IVGCVSW-6946 Add Pool3D to tflite delegate

 * Add new test and test helper for Pool3d
 * Add new custom operator to switch in armnn_delegate.cpp
 * Add new pool3d function to pooling.hpp
 * Update doxygen

Signed-off-by: Ryan OShea <ryan.oshea3@arm.com>
Change-Id: I77a541bf423b337c749e70c564cdd727efe2fd05

6 files changed, 933 insertions, 11 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index d488de4c9c..523214bb90 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -1,5 +1,5 @@
 #
-# Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+# Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
 # SPDX-License-Identifier: MIT
 #
 
@@ -177,6 +177,8 @@ if(BUILD_UNIT_TESTS)
         src/test/PadTestHelper.hpp
         src/test/Pooling2dTest.cpp
         src/test/Pooling2dTestHelper.hpp
+        src/test/Pooling3dTest.cpp
+        src/test/Pooling3dTestHelper.hpp
         src/test/PreluTest.cpp
         src/test/PreluTestHelper.hpp
         src/test/QuantizationTest.cpp
diff --git a/delegate/src/Pooling.hpp b/delegate/src/Pooling.hpp
index 4095ac4ac2..df8c3db727 100644
--- a/delegate/src/Pooling.hpp
+++ b/delegate/src/Pooling.hpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -11,15 +11,16 @@
 #include <tensorflow/lite/c/builtin_op_data.h>
 #include <tensorflow/lite/c/common.h>
 #include <tensorflow/lite/minimal_logging.h>
+#include <flatbuffers/flexbuffers.h>
 
 namespace armnnDelegate
 {
 
-TfLiteStatus VisitPoolingOperator(DelegateData& delegateData,
-                                  TfLiteContext* tfLiteContext,
-                                  TfLiteNode* tfLiteNode,
-                                  int nodeIndex,
-                                  int32_t tfLitePoolingOperatorCode)
+TfLiteStatus VisitPooling2dOperator(DelegateData& delegateData,
+                                    TfLiteContext* tfLiteContext,
+                                    TfLiteNode* tfLiteNode,
+                                    int nodeIndex,
+                                    int32_t tfLitePoolingOperatorCode)
 {
     TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
     TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
@@ -113,4 +114,167 @@ TfLiteStatus VisitPoolingOperator(DelegateData& delegateData,
     return FusedActivation(tfLiteContext, tfLiteNode, activationType, poolingLayer, 0, delegateData);
 }
 
+TfLiteStatus VisitPooling3dOperator(DelegateData& delegateData,
+                                    TfLiteContext* tfLiteContext,
+                                    TfLiteNode* tfLiteNode,
+                                    int nodeIndex,
+                                    std::string customOperatorName)
+{
+    TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
+    TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
+
+    const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors;
+    const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]];
+    if (IsDynamicTensor(tfLiteInputTensor))
+    {
+        TF_LITE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ",
+            customOperatorName.c_str(), nodeIndex);
+        return kTfLiteError;
+    }
+
+    const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
+    if (IsDynamicTensor(tfLiteOutputTensor))
+    {
+        TF_LITE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnDelegate: Dynamic output tensors are not supported in operator #%d node #%d: ",
+            customOperatorName.c_str(), nodeIndex);
+        return kTfLiteError;
+    }
+    // Set the input and output info
+    const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
+
+    // Custom Operators are defined by the name string associated to the operator. Use this to determine
+    // which pooling algorithm to create the armnn operator with. L2 Pooling3D is unsupported in TfLite.
+    armnn::PoolingAlgorithm poolingAlgorithm;
+    if (customOperatorName == "MaxPool3D")
+    {
+        poolingAlgorithm = armnn::PoolingAlgorithm::Max;
+    }
+    else if (customOperatorName == "AveragePool3D")
+    {
+        poolingAlgorithm = armnn::PoolingAlgorithm::Average;
+    }
+    else
+    {
+        return kTfLiteError;
+    }
+    // Create the armnn pool3d descriptor and set the algorithm parsed above.
