IVGCVSW-7886 Add TILE to delegate and opaque delegate

  * Adding support for Tile in classic and opaque delegates
  * CMake files updated
  * Tests added

Signed-off-by: Tianle Cheng <tianle.cheng@arm.com>
Change-Id: I9b52cea3480eb71961cbccb1a346805f73b5661a

10 files changed, 621 insertions, 4 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index e46ac04092..c1bf73a6ab 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -201,6 +201,8 @@ if(BUILD_UNIT_TESTS)
         test/SplitTestHelper.hpp
         test/TestUtils.hpp
         test/TestUtils.cpp
+        test/TileTest.cpp
+        test/TileTestHelper.hpp
         test/TransposeConvolution2dTest.cpp
         test/TransposeTest.cpp
         test/TransposeTestHelper.hpp
diff --git a/delegate/classic/CMakeLists.txt b/delegate/classic/CMakeLists.txt
index 8f872d6adc..7807153359 100644
--- a/delegate/classic/CMakeLists.txt
+++ b/delegate/classic/CMakeLists.txt
@@ -44,6 +44,7 @@ list(APPEND armnnClassicDelegateObject_sources
         src/SpaceDepth.hpp
         src/Split.hpp
         src/Unpack.hpp
+        src/Tile.hpp
         src/Transpose.hpp)
 
 add_library(armnnClassicDelegateObject OBJECT ${armnnClassicDelegateObject_sources})
diff --git a/delegate/classic/src/Tile.hpp b/delegate/classic/src/Tile.hpp
new file mode 100644
index 0000000000..974c771a7e
--- /dev/null
+++ b/delegate/classic/src/Tile.hpp
@@ -0,0 +1,169 @@
+//
+// Copyright © 2023 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#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>
+#include <tensorflow/lite/minimal_logging.h>
+#include <tensorflow/lite/kernels/internal/tensor_ctypes.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+
+namespace armnnDelegate
+{
+TfLiteStatus ValidateTileOperator(DelegateData& delegateData,
+                                  TfLiteContext* tfLiteContext,
+                                  const armnn::TensorInfo& inputInfo,
+                                  const armnn::TensorInfo& outputInfo,
+                                  const armnn::TileDescriptor& descriptor)
+{
+    bool isSupported = false;
+    FORWARD_LAYER_SUPPORT_FUNC("TILE",
+                               tfLiteContext,
+                               IsTileSupported,
+                               delegateData.m_Backends,
+                               isSupported,
+                               armnn::BackendId(),
+                               inputInfo,
+                               outputInfo,
+                               descriptor);
+    return isSupported ? kTfLiteOk : kTfLiteError;
+}
+
+TfLiteStatus VisitTileOperator(DelegateData& delegateData,
+                               TfLiteContext* tfLiteContext,
+                               TfLiteNode* tfLiteNode,
+                               int nodeIndex,
+                               int32_t tileOperatorCode)
+{
+    TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 2, nodeIndex));
+    TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
+
+    const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors;
+
+    // The input contains the data that should be tiled
+    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: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // The multiples tensor contains the number of copies for each axis
+    const TfLiteTensor& tfLiteMultiplesTensor = tfLiteTensors[tfLiteNode->inputs->data[1]];
+    if (IsDynamicTensor(tfLiteMultiplesTensor))
+    {
+        TF_LITE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // The output tensor
+    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: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& multiplesTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteMultiplesTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
+
+    // Multiples length must be the same as the number of dimension in input tensor
+    if (multiplesTensorInfo.GetNumElements() != inputTensorInfo.GetNumDimensions())
+    {
+        TF_LITE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnDelegate: The Multiples length must be the same as the number of dimension in input tensor",
+            "Operator: #%d node #%d: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // Get the Multiples data: In armnn, the values of the multiples input tensor is saved in the operator descriptor
+    // We have to read it from the input tensor and write it the descriptor
+    auto* multiplesTensorDataPtr = tflite::GetTensorData<int32_t>(&tfLiteMultiplesTensor);
