IVGCVSW-5401 Implement the FILL operator

 * Added FILL operator to TfLite ArmNN Delegate
 * Added unit tests

Signed-off-by: David Monahan <david.monahan@arm.com>
Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I335ef469ff773fa4305eb87f6e93ae9c03fc6997

6 files changed, 486 insertions, 7 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index 981fc9f0bf..5dbe83e014 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -131,6 +131,8 @@ if(BUILD_UNIT_TESTS)
         src/test/ElementwiseBinaryTestHelper.hpp
         src/test/ElementwiseUnaryTest.cpp
         src/test/ElementwiseUnaryTestHelper.hpp
+        src/test/FillTest.cpp
+        src/test/FillTestHelper.hpp
         src/test/FullyConnectedTest.cpp
         src/test/FullyConnectedTestHelper.hpp
         src/test/GatherTest.cpp
diff --git a/delegate/src/Fill.hpp b/delegate/src/Fill.hpp
index 99c3c625c2..c9fd159b3e 100644
--- a/delegate/src/Fill.hpp
+++ b/delegate/src/Fill.hpp
@@ -19,15 +19,99 @@ TfLiteStatus VisitFillOperator(DelegateData& delegateData,
                                TfLiteContext* tfLiteContext,
                                TfLiteNode* tfLiteNode,
                                int nodeIndex,
-                               int32_t operatorCode)
+                               int32_t tfLiteFillOperatorCode)
 {
-    armnn::IgnoreUnused(delegateData,
-                        tfLiteContext,
-                        tfLiteNode,
-                        nodeIndex,
-                        operatorCode);
+    TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
 
-    return kTfLiteError;
+    switch(tfLiteFillOperatorCode)
+    {
+        case kTfLiteBuiltinFill:
+            TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 2, nodeIndex));
+            break;
+        default:
+            return kTfLiteError;
+    }
+
+    const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors;
+    const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]];
+    if (!IsValid(tfLiteContext, tfLiteInputTensor, tfLiteFillOperatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const TfLiteTensor& tfLiteFillTensor = tfLiteTensors[tfLiteNode->inputs->data[1]];
+    if (!IsValid(tfLiteContext, tfLiteFillTensor, tfLiteFillOperatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
+    if (!IsValid(tfLiteContext, tfLiteOutputTensor, tfLiteFillOperatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    armnn::TensorInfo inputTensorInfo  = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& fillTensorInfo   = GetTensorInfoForTfLiteTensor(tfLiteFillTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
+
+    armnn::FillDescriptor descriptor;
+    switch (tfLiteFillTensor.type)
+    {
+        case kTfLiteFloat32:
+            descriptor.m_Value = tflite::GetTensorData<float>(&tfLiteFillTensor)[0];
+            break;
+        case kTfLiteInt32:
+            descriptor.m_Value = tflite::GetTensorData<int32_t>(&tfLiteFillTensor)[0];
+            break;
+        default:
+            TF_LITE_MAYBE_KERNEL_LOG(
+                tfLiteContext,
+                "TfLiteArmnnDelegate: FILL value data type is not supported in operator #%d node #%d: ",
+                tfLiteFillOperatorCode, nodeIndex);
+            return kTfLiteError;
+    }
+
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported)
+    {
+        FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                                   tfLiteContext,
+                                   IsFillSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputTensorInfo,
+                                   outInfo,
+                                   descriptor);
+    };
+
+    if (!delegateData.m_Network)
+    {
+        validateFunc(outputTensorInfo, isSupported);
+        return isSupported ? kTfLiteOk : kTfLiteError;
+    }
+
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddFillLayer(descriptor);
+    ARMNN_ASSERT(layer != nullptr);
+
+    armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
+
+    if(tflite::IsConstantTensor(&tfLiteInputTensor))
+    {
+        auto status = ConnectConstant(layer,
+                                      inputTensorInfo,
+                                      tfLiteContext,
+                                      tfLiteInputTensor,
+                                      delegateData,
+                                      tfLiteNode->inputs->data[0]);
+        if (status == kTfLiteError)
+        {
+            return status;
+        }
+    }
+
+    return Connect(layer, tfLiteNode, delegateData);
 }
 