+    armnn::Pooling3dDescriptor descriptor;
+    descriptor.m_PoolType = poolingAlgorithm;
+
+    // custom_initial_data and custom_initial_data_size are void* variables defined in the tflite registration
+    // used to access the custom option buffer for the operator.
+    auto custom_data = tfLiteNode->custom_initial_data;
+    auto custom_data_size = tfLiteNode->custom_initial_data_size;
+    // Reinterpret the void* to a byte buffer to access the options data in the flexbuffers map.
+    const flexbuffers::Map& m = flexbuffers::GetRoot(reinterpret_cast<const uint8_t*>(custom_data),
+                                                     custom_data_size).AsMap();
+    // poolDims is a vector of [ 1, Depth, Height, Width, 1 ]
+    const auto poolDims = m["ksize"].AsTypedVector();
+    descriptor.m_PoolWidth = poolDims[3].AsInt32();
+    descriptor.m_PoolHeight = poolDims[2].AsInt32();
+    descriptor.m_PoolDepth = poolDims[1].AsInt32();
+
+    // strideDimes is a vector of [ 1, Z, Y, X, 1]
+    const auto strideDims = m["strides"].AsTypedVector();
+    descriptor.m_StrideX = strideDims[3].AsInt32();
+    descriptor.m_StrideY = strideDims[2].AsInt32();
+    descriptor.m_StrideZ = strideDims[1].AsInt32();
+    descriptor.m_DataLayout = armnn::DataLayout::NDHWC;
+
+    unsigned int inputDepth = inputTensorInfo.GetShape()[1];
+    unsigned int inputHeight = inputTensorInfo.GetShape()[2];
+    unsigned int inputWidth = inputTensorInfo.GetShape()[3];
+
+    // CalcPadding expects a TfLitePadding type. Parse flexbuffers to extract padding string and create TfLitePadding.
+    std::string paddingStr = m["padding"].AsString().str();
+    TfLitePadding padding;
+    if (paddingStr == "VALID")
+    {
+        padding = kTfLitePaddingValid;
+    }
+    else if (paddingStr == "SAME")
+    {
+        padding = kTfLitePaddingSame;
+    }
+    else
+    {
+        padding = kTfLitePaddingUnknown;
+    }
+    // Calculates padding for each pooling dimension separately
+    CalcPadding(inputHeight, descriptor.m_PoolHeight, descriptor.m_StrideY, 1u,
+                descriptor.m_PadTop, descriptor.m_PadBottom, padding);
+    CalcPadding(inputWidth, descriptor.m_PoolWidth, descriptor.m_StrideX, 1u,
+                descriptor.m_PadLeft, descriptor.m_PadRight, padding);
+    CalcPadding(inputDepth, descriptor.m_PoolDepth, descriptor.m_StrideZ, 1u,
+                descriptor.m_PadFront, descriptor.m_PadBack, padding);
+
+    // Validate the output info.