+    auto multiplesTensorNum = tfLiteMultiplesTensor.dims->data[0];
+    std::vector<int32_t> multiplesIntData(multiplesTensorDataPtr, multiplesTensorDataPtr + multiplesTensorNum);
+
+    // The multiples must be positive
+    for (auto multiple : multiplesIntData)
+    {
+        if (multiple < 0)
+        {
+            TF_LITE_MAYBE_KERNEL_LOG(
+                tfLiteContext,
+                "TfLiteArmnnDelegate: The Multiples must be positive values",
+                "Operator: #%d node #%d: ",
+                tileOperatorCode, nodeIndex);
+            return kTfLiteError;
+        }
+    }
+
+    // The original input from TFLite is int32, and we have to make it as uint32 for our descriptor
+    std::vector<uint32_t> multiplesUintData;
+    std::transform(multiplesIntData.begin(),
+                   multiplesIntData.end(),
+                   std::back_inserter(multiplesUintData),
+                   [] (const int value)
+                   {
+                        return static_cast<uint32_t>(value);
+                   });
+
+    armnn::TileDescriptor tileDescriptor;
+    tileDescriptor.m_Multiples = multiplesUintData;
+
+    // Check output dimensions
+    if (inputTensorInfo.GetNumDimensions() != outputTensorInfo.GetNumDimensions())
+    {
+        TF_LITE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnDelegate: Input tensor dimension and output tensor dimension differ",
+            "Operator: #%d node #%d: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // No network pointer indicates that only support for this operator should be checked
+    if (!delegateData.m_Network)
+    {
+        return ValidateTileOperator(delegateData,
+                                    tfLiteContext,
+                                    inputTensorInfo,
+                                    outputTensorInfo,
+                                    tileDescriptor);
+    }
+
+    std::string layerName("Tile");
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddTileLayer(tileDescriptor, layerName.c_str());
+
+    if (layer == nullptr)
+    {
+        return kTfLiteError;
+    }
+
+    layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+
+    if (ProcessInputs(layer, delegateData, tfLiteContext, tfLiteNode) != kTfLiteOk)
+    {
+        return kTfLiteError;
+    }
+
+    return Connect(layer, tfLiteNode, delegateData);
+}
+
+} // namespace armnnDelegate
\ No newline at end of file
diff --git a/delegate/classic/src/armnn_delegate.cpp b/delegate/classic/src/armnn_delegate.cpp
index 0f9e8a624c..45bea3d442 100644
--- a/delegate/classic/src/armnn_delegate.cpp
+++ b/delegate/classic/src/armnn_delegate.cpp
@@ -39,6 +39,7 @@
 #include "Softmax.hpp"
 #include "SpaceDepth.hpp"
 #include "Split.hpp"
+#include "Tile.hpp"
 #include "Transpose.hpp"
 #include "UnidirectionalSequenceLstm.hpp"
 #include "Unpack.hpp"
@@ -1064,6 +1065,12 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                            tfLiteNode,
                                            nodeIndex,
                                            kTfLiteBuiltinTanh);
+        case kTfLiteBuiltinTile:
+            return VisitTileOperator(delegateData,
+                                     tfLiteContext,
+                                     tfLiteNode,
+                                     nodeIndex,
+                                     kTfLiteBuiltinTile);
         case kTfLiteBuiltinUnidirectionalSequenceLstm:
             return VisitUnidirectionalSequenceLstmOperator(delegateData,
                                                            tfLiteContext,
diff --git a/delegate/opaque/CMakeLists.txt b/delegate/opaque/CMakeLists.txt
index 787046d80c..c05bccf8c9 100644
--- a/delegate/opaque/CMakeLists.txt
+++ b/delegate/opaque/CMakeLists.txt
@@ -41,6 +41,7 @@ list(APPEND armnnOpaqueDelegateObject_sources
         src/SpaceDepth.hpp
         src/Split.hpp
         src/StridedSlice.hpp
+        src/Tile.hpp
         src/Transpose.hpp
         src/UnidirectionalSequenceLstm.hpp
         src/Unpack.hpp)
diff --git a/delegate/opaque/src/Tile.hpp b/delegate/opaque/src/Tile.hpp
new file mode 100644
index 0000000000..17cbdee7eb
--- /dev/null
+++ b/delegate/opaque/src/Tile.hpp
@@ -0,0 +1,188 @@
+//
+// Copyright © 2023 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <OpaqueDelegateUtils.hpp>
+
+namespace armnnOpaqueDelegate
+{
+TfLiteStatus ValidateTileOperator(DelegateData& delegateData,
+                                  TfLiteOpaqueContext *tfLiteContext,
