input0Handle->Allocate();
+    input1Handle->Allocate();
     outputHandle->Allocate();
 
-    CopyDataToITensorHandle(inputHandle.get(), inputData.data());
+    CopyDataToITensorHandle(input0Handle.get(), inputData.data());
+    CopyDataToITensorHandle(input1Handle.get(), kernel.data());
+    if (biasEnabled)
+    {
+        input2Handle->Allocate();
+        CopyDataToITensorHandle(input2Handle.get(), bias.data());
+    }
 
     ExecuteWorkload(*workload, memoryManager);
 
@@ -840,40 +853,44 @@ LayerTestResult<float, 5> SimpleConvolution3d3x3x3Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return SimpleConvolution3d3x3x3TestCommon<armnn::DataType::Float32, armnn::DataType::Float32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int8_t, 5> SimpleConvolution3d3x3x3Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return SimpleConvolution3d3x3x3TestCommon<armnn::DataType::QAsymmS8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<uint8_t, 5> SimpleConvolution3d3x3x3Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return SimpleConvolution3d3x3x3TestCommon<armnn::DataType::QAsymmU8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int16_t, 5> SimpleConvolution3d3x3x3Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return SimpleConvolution3d3x3x3TestCommon<armnn::DataType::QSymmS16, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 
@@ -881,158 +898,174 @@ LayerTestResult<float, 5> Convolution3d2x2x2Strides3x5x5Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Strides3x5x5TestCommon<armnn::DataType::Float32, armnn::DataType::Float32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int8_t, 5> Convolution3d2x2x2Strides3x5x5Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Strides3x5x5TestCommon<armnn::DataType::QAsymmS8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<uint8_t, 5> Convolution3d2x2x2Strides3x5x5Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Strides3x5x5TestCommon<armnn::DataType::QAsymmU8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int16_t, 5> Convolution3d2x2x2Strides3x5x5Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Strides3x5x5TestCommon<armnn::DataType::QSymmS16, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<float, 5> Convolution3d2x2x2Dilation2x2x2Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Dilation2x2x2TestCommon<armnn::DataType::Float32, armnn::DataType::Float32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int8_t, 5> Convolution3d2x2x2Dilation2x2x2Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Dilation2x2x2TestCommon<armnn::DataType::QAsymmS8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<uint8_t, 5> Convolution3d2x2x2Dilation2x2x2Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Dilation2x2x2TestCommon<armnn::DataType::QAsymmU8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int16_t, 5> Convolution3d2x2x2Dilation2x2x2Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Dilation2x2x2TestCommon<armnn::DataType::QSymmS16, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<float, 5> Convolution3dPaddingSame3x3x3Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3dPaddingSame3x3x3TestCommon<armnn::DataType::Float32, armnn::DataType::Float32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int8_t, 5> Convolution3dPaddingSame3x3x3Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3dPaddingSame3x3x3TestCommon<armnn::DataType::QAsymmS8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<uint8_t, 5> Convolution3dPaddingSame3x3x3Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3dPaddingSame3x3x3TestCommon<armnn::DataType::QAsymmU8, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<int16_t, 5> Convolution3dPaddingSame3x3x3Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3dPaddingSame3x3x3TestCommon<armnn::DataType::QSymmS16, armnn::DataType::Signed32>(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<float, 5> Convolution3dStrideDilationPadding3x3x3Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3dStrideDilationPadding3x3x3TestCommonFloat32(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<float, 5> Convolution3d2x2x2Stride3x3x3SmallFloat32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2Stride3x3x3SmallTestCommonFloat32(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<armnn::Half, 5> Convolution3d2x3x3Float16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x3x3TestCommonFloat16(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
 