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outputTensorInfo, bool& isSupported) {
+        FORWARD_LAYER_SUPPORT_FUNC("POOLING_3D",
+                                   tfLiteContext,
+                                   IsPooling3dSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputTensorInfo,
+                                   outputTensorInfo,
+                                   descriptor);
+    };
+
+    if (!delegateData.m_Network)
+    {
+        validateFunc(outputTensorInfo, isSupported);
+        return isSupported ? kTfLiteOk : kTfLiteError;
+    }
+
+    // Create the Layer
+    armnn::IConnectableLayer* poolingLayer = delegateData.m_Network->AddPooling3dLayer(descriptor);
+    ARMNN_ASSERT(poolingLayer != nullptr);
+
+    // Create and set output slots
+    armnn::IOutputSlot& outputSlot = poolingLayer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
+    Connect(poolingLayer, tfLiteNode, delegateData);
+
+    // Check activation by parsing the string from the flexbuffer map
+    std::string activationTypeStr = m["activation"].AsString().str();
+    TfLiteFusedActivation activationType;
+
+    if (activationTypeStr == "kTfLiteActRelu")
+    {
+        activationType = kTfLiteActRelu;
+    }
+    else if (activationTypeStr == "kTfLiteActReluN1To1")
+    {
+        activationType = kTfLiteActReluN1To1;
+    }
+    else if (activationTypeStr == "kTfLiteActRelu6")
+    {
+        activationType = kTfLiteActRelu6;
+    }
+    else if (activationTypeStr == "kTfLiteActTanh")
+    {
+        activationType = kTfLiteActTanh;
+    }
+    else if (activationTypeStr == "kTfLiteActSignBit")
+    {
+        activationType = kTfLiteActSignBit;
+    }
+    else if (activationTypeStr == "kTfLiteActSigmoid")
+    {
+        activationType = kTfLiteActSigmoid;
+    }
+    else
+    {
+        activationType = kTfLiteActNone;
+    }
+
+    return FusedActivation(tfLiteContext, tfLiteNode, activationType, poolingLayer, 0, delegateData);
+}
+
 } // namespace armnnDelegate
diff --git a/delegate/src/armnn_delegate.cpp b/delegate/src/armnn_delegate.cpp
index 6e1a91f9e4..bb2f3c319a 100644
--- a/delegate/src/armnn_delegate.cpp
+++ b/delegate/src/armnn_delegate.cpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -495,6 +495,32 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
 {
     switch (tfLiteRegistration->builtin_code)
     {
+        case kTfLiteBuiltinCustom:
+        {
+#if defined(ARMNN_POST_TFLITE_2_5)
+            // Custom operators are defined by the name rather than the builtin code.
+            // Parse the custom_name param in the registration to point to the correct visitor function.
+            std::string customOperatorName = tfLiteRegistration->custom_name;
+            if ( customOperatorName == "AveragePool3D" )
+            {
+                return VisitPooling3dOperator(delegateData,
+                                            tfLiteContext,
+                                            tfLiteNode,
+                                            nodeIndex,
+                                            customOperatorName);
+            }
+            else if (customOperatorName == "MaxPool3D")
+            {
+                return VisitPooling3dOperator(delegateData,
+                                            tfLiteContext,
+                                            tfLiteNode,
+                                            nodeIndex,
+                                            customOperatorName);
+            }
+#endif
+            // Invalid or unsupported custom operator
+            return kTfLiteError;
+        }
         case kTfLiteBuiltinAbs:
             return VisitElementwiseUnaryOperator(delegateData,
                                                  tfLiteContext,
@@ -520,7 +546,7 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                           nodeIndex,
                                           kTfLiteBuiltinArgMin);
         case kTfLiteBuiltinAveragePool2d:
-            return VisitPoolingOperator(delegateData,
+            return VisitPooling2dOperator(delegateData,
                                         tfLiteContext,
                                         tfLiteNode,
                                         nodeIndex,
@@ -667,7 +693,7 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                                 nodeIndex,
                                                 kTfLiteBuiltinL2Normalization);
         case kTfLiteBuiltinL2Pool2d:
-            return VisitPoolingOperator(delegateData,
+            return VisitPooling2dOperator(delegateData,
                                         tfLiteContext,
                                         tfLiteNode,
                                         nodeIndex,
@@ -729,7 +755,7 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                      nodeIndex,
                                      kTfLiteBuiltinLstm);
         case kTfLiteBuiltinMaxPool2d:
-            return VisitPoolingOperator(delegateData,
+            return VisitPooling2dOperator(delegateData,
                                         tfLiteContext,
                                         tfLiteNode,
                                         nodeIndex,
diff --git a/delegate/src/test/Pooling3dTest.cpp b/delegate/src/test/Pooling3dTest.cpp
new file mode 100644
index 0000000000..c0a186210e
--- /dev/null
+++ b/delegate/src/test/Pooling3dTest.cpp
@@ -0,0 +1,431 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "Pooling3dTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/interpreter.h>
+#include <tensorflow/lite/kernels/register.h>
+#include <tensorflow/lite/model.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+#include <tensorflow/lite/version.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+// Pool3D custom op was only added in tflite r2.6.