+                                  const armnn::TensorInfo& inputInfo,
+                                  const armnn::TensorInfo& outputInfo,
+                                  const armnn::TileDescriptor& descriptor)
+{
+    bool isSupported = false;
+    FORWARD_LAYER_OPAQUE_SUPPORT_FUNC("TILE",
+                                      tfLiteContext,
+                                      IsTileSupported,
+                                      delegateData.m_Backends,
+                                      isSupported,
+                                      armnn::BackendId(),
+                                      inputInfo,
+                                      outputInfo,
+                                      descriptor);
+    return isSupported ? kTfLiteOk : kTfLiteError;
+}
+
+TfLiteStatus VisitTileOperator(DelegateData& delegateData,
+                               TfLiteOpaqueContext* tfLiteContext,
+                               TfLiteOpaqueNode* tfLiteNode,
+                               int nodeIndex,
+                               int32_t tileOperatorCode)
+{
+    TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 2, nodeIndex));
+    TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
+
+    // Gather input tensors
+    auto numInputs = TfLiteOpaqueNodeNumberOfInputs(tfLiteNode);
+    const int* inputTensors;
+    if (TfLiteOpaqueNodeInputs(tfLiteNode, &inputTensors, &numInputs) != kTfLiteOk)
+    {
+        TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnOpaqueDelegate: Unable to gather input tensor indices from node #%d: ",
+            nodeIndex);
+        return kTfLiteError;
+    }
+
+    // Gather output tensors
+    int numOutputs = 0;
+    const int* outputTensors;
+    if (TfLiteOpaqueNodeOutputs(tfLiteNode, &outputTensors, &numOutputs) != kTfLiteOk)
+    {
+        TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnOpaqueDelegate: Unable to gather output tensor indices from node #%d: ",
+            nodeIndex);
+        return kTfLiteError;
+    }
+
+    // The input contains the data that should be tiled
+    const TfLiteOpaqueTensor* tfLiteInputTensor =
+            TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[0]);
+    if (IsDynamicTensor(tfLiteInputTensor))
+    {
+        TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnOpaqueDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // The multiples tensor contains the number of copies for each axis
+    const TfLiteOpaqueTensor* tfLiteMultiplesTensor =
+            TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[1]);;
+    if (IsDynamicTensor(tfLiteMultiplesTensor))
+    {
+        TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnOpaqueDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // The output tensor
+    const TfLiteOpaqueTensor* tfLiteOutputTensor =
+            TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, outputTensors[0]);
+    if (IsDynamicTensor(tfLiteOutputTensor))
+    {
+        TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnOpaqueDelegate: Dynamic output tensors are not supported in operator #%d node #%d: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& multiplesTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(tfLiteMultiplesTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(tfLiteOutputTensor, true);
+
+    // Multiples length must be the same as the number of dimension in input tensor
+    if (multiplesTensorInfo.GetNumElements() != inputTensorInfo.GetNumDimensions())
+    {
+        TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnOpaqueDelegate:",
+            "The Multiples length must be the same as the number of dimension in input tensor",
+            "Operator: #%d node #%d: ",
+            tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // Get the Multiples data: In armnn, the values of the multiples input tensor is saved in the operator descriptor
+    // We have to read it from the input tensor and write it the descriptor
+    auto* multiplesTensorDataPtr = static_cast<int32_t*>(TfLiteOpaqueTensorData(tfLiteMultiplesTensor));
+    auto multiplesTensorNum = TfLiteOpaqueTensorDim(tfLiteMultiplesTensor, 0);
+    std::vector<int32_t> multiplesIntData(multiplesTensorDataPtr, multiplesTensorDataPtr + multiplesTensorNum);
+
+    // The multiples must be positive
+    for (auto multiple : multiplesIntData)
+    {
+        if (multiple < 0)
+        {
+            TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+                tfLiteContext,
