IVGCVSW-5391 'ArmNN TfLiteDelegate: Implement the Space/Depth operators'

* Added DEPTH_TO_SPACE and SPACE_TO_DEPTH operators support

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I2595f759181bd7339127e7b114b850b534210dd5

5 files changed, 479 insertions, 14 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index 2862faf9e6..eda5b935e3 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -145,6 +145,8 @@ if(BUILD_UNIT_TESTS)
         src/test/ResizeTestHelper.hpp
         src/test/SoftmaxTest.cpp
         src/test/SoftmaxTestHelper.hpp
+        src/test/SpaceDepthTest.cpp
+        src/test/SpaceDepthTestHelper.hpp
         src/test/SplitTest.cpp
         src/test/SplitTestHelper.hpp
         src/test/TestUtils.hpp
diff --git a/delegate/TensorFlowLiteDelegateSupport.md b/delegate/TensorFlowLiteDelegateSupport.md
index ed1124a1ff..a5d4faf3ef 100644
--- a/delegate/TensorFlowLiteDelegateSupport.md
+++ b/delegate/TensorFlowLiteDelegateSupport.md
@@ -20,6 +20,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 * CONV_2D, Supported Fused Activation: RELU , RELU6 , TANH, NONE
 
+* DEPTH_TO_SPACE
+
 * DEPTHWISE_CONV_2D, Supported Fused Activation: RELU , RELU6 , TANH, NONE
 
 * DEQUANTIZE
@@ -94,6 +96,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 * SOFTMAX
 
+* SPACE_TO_DEPTH
+
 * SPLIT
 
 * SPLIT_V
diff --git a/delegate/src/SpaceDepth.hpp b/delegate/src/SpaceDepth.hpp
index 603e0f2fff..03859b6fcb 100644
--- a/delegate/src/SpaceDepth.hpp
+++ b/delegate/src/SpaceDepth.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>
@@ -21,13 +19,57 @@ TfLiteStatus VisitSpaceToDepthOperator(DelegateData& delegateData,
                                        int nodeIndex,
                                        int32_t operatorCode)
 {
-    armnn::IgnoreUnused(delegateData,
-                        tfLiteContext,
-                        tfLiteNode,
-                        nodeIndex,
-                        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::SpaceToDepthDescriptor descriptor;
+    auto* params = reinterpret_cast<TfLiteSpaceToDepthParams*>(tfLiteNode->builtin_data);
+    descriptor.m_BlockSize = params->block_size;
+
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported)
+    {
+        FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                                   tfLiteContext,
+                                   IsSpaceToDepthSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputTensorInfo,
+                                   outInfo,
+                                   descriptor);
+    };
+
+    if (!delegateData.m_Network)
+    {
+        validateFunc(outputTensorInfo, isSupported);
+        return isSupported ? kTfLiteOk : kTfLiteError;
+    }
+
+    // Add a SpaceToDepth layer
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddSpaceToDepthLayer(descriptor);
+    ARMNN_ASSERT(layer != nullptr);
+
+    armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
 
-    return kTfLiteError;
+    // Connect
+    return Connect(layer, tfLiteNode, delegateData);
 }
 