+#if defined(ARMNN_POST_TFLITE_2_5)
+
+void MaxPool3dFP32PaddingValidTest(std::vector<armnn::BackendId>& backends)
+{
+    // Set input and expected output data
+    std::vector<int32_t> inputShape = { 1, 2, 3, 4, 1 };
+    std::vector<int32_t> outputShape = { 1, 1, 2, 3, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 6, 6, 4 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kMax";
+    TfLitePadding padding = kTfLitePaddingValid;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         1,
+                         1,
+                         1,
+                         2,
+                         2,
+                         2);
+}
+
+void MaxPool3dFP32PaddingSameTest(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data and expected output data
+    std::vector<int32_t> inputShape = { 1, 2, 3, 4, 1 };
+    std::vector<int32_t> outputShape = { 1, 2, 3, 4, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 6, 6, 4, 4, 6, 6, 6, 6, 4, 5, 6, 6, 6, 6, 4, 4 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kMax";
+    TfLitePadding padding = kTfLitePaddingSame;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         1,
+                         1,
+                         1,
+                         2,
+                         2,
+                         2);
+}
+
+void MaxPool3dFP32H1Test(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data and expected output data
+    std::vector<int32_t> inputShape = { 1, 2, 3, 4, 1 };
+    std::vector<int32_t> outputShape = { 1, 1, 3, 3, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 2, 3 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kMax";
+    TfLitePadding padding = kTfLitePaddingValid;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         1,
+                         1,
+                         1,
+                         2,
+                         1,
+                         2);
+}
+
+void MaxPool3dFP32Test(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data and expected output data
+    std::vector<int32_t> inputShape = { 1, 2, 3, 4, 1 };
+    std::vector<int32_t> outputShape = { 1, 2, 3, 4, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 6, 6 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kMax";
+    TfLitePadding padding = kTfLitePaddingUnknown;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         1,
+                         1,
+                         1,
+                         2,
+                         2,
+                         2);
+}
+
+void AveragePool3dFP32PaddingValidTest(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data and expected output data.
+    std::vector<int32_t> inputShape = { 1, 2, 3, 4, 1 };
+    std::vector<int32_t> outputShape = { 1, 1, 2, 3, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 3.5, 3, 2.5 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kAverage";
+    TfLitePadding padding = kTfLitePaddingValid;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         1,
+                         1,
+                         1,
+                         2,
+                         2,
+                         2);
+}
+
+void AveragePool3dFP32PaddingSameTest(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data and expected output data
+    std::vector<int32_t> inputShape = { 4, 2, 3, 1, 1 };
+    std::vector<int32_t> outputShape = { 4, 2, 3, 1, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 3, 4, 4.5, 4.5, 5.5, 6, 3, 4, 4.5, 4.5, 5.5, 6, 3, 4, 4.5, 4.5 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kAverage";
+    TfLitePadding padding = kTfLitePaddingSame;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         1,
+                         1,
+                         1,
+                         2,
+                         2,
+                         2);
+}
+
+void AveragePool3dFP32H1Test(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data and expected output data
+    std::vector<int32_t> inputShape = { 1, 2, 3, 4, 1 };
+    std::vector<int32_t> outputShape = { 1, 1, 2, 2, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 1.5, 3.5 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kAverage";
+    TfLitePadding padding = kTfLitePaddingUnknown;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         2,
+                         2,
+                         2,
+                         2,
+                         1,
+                         2);
+}
+
+void AveragePool3dFP32Test(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data and expected output data
+    std::vector<int32_t> inputShape = { 4, 3, 2, 1, 1 };
+    std::vector<int32_t> outputShape = { 1, 2, 2, 4, 1 };
