IVGCVSW-5390 'TfLiteDelegate: Implement the Normalisation operators'

* Added L2_NORMALIZATION and LOCAL_RESPONSE_NORMALIZATION opertor support

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: Ic9e66879cf6469fa8761fb1c9dd5950771f629b0

6 files changed, 487 insertions, 23 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index ba8ba6de00..2862faf9e6 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -131,6 +131,8 @@ if(BUILD_UNIT_TESTS)
         src/test/GatherTestHelper.hpp
         src/test/LogicalTest.cpp
         src/test/LogicalTestHelper.hpp
+        src/test/NormalizationTest.cpp
+        src/test/NormalizationTestHelper.hpp
         src/test/PadTest.cpp
         src/test/PadTestHelper.hpp
         src/test/Pooling2dTest.cpp
diff --git a/delegate/TensorFlowLiteDelegateSupport.md b/delegate/TensorFlowLiteDelegateSupport.md
index dd4cbace7b..ed1124a1ff 100644
--- a/delegate/TensorFlowLiteDelegateSupport.md
+++ b/delegate/TensorFlowLiteDelegateSupport.md
@@ -46,6 +46,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 * LESS_OR_EQUAL
 
+* LOCAL_RESPONSE_NORMALIZATION
+
 * LOGICAL_AND
   
 * LOGICAL_NOT
@@ -56,6 +58,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 * LOG_SOFTMAX
 
+* L2_NORMALIZATION
+
 * L2_POOL_2D
 
 * MAXIMUM
diff --git a/delegate/src/Normalization.hpp b/delegate/src/Normalization.hpp
index 4c18b364cc..68ff3af32d 100644
--- a/delegate/src/Normalization.hpp
+++ b/delegate/src/Normalization.hpp
@@ -5,8 +5,6 @@
 
 #pragma once
 
-#include <armnn/utility/IgnoreUnused.hpp>
-
 #include <tensorflow/lite/builtin_ops.h>
 #include <tensorflow/lite/c/builtin_op_data.h>
 #include <tensorflow/lite/c/common.h>
@@ -15,19 +13,132 @@
 namespace armnnDelegate
 {
 
-TfLiteStatus VisitNormalizationOperator(DelegateData& delegateData,
-                                        TfLiteContext* tfLiteContext,
-                                        TfLiteNode* tfLiteNode,
-                                        int nodeIndex,
-                                        int32_t normalizationOperatorCode)
+TfLiteStatus VisitL2NormalizationOperator(DelegateData& delegateData,
+                                          TfLiteContext* tfLiteContext,
+                                          TfLiteNode* tfLiteNode,
+                                          int nodeIndex,
+                                          int32_t operatorCode)
+{
+    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 (!IsValid(tfLiteContext, tfLiteInputTensor, operatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
+    if (!IsValid(tfLiteContext, tfLiteOutputTensor, operatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const armnn::TensorInfo& inputTensorInfo  = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
+
+    armnn::L2NormalizationDescriptor descriptor;
+    descriptor.m_DataLayout = armnn::DataLayout::NHWC;
+
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported)
+    {
+        FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                                   tfLiteContext,
+                                   IsL2NormalizationSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputTensorInfo,
+                                   outInfo,
+                                   descriptor);
+    };
+
+    if (!delegateData.m_Network)
+    {
+        validateFunc(outputTensorInfo, isSupported);
+        return isSupported ? kTfLiteOk : kTfLiteError;
+    }
+
+    // Add a L2Normalization layer
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddL2NormalizationLayer(descriptor);
+    ARMNN_ASSERT(layer != nullptr);
+
+    armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
+
+    // Connect
+    return Connect(layer, tfLiteNode, delegateData);
+}
+
+
+TfLiteStatus VisitLocalResponseNormalizationOperator(DelegateData& delegateData,
+                                                     TfLiteContext* tfLiteContext,
+                                                     TfLiteNode* tfLiteNode,
+                                                     int nodeIndex,
+                                                     int32_t normalizationOperatorCode)
 {
-    armnn::IgnoreUnused(delegateData,
-                        tfLiteContext,
-                        tfLiteNode,
-                        nodeIndex,
-                        normalizationOperatorCode);
+    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 (!IsValid(tfLiteContext, tfLiteInputTensor, normalizationOperatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
+    if (!IsValid(tfLiteContext, tfLiteOutputTensor, normalizationOperatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const armnn::TensorInfo& inputTensorInfo  = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
+
+    armnn::NormalizationDescriptor descriptor;
+    descriptor.m_DataLayout = armnn::DataLayout::NHWC;
+    descriptor.m_NormChannelType = armnn::NormalizationAlgorithmChannel::Across;
+    descriptor.m_NormMethodType  = armnn::NormalizationAlgorithmMethod::LocalBrightness;
+
+    auto* params = reinterpret_cast<TfLiteLocalResponseNormParams*>(tfLiteNode->builtin_data);
+    descriptor.m_NormSize = params->radius;
+    descriptor.m_K        = params->bias;
+    descriptor.m_Alpha    = params->alpha;
+    descriptor.m_Beta     = params->beta;
+
+    // ArmNN expects normSize to be the full size of the normalization window
+    descriptor.m_NormSize = 1 + (2 * descriptor.m_NormSize);
+
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported)
+    {
+        FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                                   tfLiteContext,
+                                   IsNormalizationSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputTensorInfo,
+                                   outInfo,
+                                   descriptor);
+    };
+
+    if (!delegateData.m_Network)
+    {
+        validateFunc(outputTensorInfo, isSupported);
+        return isSupported ? kTfLiteOk : kTfLiteError;
+    }
+
+    // Add a Normalization layer
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddNormalizationLayer(descriptor);
+    ARMNN_ASSERT(layer != nullptr);
+
+    armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
 