+                "TfLiteArmnnOpaqueDelegate: The Multiples must be positive values",
+                "Operator: #%d node #%d: ",
+                tileOperatorCode, nodeIndex);
+            return kTfLiteError;
+        }
+    }
+
+    // The original input from TFLite is int32, and we have to make it as uint32 for our descriptor
+    std::vector<uint32_t> multiplesUintData;
+    std::transform(multiplesIntData.begin(),
+                   multiplesIntData.end(),
+                   std::back_inserter(multiplesUintData),
+                   [] (const int value)
+                   {
+                       return static_cast<uint32_t>(value);
+                   });
+
+    armnn::TileDescriptor tileDescriptor;
+    tileDescriptor.m_Multiples = multiplesUintData;
+
+    // Check output dimensions
+    if (inputTensorInfo.GetNumDimensions() != outputTensorInfo.GetNumDimensions())
+    {
+        TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
+                tfLiteContext,
+                "TfLiteArmnnOpaqueDelegate: Input tensor dimension and output tensor dimension differ",
+                "Operator: #%d node #%d: ",
+                tileOperatorCode, nodeIndex);
+        return kTfLiteError;
+    }
+
+    // No network pointer indicates that only support for this operator should be checked
+    if (!delegateData.m_Network)
+    {
+        return ValidateTileOperator(delegateData,
+                                    tfLiteContext,
+                                    inputTensorInfo,
+                                    outputTensorInfo,
+                                    tileDescriptor);
+    }
+
+    std::string layerName("Tile");
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddTileLayer(tileDescriptor, layerName.c_str());
+
+    if (layer == nullptr)
+    {
+        return kTfLiteError;
+    }
+
+    layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+
+    if (ProcessInputs(layer, delegateData, tfLiteContext, tfLiteNode) != kTfLiteOk)
+    {
+        return kTfLiteError;
+    }
+
+    return Connect(layer, tfLiteContext, tfLiteNode, delegateData);
+}
+
+} // namespace armnnOpaqueDelegate
\ No newline at end of file
diff --git a/delegate/opaque/src/armnn_delegate.cpp b/delegate/opaque/src/armnn_delegate.cpp
index 510352eae9..49fa30d8f0 100644
--- a/delegate/opaque/src/armnn_delegate.cpp
+++ b/delegate/opaque/src/armnn_delegate.cpp
@@ -38,6 +38,7 @@
 #include "Softmax.hpp"
 #include "SpaceDepth.hpp"
 #include "Split.hpp"
+#include "Tile.hpp"
 #include "Transpose.hpp"
 #include "UnidirectionalSequenceLstm.hpp"
 #include "Unpack.hpp"
@@ -1138,12 +1139,18 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                            tfLiteNode,
                                            nodeIndex,
                                            kTfLiteBuiltinTanh);
+        case kTfLiteBuiltinTile:
+            return VisitTileOperator(delegateData,
+                                     tfLiteContext,
+                                     tfLiteNode,
+                                     nodeIndex,
+                                     kTfLiteBuiltinTile);
         case kTfLiteBuiltinTranspose:
             return VisitTransposeOperator(delegateData,
-                                            tfLiteContext,
-                                            tfLiteNode,
-                                            nodeIndex,
-                                            kTfLiteBuiltinTranspose);
+                                          tfLiteContext,
+                                          tfLiteNode,
+                                          nodeIndex,
+                                          kTfLiteBuiltinTranspose);
         case kTfLiteBuiltinTransposeConv:
             return VisitConvolutionOperator(delegateData,
                                             tfLiteContext,
diff --git a/delegate/test/TileTest.cpp b/delegate/test/TileTest.cpp
new file mode 100644
index 0000000000..2e20859f5b
--- /dev/null
+++ b/delegate/test/TileTest.cpp
@@ -0,0 +1,91 @@
+//
+// Copyright © 2023 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "TileTestHelper.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 <schema_generated.h>
+#include <tensorflow/lite/version.h>
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+void TileFloat32Test(std::vector<armnn::BackendId>& backends)
+{
+    // Set input data
+    std::vector<float> inputValues =
+    {