+
+    std::vector<float> inputValues = { 1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6,
+                                       1, 2, 3, 4, 5, 6 };
+    std::vector<float> expectedOutputValues = { 3.125, 4.25 };
+
+    // poolType string required to create the correct pooling operator
+    // Padding type required to create the padding in custom options
+    std::string poolType = "kMax";
+    TfLitePadding padding = kTfLitePaddingUnknown;
+
+    Pooling3dTest<float>(poolType,
+                         ::tflite::TensorType_FLOAT32,
+                         backends,
+                         inputShape,
+                         outputShape,
+                         inputValues,
+                         expectedOutputValues,
+                         padding,
+                         2,
+                         2,
+                         2,
+                         2,
+                         2,
+                         2);
+}
+
+TEST_SUITE("Pooling3d_GpuAccTests")
+{
+
+TEST_CASE ("MaxPooling3d_FP32_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    MaxPool3dFP32Test(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_PaddingValid_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    MaxPool3dFP32PaddingValidTest(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_PaddingSame_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    MaxPool3dFP32PaddingSameTest(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_H1_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    MaxPool3dFP32H1Test(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_PaddingValid_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    AveragePool3dFP32PaddingValidTest(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_PaddingSame_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    AveragePool3dFP32PaddingSameTest(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_H1_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    AveragePool3dFP32H1Test(backends);
+}
+
+} // TEST_SUITE("Pooling3d_GpuAccTests")
+
+TEST_SUITE("Pooling3d_CpuAccTests")
+{
+
+TEST_CASE ("MaxPooling3d_FP32_PaddingValid_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    MaxPool3dFP32PaddingValidTest(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_PaddingSame_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    MaxPool3dFP32PaddingSameTest(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    MaxPool3dFP32Test(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_H1_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    MaxPool3dFP32H1Test(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_PaddingValid_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    AveragePool3dFP32PaddingValidTest(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_PaddingSame_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    AveragePool3dFP32PaddingSameTest(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_H1_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    AveragePool3dFP32H1Test(backends);
+}
+
+} // TEST_SUITE("Pooling3d_CpuAccTests")
+
+TEST_SUITE("Pooling3d_CpuRefTests")
+{
+TEST_CASE ("MaxPooling3d_FP32_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    MaxPool3dFP32Test(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_PaddingValid_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    MaxPool3dFP32PaddingValidTest(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_PaddingSame_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    MaxPool3dFP32PaddingSameTest(backends);
+}
+
+TEST_CASE ("MaxPooling3d_FP32_H1_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    MaxPool3dFP32H1Test(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_PaddingValid_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    AveragePool3dFP32PaddingValidTest(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_PaddingSame_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    AveragePool3dFP32PaddingSameTest(backends);
+}
+
+TEST_CASE ("AveragePooling3d_FP32_H1_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    AveragePool3dFP32H1Test(backends);
+}
+
+} // TEST_SUITE("Pooling3d_CpuRefTests")
+
+#endif
+
+}
\ No newline at end of file
diff --git a/delegate/src/test/Pooling3dTestHelper.hpp b/delegate/src/test/Pooling3dTestHelper.hpp
new file mode 100644
index 0000000000..f5f5cc3809
--- /dev/null
+++ b/delegate/src/test/Pooling3dTestHelper.hpp
@@ -0,0 +1,295 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "TestUtils.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <flatbuffers/flexbuffers.h>
+#include <tensorflow/lite/interpreter.h>
