IVGCVSW-5382 TfLiteDelegate: Implement the Activation operators

 * Added TfLiteDelegate implementations for ReLu, Relu6, Logistic, and TanH
Activation Functions

Signed-off-by: David Monahan <david.monahan@arm.com>
Change-Id: Id021b4ec9c10fd4357535fe2a665f32c053dad61

4 files changed, 447 insertions, 8 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index 1ea0cdd644..3c77dcf2f3 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -104,6 +104,8 @@ option(BUILD_UNIT_TESTS "Build unit tests" ON)
 if(BUILD_UNIT_TESTS)
     set(armnnDelegate_unittest_sources)
     list(APPEND armnnDelegate_unittest_sources
+        src/test/ActivationTest.cpp
+        src/test/ActivationTestHelper.hpp
         src/test/ArmnnDelegateTest.cpp
         src/test/ComparisonTest.cpp
         src/test/ComparisonTestHelper.hpp
diff --git a/delegate/src/Activation.hpp b/delegate/src/Activation.hpp
index 1ffa7d7f8c..5e8d876110 100644
--- a/delegate/src/Activation.hpp
+++ b/delegate/src/Activation.hpp
@@ -5,7 +5,7 @@
 
 #pragma once
 
-#include <armnn/utility/IgnoreUnused.hpp>
+#include "DelegateUtils.hpp"
 
 #include <tensorflow/lite/builtin_ops.h>
 #include <tensorflow/lite/c/builtin_op_data.h>
@@ -15,20 +15,108 @@
 namespace armnnDelegate
 {
 
+TfLiteStatus ValidateActivationOperator(DelegateData& delegateData,
+                                        TfLiteContext* tfLiteContext,
+                                        const armnn::TensorInfo& inputInfo,
+                                        const armnn::TensorInfo& outputInfo,
+                                        armnn::ActivationDescriptor& activationDesc)
+{
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported)
+    {
+        FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                                   tfLiteContext,
+                                   IsActivationSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputInfo,
+                                   outputInfo,
+                                   activationDesc);
+    };
+
+    validateFunc(outputInfo, isSupported);
+    return isSupported ? kTfLiteOk : kTfLiteError;
+}
+
 TfLiteStatus VisitActivationOperator(DelegateData& delegateData,
                                      TfLiteContext* tfLiteContext,
                                      TfLiteNode* tfLiteNode,
                                      int nodeIndex,
-                                     int32_t comparisonOperatorCode)
+                                     int32_t operatorCode)
 {
-    armnn::IgnoreUnused(delegateData,
-                        tfLiteContext,
-                        tfLiteNode,
-                        nodeIndex,
-                        comparisonOperatorCode);
+    TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
+    TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
+
+    const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors;
+    const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]];
+    if (IsDynamicTensor(tfLiteInputTensor))
+    {
+        TF_LITE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnDelegate: Dynamic input tensors are not supported in node #%d: ",
+            nodeIndex);
+        return kTfLiteError;
+    }
+    const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
+    if (IsDynamicTensor(tfLiteOutputTensor))
+    {
+        TF_LITE_MAYBE_KERNEL_LOG(
+            tfLiteContext,
+            "TfLiteArmnnDelegate: Dynamic output tensors are not supported in node #%d: ",
+            nodeIndex);
+        return kTfLiteError;
+    }
+
+    const armnn::TensorInfo& inputTensorInfo  = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
+
+    armnn::ActivationDescriptor activationDesc;
+    switch(operatorCode)
+    {
+        case kTfLiteBuiltinRelu:
+        {
+            activationDesc.m_Function = armnn::ActivationFunction::ReLu;
+            break;
+        }
+        case kTfLiteBuiltinRelu6:
+        {
+            activationDesc.m_Function = armnn::ActivationFunction::BoundedReLu;
+            activationDesc.m_A = 6.0f;
+            break;
+        }
+        case kTfLiteBuiltinLogistic:
+        {
+            activationDesc.m_Function = armnn::ActivationFunction::Sigmoid;
+            break;
+        }
+        case kTfLiteBuiltinTanh:
+        {
+            activationDesc.m_Function = armnn::ActivationFunction::TanH;
+            activationDesc.m_A = 1.0f;
+            activationDesc.m_B = 1.0f;
+            break;
+        }
+        default:
+        {
+            return kTfLiteError;
+        }
+    }
+    if (!delegateData.m_Network)
+    {
+        return ValidateActivationOperator(delegateData,
+                                          tfLiteContext,
+                                          inputTensorInfo,
+                                          outputTensorInfo,
+                                          activationDesc);
+    }
+    armnn::IConnectableLayer* activationLayer = delegateData.m_Network->AddActivationLayer(activationDesc);
+    ARMNN_ASSERT(activationLayer != nullptr);
 