-    return kTfLiteError;
+    // Connect
+    return Connect(layer, tfLiteNode, delegateData);
 }
 
 } // namespace armnnDelegate
diff --git a/delegate/src/armnn_delegate.cpp b/delegate/src/armnn_delegate.cpp
index 6dba890509..6f72d864b9 100644
--- a/delegate/src/armnn_delegate.cpp
+++ b/delegate/src/armnn_delegate.cpp
@@ -575,11 +575,11 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                            nodeIndex,
                                            kTfLiteBuiltinHardSwish);
         case kTfLiteBuiltinL2Normalization:
-            return VisitNormalizationOperator(delegateData,
-                                              tfLiteContext,
-                                              tfLiteNode,
-                                              nodeIndex,
-                                              kTfLiteBuiltinL2Normalization);
+            return VisitL2NormalizationOperator(delegateData,
+                                                tfLiteContext,
+                                                tfLiteNode,
+                                                nodeIndex,
+                                                kTfLiteBuiltinL2Normalization);
         case kTfLiteBuiltinL2Pool2d:
             return VisitPoolingOperator(delegateData,
                                         tfLiteContext,
@@ -599,11 +599,11 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                            nodeIndex,
                                            kTfLiteBuiltinLessEqual);
         case kTfLiteBuiltinLocalResponseNormalization:
-            return VisitNormalizationOperator(delegateData,
-                                              tfLiteContext,
-                                              tfLiteNode,
-                                              nodeIndex,
-                                              kTfLiteBuiltinLocalResponseNormalization);
+            return VisitLocalResponseNormalizationOperator(delegateData,
+                                                           tfLiteContext,
+                                                           tfLiteNode,
+                                                           nodeIndex,
+                                                           kTfLiteBuiltinLocalResponseNormalization);
         case kTfLiteBuiltinLogicalAnd:
             return VisitLogicalBinaryOperator(delegateData,
                                               tfLiteContext,
diff --git a/delegate/src/test/NormalizationTest.cpp b/delegate/src/test/NormalizationTest.cpp
new file mode 100644
index 0000000000..058394edb7
--- /dev/null
+++ b/delegate/src/test/NormalizationTest.cpp
@@ -0,0 +1,166 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NormalizationTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+void L2NormalizationTest(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data
+    std::vector<int32_t> inputShape  { 1, 1, 1, 10 };
+    std::vector<int32_t> outputShape { 1, 1, 1, 10 };
+
+    std::vector<float> inputValues
+    {
+        1.0f,
+        2.0f,
+        3.0f,
+        4.0f,
+        5.0f,
+        6.0f,
+        7.0f,
+        8.0f,
+        9.0f,
+        10.0f
+    };
+
+    const float approxInvL2Norm = 0.050964719f;
+    std::vector<float> expectedOutputValues
+    {
+        1.0f  * approxInvL2Norm,
+        2.0f  * approxInvL2Norm,
+        3.0f  * approxInvL2Norm,
+        4.0f  * approxInvL2Norm,
+        5.0f  * approxInvL2Norm,
+        6.0f  * approxInvL2Norm,
+        7.0f  * approxInvL2Norm,
+        8.0f  * approxInvL2Norm,
+        9.0f  * approxInvL2Norm,
+        10.0f * approxInvL2Norm
+    };
+
+    NormalizationTest<float>(tflite::BuiltinOperator_L2_NORMALIZATION,
+                             ::tflite::TensorType_FLOAT32,
+                             backends,
+                             inputShape,
+                             outputShape,
+                             inputValues,
+                             expectedOutputValues);
+}
+
+void LocalResponseNormalizationTest(std::vector<armnn::BackendId>& backends,
+                                    int32_t radius,
+                                    float bias,
+                                    float alpha,
+                                    float beta)
+{
+    // Set input data
+    std::vector<int32_t> inputShape  { 2, 2, 2, 1 };
+    std::vector<int32_t> outputShape { 2, 2, 2, 1 };
+
+    std::vector<float> inputValues
+    {
+        1.0f, 2.0f,
+        3.0f, 4.0f,
+        5.0f, 6.0f,
+        7.0f, 8.0f
+    };
+
+    std::vector<float> expectedOutputValues
+    {