+#include <tensorflow/lite/kernels/custom_ops_register.h>
+#include <tensorflow/lite/kernels/register.h>
+#include <tensorflow/lite/model.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+#include <tensorflow/lite/version.h>
+
+#include <doctest/doctest.h>
+
+namespace
+{
+#if defined(ARMNN_POST_TFLITE_2_5)
+
+std::vector<uint8_t> CreateCustomOptions(int, int, int, int, int, int, TfLitePadding);
+
+std::vector<char> CreatePooling3dTfLiteModel(
+    std::string poolType,
+    tflite::TensorType tensorType,
+    const std::vector<int32_t>& inputTensorShape,
+    const std::vector<int32_t>& outputTensorShape,
+    TfLitePadding padding = kTfLitePaddingSame,
+    int32_t strideWidth = 0,
+    int32_t strideHeight = 0,
+    int32_t strideDepth = 0,
+    int32_t filterWidth = 0,
+    int32_t filterHeight = 0,
+    int32_t filterDepth = 0,
+    tflite::ActivationFunctionType fusedActivation = tflite::ActivationFunctionType_NONE,
+    float quantScale = 1.0f,
+    int quantOffset = 0)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
+    buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+
+    auto quantizationParameters =
+        CreateQuantizationParameters(flatBufferBuilder,
+                                     0,
+                                     0,
+                                     flatBufferBuilder.CreateVector<float>({ quantScale }),
+                                     flatBufferBuilder.CreateVector<int64_t>({ quantOffset }));
+
+    // Create the input and output tensors
+    std::array<flatbuffers::Offset<Tensor>, 2> tensors;
+    tensors[0] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(),
+                                                                      inputTensorShape.size()),
+                              tensorType,
+                              0,
+                              flatBufferBuilder.CreateString("input"),
+                              quantizationParameters);
+
+    tensors[1] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(),
+                                                                      outputTensorShape.size()),
+                              tensorType,
+                              0,
+                              flatBufferBuilder.CreateString("output"),
+                              quantizationParameters);
+
+    // Create the custom options from the function below
+    std::vector<uint8_t> customOperatorOptions = CreateCustomOptions(strideHeight, strideWidth, strideDepth,
+                                                                     filterHeight, filterWidth, filterDepth, padding);
+    // opCodeIndex is created as a uint8_t to avoid map lookup
+    uint8_t opCodeIndex = 0;
+    // Set the operator name based on the PoolType passed in from the test case
+    std::string opName = "";
+    if (poolType == "kMax")
+    {
+        opName = "MaxPool3D";
+    }
+    else
+    {
+        opName = "AveragePool3D";
+    }
+    // To create a custom operator code you pass in the builtin code for custom operators and the name of the custom op
+    flatbuffers::Offset<OperatorCode> operatorCode = CreateOperatorCodeDirect(flatBufferBuilder,
+                                                                              tflite::BuiltinOperator_CUSTOM,
+                                                                              opName.c_str());
+
+    // Create the Operator using the opCodeIndex and custom options. Also sets builtin options to none.
+    const std::vector<int32_t> operatorInputs{ 0 };
+    const std::vector<int32_t> operatorOutputs{ 1 };
+    flatbuffers::Offset<Operator> poolingOperator =
+        CreateOperator(flatBufferBuilder,
+                       opCodeIndex,
+                       flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                       tflite::BuiltinOptions_NONE,
+                       0,
+                       flatBufferBuilder.CreateVector<uint8_t>(customOperatorOptions),
+                       tflite::CustomOptionsFormat_FLEXBUFFERS);
+
+    // Create the subgraph using the operator created above.
+    const std::vector<int> subgraphInputs{ 0 };
+    const std::vector<int> subgraphOutputs{ 1 };
+    flatbuffers::Offset<SubGraph> subgraph =
+        CreateSubGraph(flatBufferBuilder,
+                       flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(), subgraphOutputs.size()),
+                       flatBufferBuilder.CreateVector(&poolingOperator, 1));
+
+    flatbuffers::Offset<flatbuffers::String> modelDescription =
+        flatBufferBuilder.CreateString("ArmnnDelegate: Pooling3d Operator Model");
+
+    // Create the model using operatorCode and the subgraph.