 } // namespace armnnDelegate
diff --git a/delegate/src/test/FillTest.cpp b/delegate/src/test/FillTest.cpp
new file mode 100644
index 0000000000..50f7f53d56
--- /dev/null
+++ b/delegate/src/test/FillTest.cpp
@@ -0,0 +1,221 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "FillTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+void Fill2dTest(std::vector<armnn::BackendId>& backends,
+               tflite::BuiltinOperator fillOperatorCode = tflite::BuiltinOperator_FILL,
+               float fill = 2.0f )
+{
+    std::vector<int32_t> inputShape { 2 };
+    std::vector<int32_t> tensorShape { 2, 2 };
+    std::vector<float> expectedOutputValues = { fill, fill,
+                                                fill, fill };
+
+    FillTest<float>(fillOperatorCode,
+                    ::tflite::TensorType_FLOAT32,
+                    backends,
+                    inputShape,
+                    tensorShape,
+                    expectedOutputValues,
+                    fill);
+}
+
+void Fill3dTest(std::vector<armnn::BackendId>& backends,
+               tflite::BuiltinOperator fillOperatorCode = tflite::BuiltinOperator_FILL,
+               float fill = 5.0f )
+{
+    std::vector<int32_t> inputShape { 3 };
+    std::vector<int32_t> tensorShape { 3, 3, 3 };
+    std::vector<float> expectedOutputValues = { fill, fill, fill,
+                                                fill, fill, fill,
+                                                fill, fill, fill,
+
+                                                fill, fill, fill,
+                                                fill, fill, fill,
+                                                fill, fill, fill,
+
+                                                fill, fill, fill,
+                                                fill, fill, fill,
+                                                fill, fill, fill };
+
+    FillTest<float>(fillOperatorCode,
+                    ::tflite::TensorType_FLOAT32,
+                    backends,
+                    inputShape,
+                    tensorShape,
+                    expectedOutputValues,
+                    fill);
+}
+
+void Fill4dTest(std::vector<armnn::BackendId>& backends,
+               tflite::BuiltinOperator fillOperatorCode = tflite::BuiltinOperator_FILL,
+               float fill = 3.0f )
+{
+    std::vector<int32_t> inputShape { 4 };
+    std::vector<int32_t> tensorShape { 2, 2, 4, 4 };
+    std::vector<float> expectedOutputValues = { fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill,
+                                                fill, fill, fill, fill };
+
+    FillTest<float>(fillOperatorCode,
+                    ::tflite::TensorType_FLOAT32,
+                    backends,
+                    inputShape,
+                    tensorShape,
+                    expectedOutputValues,
+                    fill);
+}
+
+void FillInt32Test(std::vector<armnn::BackendId>& backends,
+                  tflite::BuiltinOperator fillOperatorCode = tflite::BuiltinOperator_FILL,
+                  int32_t fill = 2 )
+{
+    std::vector<int32_t> inputShape { 2 };
+    std::vector<int32_t> tensorShape { 2, 2 };
+    std::vector<int32_t> expectedOutputValues = { fill, fill,
+                                                  fill, fill };
+
+    FillTest<int32_t>(fillOperatorCode,
+                      ::tflite::TensorType_INT32,
+                      backends,
+                      inputShape,
+                      tensorShape,
+                      expectedOutputValues,
+                      fill);
+}
+
+TEST_SUITE("Fill_CpuRefTests")
+{
+
+TEST_CASE ("Fill2d_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    Fill2dTest(backends);
+}
+
+TEST_CASE ("Fill3d_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    Fill3dTest(backends);
+}
+
+TEST_CASE ("Fill3d_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    Fill3dTest(backends);
+}
+
+TEST_CASE ("Fill4d_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    Fill4dTest(backends);
+}
+
+TEST_CASE ("FillInt32_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    FillInt32Test(backends);
+}
+
+}
+
+TEST_SUITE("Fill_CpuAccTests")
+{
+
+TEST_CASE ("Fill2d_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    Fill2dTest(backends);
+}
+
+TEST_CASE ("Fill3d_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    Fill3dTest(backends);
+}
+
+TEST_CASE ("Fill3d_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    Fill3dTest(backends);
+}
+
+TEST_CASE ("Fill4d_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    Fill4dTest(backends);
+}
+
+TEST_CASE ("FillInt32_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    FillInt32Test(backends);
+}
+
+}
+
+TEST_SUITE("Fill_GpuAccTests")
+{
+
+TEST_CASE ("Fill2d_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    Fill2dTest(backends);
+}
+
+TEST_CASE ("Fill3d_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    Fill3dTest(backends);
+}
+
+TEST_CASE ("Fill3d_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    Fill3dTest(backends);
+}
+
+TEST_CASE ("Fill4d_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    Fill4dTest(backends);
+}
+
+TEST_CASE ("FillInt32_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    FillInt32Test(backends);
+}
+
+}
+
+} // namespace armnnDelegate
\ No newline at end of file
diff --git a/delegate/src/test/FillTestHelper.hpp b/delegate/src/test/FillTestHelper.hpp
new file mode 100644
index 0000000000..e6890a2b2d
--- /dev/null
+++ b/delegate/src/test/FillTestHelper.hpp
@@ -0,0 +1,160 @@
+//
+// 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
+{
+
+template <typename T>
+std::vector<char> CreateFillTfLiteModel(tflite::BuiltinOperator fillOperatorCode,
+                                        tflite::TensorType tensorType,
+                                        const std::vector<int32_t>& inputShape,
+                                        const std::vector <int32_t>& tensorShape,
+                                        const std::vector<T> fillValue)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
+    buffers.push_back(
+        CreateBuffer(flatBufferBuilder,
+                     flatBufferBuilder.CreateVector({})));
+    buffers.push_back(
+        CreateBuffer(flatBufferBuilder,
+                     flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(tensorShape.data()),
+                                                    sizeof(int32_t) * tensorShape.size())));
+    buffers.push_back(
+        CreateBuffer(flatBufferBuilder,
+                     flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(fillValue.data()),
+                                                    sizeof(T) * fillValue.size())));
+
+    std::array<flatbuffers::Offset<Tensor>, 3> tensors;
+    tensors[0] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(inputShape.data(),
+                                                                      inputShape.size()),
+                              tflite::TensorType_INT32,
+                              1,
+                              flatBufferBuilder.CreateString("dims"));
+
+    std::vector<int32_t> fillShape = {};
+    tensors[1] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(fillShape.data(),
+                                                                      fillShape.size()),
+                              tensorType,
+                              2,
+                              flatBufferBuilder.CreateString("value"));
+
+    tensors[2] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(),
+                                                                      tensorShape.size()),
+                              tensorType,
+                              0,
+                              flatBufferBuilder.CreateString("output"));
+
+    tflite::BuiltinOptions operatorBuiltinOptionsType = BuiltinOptions_FillOptions;
+    flatbuffers::Offset<void> operatorBuiltinOptions = CreateFillOptions(flatBufferBuilder).Union();
+
+    // create operator
+    const std::vector<int> operatorInputs{ {0, 1} };
+    const std::vector<int> operatorOutputs{ 2 };
+    flatbuffers::Offset <Operator> fillOperator =
+        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> 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(&fillOperator, 1));
+
+    flatbuffers::Offset <flatbuffers::String> modelDescription =
+        flatBufferBuilder.CreateString("ArmnnDelegate: Fill Operator Model");
+    flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder,
+                                                                         fillOperatorCode);
+
+    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 FillTest(tflite::BuiltinOperator fillOperatorCode,
+              tflite::TensorType tensorType,
+              const std::vector<armnn::BackendId>& backends,
+              std::vector<int32_t >& inputShape,
+              std::vector<int32_t >& tensorShape,
+              std::vector<T>& expectedOutputValues,
+              T fillValue)
+{
+    using namespace tflite;
+    std::vector<char> modelBuffer = CreateFillTfLiteModel<T>(fillOperatorCode,
+                                                             tensorType,
+                                                             inputShape,
+                                                             tensorShape,
+                                                             {fillValue});
+
+    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);
+
+    // Run EnqueueWorkload
+    CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
+
+    armnnDelegate::CompareOutputData<T>(tfLiteInterpreter, armnnDelegateInterpreter, tensorShape, expectedOutputValues);
+}
+
+} // anonymous namespace
diff --git a/delegate/src/test/TestUtils.cpp b/delegate/src/test/TestUtils.cpp
index 1bc5786112..bbe89904eb 100644
--- a/delegate/src/test/TestUtils.cpp
+++ b/delegate/src/test/TestUtils.cpp
@@ -52,6 +52,15 @@ void CompareData(int16_t tensor1[], int16_t tensor2[], size_t tensorSize)
     }
 }
 