+        0.5f, 0.400000006f, 0.300000012f, 0.235294119f,
+        0.192307696f, 0.16216217f, 0.140000001f, 0.123076923f
+    };
+
+    NormalizationTest<float>(tflite::BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION,
+                             ::tflite::TensorType_FLOAT32,
+                             backends,
+                             inputShape,
+                             outputShape,
+                             inputValues,
+                             expectedOutputValues,
+                             radius,
+                             bias,
+                             alpha,
+                             beta);
+}
+
+
+TEST_SUITE("L2Normalization_CpuRefTests")
+{
+
+TEST_CASE ("L2NormalizationFp32Test_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    L2NormalizationTest(backends);
+}
+
+} // TEST_SUITE("L2Normalization_CpuRefTests")
+
+TEST_SUITE("L2Normalization_CpuAccTests")
+{
+
+TEST_CASE ("L2NormalizationFp32Test_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    L2NormalizationTest(backends);
+}
+
+} // TEST_SUITE("L2NormalizationFp32Test_CpuAcc_Test")
+
+TEST_SUITE("L2Normalization_GpuAccTests")
+{
+
+TEST_CASE ("L2NormalizationFp32Test_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    L2NormalizationTest(backends);
+}
+
+} // TEST_SUITE("L2Normalization_GpuAccTests")
+
+TEST_SUITE("LocalResponseNormalization_CpuRefTests")
+{
+
+TEST_CASE ("LocalResponseNormalizationTest_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    LocalResponseNormalizationTest(backends, 3, 1.f, 1.f, 1.f);
+}
+
+} // TEST_SUITE("LocalResponseNormalization_CpuRefTests")
+
+TEST_SUITE("LocalResponseNormalization_CpuAccTests")
+{
+
+TEST_CASE ("LocalResponseNormalizationTest_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    LocalResponseNormalizationTest(backends, 3, 1.f, 1.f, 1.f);
+}
+
+} // TEST_SUITE("LocalResponseNormalization_CpuAccTests")
+
+TEST_SUITE("LocalResponseNormalization_GpuAccTests")
+{
+
+TEST_CASE ("LocalResponseNormalizationTest_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    LocalResponseNormalizationTest(backends, 3, 1.f, 1.f, 1.f);
+}
+
+} // TEST_SUITE("LocalResponseNormalization_GpuAccTests")
+
+} // namespace armnnDelegate
\ No newline at end of file
diff --git a/delegate/src/test/NormalizationTestHelper.hpp b/delegate/src/test/NormalizationTestHelper.hpp
new file mode 100644
index 0000000000..26286b1c88
--- /dev/null
+++ b/delegate/src/test/NormalizationTestHelper.hpp
@@ -0,0 +1,181 @@
+//
+// Copyright © 2021 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 <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
+{
+
+std::vector<char> CreateNormalizationTfLiteModel(tflite::BuiltinOperator normalizationOperatorCode,
+                                                 tflite::TensorType tensorType,
+                                                 const std::vector<int32_t>& inputTensorShape,
+                                                 const std::vector<int32_t>& outputTensorShape,
+                                                 int32_t radius,
+                                                 float bias,
+                                                 float alpha,
+                                                 float beta,
+                                                 float quantScale = 1.0f,
+                                                 int quantOffset  = 0)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    auto quantizationParameters =
+        CreateQuantizationParameters(flatBufferBuilder,
+                                     0,
+                                     0,
+                                     flatBufferBuilder.CreateVector<float>({ quantScale }),
+                                     flatBufferBuilder.CreateVector<int64_t>({ quantOffset }));
+
+    auto inputTensor = CreateTensor(flatBufferBuilder,
+                                    flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(),
+                                                                            inputTensorShape.size()),
+                                    tensorType,
+                                    0,
+                                    flatBufferBuilder.CreateString("input"),
+                                    quantizationParameters);
+
+    auto outputTensor = CreateTensor(flatBufferBuilder,
+                                     flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(),
+                                                                             outputTensorShape.size()),
+                                     tensorType,