+    flatbuffers::Offset<Model> flatbufferModel =
+        CreateModel(flatBufferBuilder,
+                    TFLITE_SCHEMA_VERSION,
+                    flatBufferBuilder.CreateVector(&operatorCode, 1),
+                    flatBufferBuilder.CreateVector(&subgraph, 1),
+                    modelDescription,
+                    flatBufferBuilder.CreateVector(buffers.data(), buffers.size()));
+
+    flatBufferBuilder.Finish(flatbufferModel);
+
+    return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
+                             flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
+}
+
+template<typename T>
+void Pooling3dTest(std::string poolType,
+                   tflite::TensorType tensorType,
+                   std::vector<armnn::BackendId>& backends,
+                   std::vector<int32_t>& inputShape,
+                   std::vector<int32_t>& outputShape,
+                   std::vector<T>& inputValues,
+                   std::vector<T>& expectedOutputValues,
+                   TfLitePadding padding = kTfLitePaddingSame,
+                   int32_t strideWidth = 0,
+                   int32_t strideHeight = 0,
+                   int32_t strideDepth = 0,
+                   int32_t filterWidth = 0,
+                   int32_t filterHeight = 0,
+                   int32_t filterDepth = 0,
+                   tflite::ActivationFunctionType fusedActivation = tflite::ActivationFunctionType_NONE,
+                   float quantScale = 1.0f,
+                   int quantOffset = 0)
+{
+    using namespace tflite;
+    // Create the single op model buffer
+    std::vector<char> modelBuffer = CreatePooling3dTfLiteModel(poolType,
+                                                               tensorType,
+                                                               inputShape,
+                                                               outputShape,
+                                                               padding,
+                                                               strideWidth,
+                                                               strideHeight,
+                                                               strideDepth,
+                                                               filterWidth,
+                                                               filterHeight,
+                                                               filterDepth,
+                                                               fusedActivation,
+                                                               quantScale,
+                                                               quantOffset);
+
+    const Model* tfLiteModel = GetModel(modelBuffer.data());
+    CHECK(tfLiteModel != nullptr);
+    // Create TfLite Interpreters
+    std::unique_ptr<Interpreter> armnnDelegateInterpreter;
+
+    // Custom ops need to be added to the BuiltinOp resolver before the interpreter is created
+    // Based on the poolType from the test case add the custom operator using the name and the tflite
+    // registration function
+    tflite::ops::builtin::BuiltinOpResolver armnn_op_resolver;
+    if (poolType == "kMax")
+    {
+        armnn_op_resolver.AddCustom("MaxPool3D", tflite::ops::custom::Register_MAX_POOL_3D());
+    }
+    else
+    {
+        armnn_op_resolver.AddCustom("AveragePool3D", tflite::ops::custom::Register_AVG_POOL_3D());
+    }
+
+    CHECK(InterpreterBuilder(tfLiteModel, armnn_op_resolver)
+              (&armnnDelegateInterpreter) == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter != nullptr);
+    CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk);
+
+    std::unique_ptr<Interpreter> tfLiteInterpreter;
+
+    // Custom ops need to be added to the BuiltinOp resolver before the interpreter is created
+    // Based on the poolType from the test case add the custom operator using the name and the tflite
+    // registration function
+    tflite::ops::builtin::BuiltinOpResolver tflite_op_resolver;