+void CompareData(int32_t tensor1[], int32_t tensor2[], size_t tensorSize)
+{
+    int32_t tolerance = 1;
+    for (size_t i = 0; i < tensorSize; i++)
+    {
+        CHECK(std::max(tensor1[i], tensor2[i]) - std::min(tensor1[i], tensor2[i]) <= tolerance);
+    }
+}
+
 void CompareData(int8_t tensor1[], int8_t tensor2[], size_t tensorSize)
 {
     int8_t tolerance = 1;
diff --git a/delegate/src/test/TestUtils.hpp b/delegate/src/test/TestUtils.hpp
index d14e1edb45..8a2756f4c5 100644
--- a/delegate/src/test/TestUtils.hpp
+++ b/delegate/src/test/TestUtils.hpp
@@ -51,6 +51,9 @@ void CompareData(uint8_t tensor1[], uint8_t tensor2[], size_t tensorSize);
 /// Can be used to compare int16_t data coming from a tflite interpreter with a tolerance of 1
 void CompareData(int16_t tensor1[], int16_t tensor2[], size_t tensorSize);
 
+/// Can be used to compare int32_t data coming from a tflite interpreter with a tolerance of 1
+void CompareData(int32_t tensor1[], int32_t tensor2[], size_t tensorSize);
+
 /// Can be used to compare Half (Float16) data with a tolerance of limit_of_float*100
 void CompareData(Half tensor1[], Half tensor2[], size_t tensorSize);