 LayerTestResult<armnn::Half, 5> Convolution3d2x2x2SmallFloat16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled)
+        bool biasEnabled,
+        armnn::DataLayout dataLayout)
 {
     return Convolution3d2x2x2SmallTestCommonFloat16(
-            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, armnn::DataLayout::NDHWC);
+            workloadFactory, memoryManager, tensorHandleFactory, biasEnabled, dataLayout);
 }
diff --git a/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.hpp b/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.hpp
index a07c183c76..c612e19c9b 100644
--- a/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.hpp
+++ b/src/backends/backendsCommon/test/layerTests/Conv3dTestImpl.hpp
@@ -24,118 +24,138 @@ LayerTestResult<float, 5> SimpleConvolution3d3x3x3Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int8_t , 5> SimpleConvolution3d3x3x3Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<uint8_t, 5> SimpleConvolution3d3x3x3Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int16_t, 5> SimpleConvolution3d3x3x3Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<float, 5> Convolution3d2x2x2Strides3x5x5Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int8_t , 5> Convolution3d2x2x2Strides3x5x5Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<uint8_t, 5> Convolution3d2x2x2Strides3x5x5Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int16_t, 5> Convolution3d2x2x2Strides3x5x5Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<float, 5> Convolution3d2x2x2Dilation2x2x2Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int8_t , 5> Convolution3d2x2x2Dilation2x2x2Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<uint8_t, 5> Convolution3d2x2x2Dilation2x2x2Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int16_t, 5> Convolution3d2x2x2Dilation2x2x2Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<float, 5> Convolution3dPaddingSame3x3x3Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int8_t , 5> Convolution3dPaddingSame3x3x3Int8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<uint8_t, 5> Convolution3dPaddingSame3x3x3Uint8Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<int16_t, 5> Convolution3dPaddingSame3x3x3Int16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<float, 5> Convolution3dStrideDilationPadding3x3x3Float32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<float, 5> Convolution3d2x2x2Stride3x3x3SmallFloat32Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<armnn::Half, 5> Convolution3d2x3x3Float16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
 
 LayerTestResult<armnn::Half, 5> Convolution3d2x2x2SmallFloat16Test(
         armnn::IWorkloadFactory& workloadFactory,
         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
         const armnn::ITensorHandleFactory& tensorHandleFactory,
-        bool biasEnabled);
+        bool biasEnabled,
+        armnn::DataLayout dataLayout);
diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp
index 0cc8f4aa10..dc4dcecd81 100644
--- a/src/backends/reference/test/RefEndToEndTests.cpp
+++ b/src/backends/reference/test/RefEndToEndTests.cpp
@@ -11,6 +11,7 @@
 #include <backendsCommon/test/ChannelShuffleEndToEndTestImpl.hpp>
 #include <backendsCommon/test/ComparisonEndToEndTestImpl.hpp>
 #include <backendsCommon/test/ConcatEndToEndTestImpl.hpp>
+#include <backendsCommon/test/Convolution3dEndToEndTestImpl.hpp>
 #include <backendsCommon/test/DepthToSpaceEndToEndTestImpl.hpp>
 #include <backendsCommon/test/DequantizeEndToEndTestImpl.hpp>
 #include <backendsCommon/test/DetectionPostProcessEndToEndTestImpl.hpp>
@@ -566,6 +567,36 @@ TEST_CASE("RefConcatEndToEndDim3Uint8Test")
     ConcatDim3EndToEnd<armnn::DataType::QAsymmU8>(defaultBackends);
 }
 
+TEST_CASE("RefConvolution3dFloat32Test")
+{
+    Convolution3dEndToEnd<armnn::DataType::Float32, armnn::DataType::Float32>(defaultBackends,
+                                                                              armnn::DataLayout::NDHWC);
+}
+
+TEST_CASE("RefConvolution3dNcdhwFloat32Test")
+{
+    Convolution3dEndToEnd<armnn::DataType::Float32, armnn::DataType::Float32>(defaultBackends,
+                                                                              armnn::DataLayout::NCDHW);
+}
+
+TEST_CASE("RefConvolution3dFloat16Test")
+{
+    Convolution3dEndToEnd<armnn::DataType::Float16, armnn::DataType::Float16>(defaultBackends,
+                                                                              armnn::DataLayout::NDHWC);
+}
+
+TEST_CASE("RefConvolution3dUint8Test")
+{
+    Convolution3dEndToEnd<armnn::DataType::QAsymmU8, armnn::DataType::Signed32>(defaultBackends,
+                                                                                armnn::DataLayout::NDHWC);
+}
+
+TEST_CASE("RefConvolution3dInt8Test")
+{
+    Convolution3dEndToEnd<armnn::DataType::QAsymmS8, armnn::DataType::Signed32>(defaultBackends,
+                                                                                armnn::DataLayout::NDHWC);
+}
+
 TEST_CASE("RefEluEndToEndTestFloat32")
 {
     EluEndToEndTest<armnn::DataType::Float32>(defaultBackends);
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index f5d388d007..cb31b37161 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -208,37 +208,119 @@ ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution2d3x3Stride2x2BFloat16SmallValue,
                      false,
                      DataLayout::NHWC);
 