 TfLiteStatus VisitDepthToSpaceOperator(DelegateData& delegateData,
@@ -36,13 +78,57 @@ TfLiteStatus VisitDepthToSpaceOperator(DelegateData& delegateData,
                                        int nodeIndex,
                                        int32_t operatorCode)
 {
-    armnn::IgnoreUnused(delegateData,
-                        tfLiteContext,
-                        tfLiteNode,
-                        nodeIndex,
-                        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::DepthToSpaceDescriptor descriptor;
+    auto* params = reinterpret_cast<TfLiteDepthToSpaceParams*>(tfLiteNode->builtin_data);
+    descriptor.m_BlockSize = params->block_size;
+
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported)
+    {
+        FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                                   tfLiteContext,
+                                   IsDepthToSpaceSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputTensorInfo,
+                                   outInfo,
+                                   descriptor);
+    };
+
+    if (!delegateData.m_Network)
+    {
+        validateFunc(outputTensorInfo, isSupported);
+        return isSupported ? kTfLiteOk : kTfLiteError;
+    }
+
+    // Add a DepthToSpace layer
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddDepthToSpaceLayer(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/test/SpaceDepthTest.cpp b/delegate/src/test/SpaceDepthTest.cpp
new file mode 100644
index 0000000000..f80e749b87
--- /dev/null
+++ b/delegate/src/test/SpaceDepthTest.cpp
@@ -0,0 +1,207 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "SpaceDepthTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+void DepthToSpaceFp32Test(std::vector<armnn::BackendId>& backends, int blockSize)
+{
+    // Set input data
+    std::vector<int32_t> inputShape { 1, 2, 2, 4 };
+    std::vector<int32_t> outputShape { 1, 4, 4, 1 };
+
+    std::vector<float> inputValues = { 1.f,  2.f,  3.f,  4.f,
+                                       5.f,  6.f,  7.f,  8.f,
+                                       9.f, 10.f, 11.f, 12.f,
+                                       13.f, 14.f, 15.f, 16.f };
+
+    std::vector<float> expectedOutputValues = { 1.f,   2.f,   5.f,   6.f,
+                                                3.f,   4.f,   7.f,   8.f,
+                                                9.f,  10.f,  13.f,  14.f,
+                                                11.f,  12.f,  15.f,  16.f };
+
+    SpaceDepthTest<float>(tflite::BuiltinOperator_DEPTH_TO_SPACE,
+                          ::tflite::TensorType_FLOAT32,
+                          backends,
+                          inputShape,
+                          outputShape,
+                          inputValues,
+                          expectedOutputValues,
+                          blockSize);
+}
+
+void DepthToSpaceUint8Test(std::vector<armnn::BackendId>& backends, int blockSize)
+{
+    // Set input data
+    std::vector<int32_t> inputShape { 2, 1, 1, 4 };
+    std::vector<int32_t> outputShape { 2, 2, 2, 1 };
+
+    std::vector<uint8_t> inputValues = { 1,  2,  3,  4,
+                                         5,  6,  7,  8 };
+
+    std::vector<uint8_t> expectedOutputValues = { 1,  2,  3,  4,
+                                                  5,  6,  7,  8 };
+
+    SpaceDepthTest<uint8_t>(tflite::BuiltinOperator_DEPTH_TO_SPACE,
+                            ::tflite::TensorType_UINT8,
+                            backends,
+                            inputShape,
+                            outputShape,
+                            inputValues,
+                            expectedOutputValues,
+                            blockSize);
+}
+
+void SpaceToDepthFp32Test(std::vector<armnn::BackendId>& backends, int blockSize)
+{
+    // Set input data
+    std::vector<int32_t> inputShape { 1, 2, 2, 2 };
+    std::vector<int32_t> outputShape { 1, 1, 1, 8 };
+
+    std::vector<float> inputValues = { 1.4f, 2.3f, 3.2f, 4.1f, 5.4f, 6.3f, 7.2f, 8.1f };
+    std::vector<float> expectedOutputValues = { 1.4f, 2.3f, 3.2f, 4.1f, 5.4f, 6.3f, 7.2f, 8.1f };
+
+    SpaceDepthTest<float>(tflite::BuiltinOperator_SPACE_TO_DEPTH,
+                          ::tflite::TensorType_FLOAT32,
+                          backends,
+                          inputShape,
+                          outputShape,
+                          inputValues,
+                          expectedOutputValues,
+                          blockSize);
+}
+
+void SpaceToDepthUint8Test(std::vector<armnn::BackendId>& backends, int blockSize)
+{
+    // Set input data
+    std::vector<int32_t> inputShape { 1, 2, 2, 1 };
+    std::vector<int32_t> outputShape { 1, 1, 1, 4 };
+
+    std::vector<uint8_t> inputValues = { 1, 2, 3, 2 };
+    std::vector<uint8_t> expectedOutputValues = { 1, 2, 3, 2 };
+
+    SpaceDepthTest<uint8_t>(tflite::BuiltinOperator_SPACE_TO_DEPTH,
+                            ::tflite::TensorType_UINT8,
+                            backends,
+                            inputShape,
+                            outputShape,
+                            inputValues,
+                            expectedOutputValues,
+                            blockSize);
+}
+
+TEST_SUITE("DepthToSpace_CpuRefTests")
+{
+
+TEST_CASE ("DepthToSpaceFp32Test_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    DepthToSpaceFp32Test(backends, 2);
+}
+
+TEST_CASE ("DepthToSpaceUint8Test_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    DepthToSpaceUint8Test(backends, 2);
+}
+
+} // TEST_SUITE("DepthToSpace_CpuRefTests")
+
+
+TEST_SUITE("DepthToSpace_CpuAccTests")
+{
+
+TEST_CASE ("DepthToSpaceFp32Test_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    DepthToSpaceFp32Test(backends, 2);
+}
+
+TEST_CASE ("DepthToSpaceUint8Test_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    DepthToSpaceUint8Test(backends, 2);
+}
+
+} // TEST_SUITE("DepthToSpace_CpuAccTests")