+        0.f, 1.f, 2.f,
+        3.f, 4.f, 5.f
+    };
+
+    // Set output data
+    std::vector<float> expectedOutputValues =
+    {
+        0.f, 1.f, 2.f, 0.f, 1.f, 2.f,
+        3.f, 4.f, 5.f, 3.f, 4.f, 5.f,
+
+        0.f, 1.f, 2.f, 0.f, 1.f, 2.f,
+        3.f, 4.f, 5.f, 3.f, 4.f, 5.f
+    };
+
+    // The multiples
+    const std::vector<int32_t> multiplesValues = { 2, 2 };
+
+    // Set shapes
+    const std::vector<int32_t> inputShape = { 2, 3 };
+    const std::vector<int32_t> multiplesShape = { 2 };
+    const std::vector<int32_t> expectedOutputShape = { 4, 6 };
+
+    TileFP32TestImpl(tflite::BuiltinOperator_TILE,
+                     backends,
+                     inputValues,
+                     inputShape,
+                     multiplesValues,
+                     multiplesShape,
+                     expectedOutputValues,
+                     expectedOutputShape);
+}
+
+#if defined(TILE_GPUACC)
+TEST_SUITE("TileTests_GpuAccTests")
+{
+
+    TEST_CASE ("Tile_Float32_GpuAcc_Test")
+    {
+        std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+        TileFloat32Test(backends);
+    }
+
+} // TEST_SUITE("Tile_Float32_GpuAcc_Test")
+#endif
+
+TEST_SUITE("TileTests_CpuAccTests")
+{
+
+    TEST_CASE ("Tile_Float32_CpuAcc_Test")
+    {
+        std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+        TileFloat32Test(backends);
+    }
+
+} // TEST_SUITE("Tile_Float32_CpuAcc_Test")
+
+TEST_SUITE("TileTests_CpuRefTests")
+{
+
+    TEST_CASE ("Tile_Float32_CpuRef_Test")
+    {
+        std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+        TileFloat32Test(backends);
+    }
+
+} // TEST_SUITE("Tile_Float32_CpuRef_Test")
+
+} // namespace armnnDelegate
\ No newline at end of file
diff --git a/delegate/test/TileTestHelper.hpp b/delegate/test/TileTestHelper.hpp
new file mode 100644
index 0000000000..f376612f36
--- /dev/null
+++ b/delegate/test/TileTestHelper.hpp
@@ -0,0 +1,149 @@
+//
+// Copyright © 2023 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "TestUtils.hpp"
+
+#include <armnn_delegate.hpp>
+#include <DelegateTestInterpreter.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/kernels/register.h>
+#include <tensorflow/lite/version.h>
+
+#include <schema_generated.h>
+
+#include <doctest/doctest.h>
+
+namespace
+{
+std::vector<char> CreateTileTfLiteModel(tflite::BuiltinOperator operatorCode,
+                                        tflite::TensorType inputTensorType,
+                                        const std::vector<int32_t>& inputTensorShape,
+                                        const std::vector<int32_t>& multiplesTensorData,
+                                        const std::vector<int32_t>& multiplesTensorShape,
+                                        const std::vector<int32_t>& outputTensorShape)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
+    buffers.push_back(CreateBuffer(flatBufferBuilder));
+    buffers.push_back(CreateBuffer(flatBufferBuilder));
+    buffers.push_back(CreateBuffer(flatBufferBuilder,
+                                   flatBufferBuilder.CreateVector(
+                                           reinterpret_cast<const uint8_t*>(multiplesTensorData.data()),
+                                           sizeof(int32_t) * multiplesTensorData.size())));
+    buffers.push_back(CreateBuffer(flatBufferBuilder));
+
+    std::array<flatbuffers::Offset<Tensor>, 3> tensors;
+    tensors[0] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(),
+                                                                      inputTensorShape.size()),
+                              inputTensorType,
+                              1,
+                              flatBufferBuilder.CreateString("input_tensor"));
+
+    tensors[1] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(multiplesTensorShape.data(),
+                                                                      multiplesTensorShape.size()),
+                              TensorType_INT32,
+                              2,
+                              flatBufferBuilder.CreateString("axis_input_tensor"));
+
+    tensors[2] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(),
+                                                                      outputTensorShape.size()),
+                              inputTensorType,
+                              3,
+                              flatBufferBuilder.CreateString("output_tensor"));
+
+    // Create Operator