-// Convolution 3d
-ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Float32, SimpleConvolution3d3x3x3Float32Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Int8, SimpleConvolution3d3x3x3Int8Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Uint8, SimpleConvolution3d3x3x3Uint8Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Int16, SimpleConvolution3d3x3x3Int16Test, false)
-
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5Float32, Convolution3d2x2x2Strides3x5x5Float32Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5TestInt8, Convolution3d2x2x2Strides3x5x5Int8Test, true)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5TestUint8, Convolution3d2x2x2Strides3x5x5Uint8Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5TestInt16, Convolution3d2x2x2Strides3x5x5Int16Test, true)
-
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3Float32, Convolution3dPaddingSame3x3x3Float32Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3TestInt8, Convolution3dPaddingSame3x3x3Int8Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3TestUint8, Convolution3dPaddingSame3x3x3Uint8Test, false)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3TestInt16, Convolution3dPaddingSame3x3x3Int16Test, false)
-
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2Float32, Convolution3d2x2x2Dilation2x2x2Float32Test, true)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2TestInt8, Convolution3d2x2x2Dilation2x2x2Int8Test, true)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2TestUint8, Convolution3d2x2x2Dilation2x2x2Uint8Test, true)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2TestInt16, Convolution3d2x2x2Dilation2x2x2Int16Test, true)
+// Convolution 3d - NDHWC
+ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Float32,
+                              SimpleConvolution3d3x3x3Float32Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Int8,
+                              SimpleConvolution3d3x3x3Int8Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Uint8,
+                              SimpleConvolution3d3x3x3Uint8Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3Int16,
+                              SimpleConvolution3d3x3x3Int16Test,
+                              false,
+                              DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5Float32,
+                              Convolution3d2x2x2Strides3x5x5Float32Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5TestInt8,
+                              Convolution3d2x2x2Strides3x5x5Int8Test,
+                              true,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5TestUint8,
+                              Convolution3d2x2x2Strides3x5x5Uint8Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5TestInt16,
+                              Convolution3d2x2x2Strides3x5x5Int16Test,
+                              true,
+                              DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3Float32,
+                              Convolution3dPaddingSame3x3x3Float32Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3TestInt8,
+                              Convolution3dPaddingSame3x3x3Int8Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3TestUint8,
+                              Convolution3dPaddingSame3x3x3Uint8Test,
+                              false,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3TestInt16,
+                              Convolution3dPaddingSame3x3x3Int16Test,
+                              false,
+                              DataLayout::NDHWC)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2Float32,
+                              Convolution3d2x2x2Dilation2x2x2Float32Test,
+                              true,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2TestInt8,
+                              Convolution3d2x2x2Dilation2x2x2Int8Test,
+                              true,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2TestUint8,
+                              Convolution3d2x2x2Dilation2x2x2Uint8Test,
+                              true,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2TestInt16,
+                              Convolution3d2x2x2Dilation2x2x2Int16Test,
+                              true,
+                              DataLayout::NDHWC)
 
 ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dStrideDilationPadding3x3x3Float32,
                               Convolution3dStrideDilationPadding3x3x3Float32Test,
-                              true)
+                              true,
+                              DataLayout::NDHWC)
 
 ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Stride3x3x3SmallTestFloat32,
                               Convolution3d2x2x2Stride3x3x3SmallFloat32Test,
-                              false)
+                              false,
+                              DataLayout::NDHWC)
 
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x3x3TestFloat16, Convolution3d2x3x3Float16Test, true)
-ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2SmallTestFloat16, Convolution3d2x2x2SmallFloat16Test, false)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x3x3TestFloat16,
+                              Convolution3d2x3x3Float16Test,
+                              true,
+                              DataLayout::NDHWC)
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2SmallTestFloat16,
+                              Convolution3d2x2x2SmallFloat16Test,
+                              false,
+                              DataLayout::NDHWC)
+
+// Convolution 3d - NCDHW
+ARMNN_AUTO_TEST_CASE_WITH_THF(SimpleConvolution3d3x3x3NcdhwFloat32,
+                              SimpleConvolution3d3x3x3Float32Test,
+                              false,
+                              DataLayout::NCDHW)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x3x3TestNcdhwFloat16,
+                              Convolution3d2x3x3Float16Test,
+                              false,
+                              DataLayout::NCDHW)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Strides3x5x5NcdhwTestInt8,
+                              Convolution3d2x2x2Strides3x5x5Int8Test,
+                              true,
+                              DataLayout::NCDHW)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3dPaddingSame3x3x3NcdhwTestUint8,
+                              Convolution3dPaddingSame3x3x3Uint8Test,
+                              false,
+                              DataLayout::NCDHW)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(Convolution3d2x2x2Dilation2x2x2NcdhwTestInt16,
+                              Convolution3d2x2x2Dilation2x2x2Int16Test,
+                              true,
+                              DataLayout::NCDHW)
 