+
+TEST_SUITE("DepthToSpace_GpuAccTests")
+{
+
+TEST_CASE ("DepthToSpaceFp32Test_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    DepthToSpaceFp32Test(backends, 2);
+}
+
+TEST_CASE ("DepthToSpaceUint8Test_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    DepthToSpaceUint8Test(backends, 2);
+}
+
+} // TEST_SUITE("DepthToSpace_GpuAccTests")
+
+TEST_SUITE("SpaceToDepth_CpuRefTests")
+{
+
+TEST_CASE ("SpaceToDepthFp32Test_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    SpaceToDepthFp32Test(backends, 2);
+}
+
+TEST_CASE ("SpaceToDepthUint8Test_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    SpaceToDepthUint8Test(backends, 2);
+}
+
+} // TEST_SUITE("SpaceToDepth_CpuRefTests")
+
+TEST_SUITE("SpaceToDepth_CpuAccTests")
+{
+
+TEST_CASE ("SpaceToDepthFp32Test_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    SpaceToDepthFp32Test(backends, 2);
+}
+
+TEST_CASE ("SpaceToDepthUint8Test_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    SpaceToDepthUint8Test(backends, 2);
+}
+
+} // TEST_SUITE("SpaceToDepth_CpuAccTests")
+
+TEST_SUITE("SpaceToDepth_GpuAccTests")
+{
+
+TEST_CASE ("SpaceToDepthFp32Test_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    SpaceToDepthFp32Test(backends, 2);
+}
+
+TEST_CASE ("SpaceToDepthUint8Test_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    SpaceToDepthUint8Test(backends, 2);
+}
+
+} // TEST_SUITE("SpaceToDepth_GpuAccTests")
+
+} // namespace armnnDelegate
diff --git a/delegate/src/test/SpaceDepthTestHelper.hpp b/delegate/src/test/SpaceDepthTestHelper.hpp
new file mode 100644
index 0000000000..d9a783c6a7
--- /dev/null
+++ b/delegate/src/test/SpaceDepthTestHelper.hpp
@@ -0,0 +1,166 @@
+//
+// 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> CreateSpaceDepthTfLiteModel(tflite::BuiltinOperator spaceDepthOperatorCode,
+                                              tflite::TensorType tensorType,
+                                              const std::vector <int32_t>& inputTensorShape,
+                                              const std::vector <int32_t>& outputTensorShape,
+                                              int32_t blockSize)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    auto quantizationParameters =
+        CreateQuantizationParameters(flatBufferBuilder,
+                                     0,
+                                     0,
+                                     flatBufferBuilder.CreateVector<float>({  1.0f }),
+                                     flatBufferBuilder.CreateVector<int64_t>({ 0 }));
+
+    std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
+    buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+
+    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);
+
+    const std::vector<int32_t> operatorInputs({0});
+    const std::vector<int32_t> operatorOutputs({1});
+
+    flatbuffers::Offset<Operator> spaceDepthOperator;
+    flatbuffers::Offset<flatbuffers::String> modelDescription;
+    flatbuffers::Offset<OperatorCode> operatorCode;
+
+    switch (spaceDepthOperatorCode)
+    {
+        case tflite::BuiltinOperator_SPACE_TO_DEPTH:
+            spaceDepthOperator =
+                CreateOperator(flatBufferBuilder,
+                               0,
+                               flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                               flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                               BuiltinOptions_SpaceToDepthOptions,
+                               CreateSpaceToDepthOptions(flatBufferBuilder, blockSize).Union());
+                modelDescription = flatBufferBuilder.CreateString("ArmnnDelegate: SPACE_TO_DEPTH Operator Model");
+                operatorCode = CreateOperatorCode(flatBufferBuilder,
+                                 tflite::BuiltinOperator_SPACE_TO_DEPTH);
+            break;
+        case tflite::BuiltinOperator_DEPTH_TO_SPACE:
+            spaceDepthOperator =
+                CreateOperator(flatBufferBuilder,
+                               0,
+                               flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                               flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                               BuiltinOptions_DepthToSpaceOptions,
+                               CreateDepthToSpaceOptions(flatBufferBuilder, blockSize).Union());
+                flatBufferBuilder.CreateString("ArmnnDelegate: DEPTH_TO_SPACE Operator Model");
+            operatorCode = CreateOperatorCode(flatBufferBuilder,
+                                              tflite::BuiltinOperator_DEPTH_TO_SPACE);
+            break;
+        default:
+            break;
+    }
+    const std::vector<int32_t> subgraphInputs({0});
+    const std::vector<int32_t> 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(&spaceDepthOperator, 1));
+    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 SpaceDepthTest(tflite::BuiltinOperator spaceDepthOperatorCode,
+                    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,
+                    int32_t blockSize = 2)
+{
+    using namespace tflite;
+    std::vector<char> modelBuffer = CreateSpaceDepthTfLiteModel(spaceDepthOperatorCode,
+                                                                tensorType,
+                                                                inputShape,
+                                                                outputShape,
+                                                                blockSize);
+
+    const Model* tfLiteModel = GetModel(modelBuffer.data());
+    // Create TfLite Interpreters
+    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 EnqueWorkload
+    CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
+
+    // Compare output data
+    armnnDelegate::CompareOutputData(tfLiteInterpreter, armnnDelegateInterpreter, outputShape, expectedOutputValues);
+}
+
+} // anonymous namespace