+    tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_NONE;
+    flatbuffers::Offset<void> operatorBuiltinOption = 0;
+
+    const std::vector<int> operatorInputs {0, 1};
+    const std::vector<int> operatorOutputs {2};
+
+    flatbuffers::Offset<Operator> tileOperator =
+            CreateOperator(flatBufferBuilder,
+                           0,
+                           flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                           flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                           operatorBuiltinOptionsType,
+                           operatorBuiltinOption);
+
+    const std::vector<int> subgraphInputs{0, 1};
+    const std::vector<int> subgraphOutputs{2};
+    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(&tileOperator, 1));
+
+    flatbuffers::Offset <flatbuffers::String> modelDescription =
+            flatBufferBuilder.CreateString("ArmnnDelegate: Tile Operator Model");
+    flatbuffers::Offset <OperatorCode> opCode = CreateOperatorCode(flatBufferBuilder, operatorCode);
+
+    flatbuffers::Offset <Model> flatbufferModel =
+            CreateModel(flatBufferBuilder,
+                        TFLITE_SCHEMA_VERSION,
+                        flatBufferBuilder.CreateVector(&opCode, 1),
+                        flatBufferBuilder.CreateVector(&subgraph, 1),
+                        modelDescription,
+                        flatBufferBuilder.CreateVector(buffers.data(), buffers.size()));
+
+    flatBufferBuilder.Finish(flatbufferModel, armnnDelegate::FILE_IDENTIFIER);
+
+    return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
+                             flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
+}
+
+void TileFP32TestImpl(tflite::BuiltinOperator operatorCode,
+                      std::vector<armnn::BackendId>& backends,
+                      std::vector<float>& inputValues,
+                      std::vector<int32_t> inputShape,
+                      std::vector<int32_t> multiplesValues,
+                      std::vector<int32_t> multiplesShapes,
+                      std::vector<float>& expectedOutputValues,
+                      std::vector<int32_t> expectedOutputShape)
+{
+    using namespace delegateTestInterpreter;
+
+    std::vector<char> modelBuffer = CreateTileTfLiteModel(operatorCode,
+                                                          ::tflite::TensorType::TensorType_FLOAT32,
+                                                          inputShape,
+                                                          multiplesValues,
+                                                          multiplesShapes,
+                                                          expectedOutputShape);
+
+    // Setup interpreter with just TFLite Runtime.
+    auto tfLiteInterpreter = DelegateTestInterpreter(modelBuffer);
+    CHECK(tfLiteInterpreter.AllocateTensors() == kTfLiteOk);
+    CHECK(tfLiteInterpreter.FillInputTensor<float>(inputValues, 0) == kTfLiteOk);
+    CHECK(tfLiteInterpreter.FillInputTensor<int32_t>(multiplesValues, 1) == kTfLiteOk);
+    CHECK(tfLiteInterpreter.Invoke() == kTfLiteOk);
+    std::vector<float>   tfLiteOutputValues = tfLiteInterpreter.GetOutputResult<float>(0);
+    std::vector<int32_t> tfLiteOutputShape  = tfLiteInterpreter.GetOutputShape(0);
+
+    // Setup interpreter with Arm NN Delegate applied.
+    auto armnnInterpreter = DelegateTestInterpreter(modelBuffer, backends);
+    CHECK(armnnInterpreter.AllocateTensors() == kTfLiteOk);
+    CHECK(armnnInterpreter.FillInputTensor<float>(inputValues, 0) == kTfLiteOk);
+    CHECK(armnnInterpreter.FillInputTensor<int32_t>(multiplesValues, 1) == kTfLiteOk);
+    CHECK(armnnInterpreter.Invoke() == kTfLiteOk);
+    std::vector<float>   armnnOutputValues = armnnInterpreter.GetOutputResult<float>(0);
+    std::vector<int32_t> armnnOutputShape  = armnnInterpreter.GetOutputShape(0);
+
+    armnnDelegate::CompareOutputData<float>(tfLiteOutputValues, armnnOutputValues, expectedOutputValues);
+    armnnDelegate::CompareOutputShape(tfLiteOutputShape, armnnOutputShape, expectedOutputShape);
+
+    tfLiteInterpreter.Cleanup();
+    armnnInterpreter.Cleanup();
+}
+
+} // anonymous namespace
\ No newline at end of file
diff --git a/docs/05_03_delegate.dox b/docs/05_03_delegate.dox
index 632afa0cf0..153d1eb4ca 100644
--- a/docs/05_03_delegate.dox
+++ b/docs/05_03_delegate.dox
@@ -201,6 +201,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 - TANH
 
+- TILE
+
 - TRANSPOSE
 
 - TRANSPOSE_CONV