+    if (poolType == "kMax")
+    {
+        tflite_op_resolver.AddCustom("MaxPool3D", tflite::ops::custom::Register_MAX_POOL_3D());
+    }
+    else
+    {
+        tflite_op_resolver.AddCustom("AveragePool3D", tflite::ops::custom::Register_AVG_POOL_3D());
+    }
+
+    CHECK(InterpreterBuilder(tfLiteModel, tflite_op_resolver)
+              (&tfLiteInterpreter) == kTfLiteOk);
+    CHECK(tfLiteInterpreter != nullptr);
+    CHECK(tfLiteInterpreter->AllocateTensors() == kTfLiteOk);
+
+    // Create the ArmNN Delegate
+    armnnDelegate::DelegateOptions delegateOptions(backends);
+    std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
+        theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
+                         armnnDelegate::TfLiteArmnnDelegateDelete);
+    CHECK(theArmnnDelegate != nullptr);
+
+    // Modify armnnDelegateInterpreter to use armnnDelegate
+    CHECK(armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk);
+
+    // Set input data
+    auto tfLiteDelegateInputId = tfLiteInterpreter->inputs()[0];
+    auto tfLiteDelegateInputData = tfLiteInterpreter->typed_tensor<T>(tfLiteDelegateInputId);
+    for (unsigned int i = 0; i < inputValues.size(); ++i)
+    {
+        tfLiteDelegateInputData[i] = inputValues[i];
+    }
+
+    auto armnnDelegateInputId = armnnDelegateInterpreter->inputs()[0];
+    auto armnnDelegateInputData = armnnDelegateInterpreter->typed_tensor<T>(armnnDelegateInputId);
+    for (unsigned int i = 0; i < inputValues.size(); ++i)
+    {
+        armnnDelegateInputData[i] = inputValues[i];
+    }
+
+    // Run EnqueueWorkload
+    CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
+
+    armnnDelegate::CompareOutputData(tfLiteInterpreter, armnnDelegateInterpreter, outputShape, expectedOutputValues);
+}
+
+// Function to create the flexbuffer custom options for the custom pooling3d operator.
+std::vector<uint8_t> CreateCustomOptions(int strideHeight, int strideWidth, int strideDepth,
+                                         int filterHeight, int filterWidth, int filterDepth, TfLitePadding padding)
+{
+    auto flex_builder = std::make_unique<flexbuffers::Builder>();
+    size_t map_start = flex_builder->StartMap();
+    flex_builder->String("data_format", "NDHWC");
+    // Padding is created as a key and padding type. Only VALID and SAME supported
+    if (padding == kTfLitePaddingValid)
+    {
+        flex_builder->String("padding", "VALID");
+    }
+    else
+    {
+        flex_builder->String("padding", "SAME");
+    }
+
+    // Vector of filter dimensions in order ( 1, Depth, Height, Width, 1 )
+    auto start = flex_builder->StartVector("ksize");
+    flex_builder->Add(1);
+    flex_builder->Add(filterDepth);
+    flex_builder->Add(filterHeight);
+    flex_builder->Add(filterWidth);
+    flex_builder->Add(1);
+    // EndVector( start, bool typed, bool fixed)
+    flex_builder->EndVector(start, true, false);
+
+    // Vector of stride dimensions in order ( 1, Depth, Height, Width, 1 )
+    auto stridesStart = flex_builder->StartVector("strides");
+    flex_builder->Add(1);
+    flex_builder->Add(strideDepth);
+    flex_builder->Add(strideHeight);
+    flex_builder->Add(strideWidth);
+    flex_builder->Add(1);
+    // EndVector( stridesStart, bool typed, bool fixed)
+    flex_builder->EndVector(stridesStart, true, false);
+
+    flex_builder->EndMap(map_start);
+    flex_builder->Finish();
+
+    return flex_builder->GetBuffer();
+}
+#endif
+} // anonymous namespace
+
+
+
+
diff --git a/docs/05_03_delegate.dox b/docs/05_03_delegate.dox
index 625b253992..d1c41fe213 100644
--- a/docs/05_03_delegate.dox
+++ b/docs/05_03_delegate.dox
@@ -43,6 +43,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 - AVERAGE_POOL_2D, Supported Fused Activation: RELU , RELU6 , TANH, NONE
 
+- AVERAGE_POOL_3D
+
 - BATCH_TO_SPACE_ND
 
 - CAST
@@ -107,6 +109,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 - MAX_POOL_2D, Supported Fused Activation: RELU , RELU6 , TANH, NONE
 
+- MAX_POOL_3D
+
 - MEAN
 
 - MINIMUM