+                                     1,
+                                     flatBufferBuilder.CreateString("output"),
+                                     quantizationParameters);
+
+    std::vector<flatbuffers::Offset<Tensor>> tensors = { inputTensor, outputTensor };
+
+    std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
+    buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+    buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+
+    std::vector<int32_t> operatorInputs = {{ 0 }};
+    std::vector<int> subgraphInputs = {{ 0 }};
+
+    tflite::BuiltinOptions operatorBuiltinOptionsType = BuiltinOptions_L2NormOptions;
+    flatbuffers::Offset<void> operatorBuiltinOptions = CreateL2NormOptions(flatBufferBuilder,
+                                                                           tflite::ActivationFunctionType_NONE).Union();
+
+    if (normalizationOperatorCode == tflite::BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION)
+    {
+        operatorBuiltinOptionsType = BuiltinOptions_LocalResponseNormalizationOptions;
+        operatorBuiltinOptions =
+            CreateLocalResponseNormalizationOptions(flatBufferBuilder, radius, bias, alpha, beta).Union();
+    }
+
+    // create operator
+    const std::vector<int32_t> operatorOutputs{{ 1 }};
+    flatbuffers::Offset <Operator> normalizationOperator =
+        CreateOperator(flatBufferBuilder,
+                       0,
+                       flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                       operatorBuiltinOptionsType,
+                       operatorBuiltinOptions);
+
+    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(&normalizationOperator, 1));
+
+    flatbuffers::Offset <flatbuffers::String> modelDescription =
+        flatBufferBuilder.CreateString("ArmnnDelegate: Normalization Operator Model");
+    flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder,
+                                                                         normalizationOperatorCode);
+
+    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 NormalizationTest(tflite::BuiltinOperator normalizationOperatorCode,
+                       tflite::TensorType tensorType,
+                       const std::vector<armnn::BackendId>& backends,
+                       const std::vector<int32_t>& inputShape,
+                       std::vector<int32_t>& outputShape,
+                       std::vector<T>& inputValues,
+                       std::vector<T>& expectedOutputValues,
+                       int32_t radius = 0,
+                       float bias = 0.f,
+                       float alpha = 0.f,
+                       float beta = 0.f,
+                       float quantScale = 1.0f,
+                       int quantOffset  = 0)
+{
+    using namespace tflite;
+    std::vector<char> modelBuffer = CreateNormalizationTfLiteModel(normalizationOperatorCode,
+                                                                   tensorType,
+                                                                   inputShape,
+                                                                   outputShape,
+                                                                   radius,
+                                                                   bias,
+                                                                   alpha,
+                                                                   beta,
+                                                                   quantScale,
+                                                                   quantOffset);
+
+    const Model* tfLiteModel = GetModel(modelBuffer.data());
+    CHECK(tfLiteModel != nullptr);
+
+    std::unique_ptr<Interpreter> armnnDelegateInterpreter;
+    CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
+          (&armnnDelegateInterpreter) == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter != nullptr);
+    CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk);
+
+    std::unique_ptr<Interpreter> tfLiteInterpreter;
+    CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
+          (&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
+    armnnDelegate::FillInput<T>(tfLiteInterpreter, 0, inputValues);
+    armnnDelegate::FillInput<T>(armnnDelegateInterpreter, 0, inputValues);
+
+    // Run EnqueueWorkload
+    CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
+
+    // Compare output data
+    armnnDelegate::CompareOutputData(tfLiteInterpreter, armnnDelegateInterpreter, outputShape, expectedOutputValues);
+}
+
+} // anonymous namespace
\ No newline at end of file