 
 // Depthwise Convolution
diff --git a/src/backends/reference/workloads/Conv3dImpl.cpp b/src/backends/reference/workloads/Conv3dImpl.cpp
index 484d887cfc..1c06d624a8 100644
--- a/src/backends/reference/workloads/Conv3dImpl.cpp
+++ b/src/backends/reference/workloads/Conv3dImpl.cpp
@@ -113,11 +113,25 @@ void Convolve3d(const TensorShape& rInputShape,
 
                                             // Keep this implementation, as using DataLayoutIndexed::GetIndex
                                             // causes large performance regression.
-                                            inputIndex = batchIdx * inputDepth * inputHeight * inputWidth * inChannels +
-                                                         (zInput-paddingFront) * inputHeight * inputWidth * inChannels +
-                                                         (yInput-paddingTop) * inputWidth * inChannels +
-                                                         (xInput-paddingLeft) * inChannels +
-                                                         cInput;
+                                            if (dataLayoutIndexed.GetDataLayout() == DataLayout::NDHWC)
+                                            {
+                                                inputIndex =
+                                                        batchIdx * inputDepth * inputHeight * inputWidth * inChannels +
+                                                        (zInput-paddingFront) * inputHeight * inputWidth * inChannels +
+                                                        (yInput-paddingTop) * inputWidth * inChannels +
+                                                        (xInput-paddingLeft) * inChannels +
+                                                        cInput;
+                                            }
+                                            else
+                                            {
+                                                // NCDHW DataLayout
+                                                inputIndex =
+                                                        batchIdx * inputDepth * inputHeight * inputWidth * inChannels +
+                                                        inputDepth * inputHeight * inputWidth * cInput +
+                                                        (zInput-paddingFront) * inputHeight * inputWidth +
+                                                        (yInput-paddingTop) * inputWidth +
+                                                        xInput-paddingLeft;
+                                            }
 
                                             inputValue = inputVec[inputIndex];
                                         }
@@ -133,11 +147,24 @@ void Convolve3d(const TensorShape& rInputShape,
                             sum += biasVec[cOutput];
                         }
 
-                        unsigned int outIdx = batchIdx * outputDepth * outputHeight * outputWidth * outChannels +
-                                              zOutput * outputHeight * outputWidth * outChannels +
-                                              yOutput * outputWidth * outChannels +
-                                              xOutput * outChannels +
-                                              cOutput;
+                        unsigned int outIdx;
+                        if (dataLayoutIndexed.GetDataLayout() == DataLayout::NDHWC)
+                        {
+                            outIdx = batchIdx * outputDepth * outputHeight * outputWidth * outChannels +
+                                     zOutput * outputHeight * outputWidth * outChannels +
+                                     yOutput * outputWidth * outChannels +
+                                     xOutput * outChannels +
+                                     cOutput;
+                        }
+                        else
+                        {
+                            // NCDHW DataLayout
+                            outIdx = batchIdx * outputDepth * outputHeight * outputWidth * outChannels +
+                                     cOutput * outputDepth * outputHeight * outputWidth +
+                                     zOutput * outputHeight * outputWidth +
+                                     yOutput * outputWidth +
+                                     xOutput;
+                        }
 
                         rOutputEncoder[outIdx];
                         rOutputEncoder.Set(sum);
diff --git a/src/backends/reference/workloads/RefConvolution3dWorkload.cpp b/src/backends/reference/workloads/RefConvolution3dWorkload.cpp
index ea425daec9..afab88f0a8 100644
--- a/src/backends/reference/workloads/RefConvolution3dWorkload.cpp
+++ b/src/backends/reference/workloads/RefConvolution3dWorkload.cpp
@@ -19,10 +19,10 @@ RefConvolution3dWorkload::RefConvolution3dWorkload(
     WorkloadInfo detailsInfo;
     detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
     detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
-    detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo());
+    detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]);
     if (descriptor.m_Parameters.m_BiasEnabled)
     {
-        detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo());
+        detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]);
     }
 
     // Report Profiling Details
@@ -30,18 +30,25 @@ RefConvolution3dWorkload::RefConvolution3dWorkload(
                                          descriptor.m_Parameters,
                                          detailsInfo,
                                          this->GetGuid());
+}
 
-    m_Weight = std::make_unique<ScopedTensorHandle>(*( descriptor.m_Weight ));
-    const TensorInfo& rFilterInfo = m_Weight->GetTensorInfo();
+void RefConvolution3dWorkload::PostAllocationConfigure()
+{
+    PostAllocationConfigure(m_Data.m_Inputs, m_Data.m_Outputs);
+}
 
+void RefConvolution3dWorkload::PostAllocationConfigure(std::vector<ITensorHandle*> inputs,
+                                                       std::vector<ITensorHandle*> outputs)
+{
+    IgnoreUnused(outputs);
+    const TensorInfo& rFilterInfo = GetTensorInfo(inputs[1]);
     m_FilterShape = rFilterInfo.GetShape();
-    m_FilterDecoder = MakeDecoder<float>(rFilterInfo, m_Weight.get()->Map(true));
+    m_FilterDecoder = MakeDecoder<float>(rFilterInfo);
 
-    if ( descriptor.m_Parameters.m_BiasEnabled )
+    if (m_Data.m_Parameters.m_BiasEnabled)
     {
-        m_Bias = std::make_unique<ScopedTensorHandle>(*( descriptor.m_Bias ));
-        const TensorInfo& biasInfo = m_Bias->GetTensorInfo();
-        m_BiasDecoder = MakeDecoder<float>(biasInfo, m_Bias->Map(true));
+        const TensorInfo& biasInfo = GetTensorInfo(inputs[2]);
+        m_BiasDecoder = MakeDecoder<float>(biasInfo);
     }
 }
 
@@ -52,6 +59,8 @@ void RefConvolution3dWorkload::Execute() const
 
 void RefConvolution3dWorkload::ExecuteAsync(WorkingMemDescriptor& workingMemDescriptor)
 {
+    PostAllocationConfigure(workingMemDescriptor.m_Inputs, workingMemDescriptor.m_Outputs);
+
     Execute(workingMemDescriptor.m_Inputs, workingMemDescriptor.m_Outputs);
 }
 
@@ -65,6 +74,12 @@ void RefConvolution3dWorkload::Execute(std::vector<ITensorHandle*> inputs, std::
     const TensorShape& inputShape = GetTensorInfo(inputs[0]).GetShape();
     const TensorShape& outputShape = GetTensorInfo(outputs[0]).GetShape();
 
+    m_FilterDecoder->Reset(inputs[1]->Map());
+    if (m_Data.m_Parameters.m_BiasEnabled)
+    {
+        m_BiasDecoder->Reset(inputs[2]->Map());
+    }
+
     Convolve3d(inputShape, *inputDecoder, outputShape, *outputEncoder, m_FilterShape,
                *m_FilterDecoder, m_Data.m_Parameters.m_BiasEnabled, m_BiasDecoder.get(),
                m_Data.m_Parameters.m_DataLayout,
diff --git a/src/backends/reference/workloads/RefConvolution3dWorkload.hpp b/src/backends/reference/workloads/RefConvolution3dWorkload.hpp
index 0373a8b900..4d97512095 100644
--- a/src/backends/reference/workloads/RefConvolution3dWorkload.hpp
+++ b/src/backends/reference/workloads/RefConvolution3dWorkload.hpp
@@ -19,14 +19,14 @@ public:
     explicit RefConvolution3dWorkload(const Convolution3dQueueDescriptor& descriptor,
                                       const WorkloadInfo& info);
 
+    void PostAllocationConfigure() override;
 
     void Execute() const override;
     void ExecuteAsync(WorkingMemDescriptor& workingMemDescriptor)  override;
 
 private:
+    void PostAllocationConfigure(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs);
     void Execute(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs) const;
-    std::unique_ptr<ScopedTensorHandle> m_Weight;
-    std::unique_ptr<ScopedTensorHandle> m_Bias;
 
     std::unique_ptr<Decoder<float>> m_FilterDecoder;
     std::unique_ptr<Decoder<float>> m_BiasDecoder;