+    armnn::IOutputSlot& outputSlot = activationLayer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
 
-    return kTfLiteError;
+    // Connect
+    return Connect(activationLayer, tfLiteNode, delegateData);
 }
 
 } // namespace armnnDelegate
diff --git a/delegate/src/test/ActivationTest.cpp b/delegate/src/test/ActivationTest.cpp
new file mode 100644
index 0000000000..f894d67372
--- /dev/null
+++ b/delegate/src/test/ActivationTest.cpp
@@ -0,0 +1,209 @@
+//
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ActivationTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/interpreter.h>
+#include <tensorflow/lite/kernels/register.h>
+#include <tensorflow/lite/model.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+#include <tensorflow/lite/version.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+
+void ActivationReLuTest(std::vector<armnn::BackendId>& backends)
+{
+
+    std::vector<float> inputData = {
+            -0.1f, -0.2f, -0.3f, -0.4f,
+            0.1f,  0.2f,  0.3f,  0.4f,
+            -1.0f, -2.0f, -3.0f, -4.0f,
+            1.0f,  2.0f,  3.0f,  4.0f
+    };
+
+    // Calculate output values for input.
+    auto f = [](float value)
+    {
+        return std::fmax(0.0f, value);
+    };
+    std::vector<float> outputExpectedData(inputData.size());
+    std::transform(inputData.begin(), inputData.end(), outputExpectedData.begin(), f);
+
+    ActivationTest(tflite::BuiltinOperator_RELU,
+                   backends,
+                   inputData,
+                   outputExpectedData);
+}
+
+void ActivationBoundedReluTest(std::vector<armnn::BackendId>& backends)
+{
+    std::vector<float> inputData = {
+            -0.1f, -0.2f, -0.3f, -0.4f,
+            0.1f,  0.2f,  0.3f,  0.4f,
+            -1.0f, -2.0f, -3.0f, -4.0f,
+            1.0f,  2.0f,  3.0f,  4.0f
+    };
+
+    const float a = 6.0f;
+    const float b = 0.0f;
+    // Calculate output values for input.
+    auto f = [a, b](float value)
+    {
+        return std::min(a, std::max(b, value));
+    };
+    std::vector<float> outputExpectedData(inputData.size());
+    std::transform(inputData.begin(), inputData.end(), outputExpectedData.begin(), f);
+
+    ActivationTest(tflite::BuiltinOperator_RELU6,
+                   backends,
+                   inputData,
+                   outputExpectedData);
+}
+
+void ActivationSigmoidTest(std::vector<armnn::BackendId>& backends)
+{
+    std::vector<float> inputData = {
+            -0.1f, -0.2f, -0.3f, -0.4f,
+            0.1f,  0.2f,  0.3f,  0.4f,
+            -1.0f, -2.0f, -3.0f, -4.0f,
+            1.0f,  2.0f,  3.0f,  4.0f
+    };
+
+    // Calculate output values for input.
+    auto f = [](float value)
+    {
+        return 1.0f / (1.0f + std::exp(-value));
+    };
+    std::vector<float> outputExpectedData(inputData.size());
+    std::transform(inputData.begin(), inputData.end(), outputExpectedData.begin(), f);
+
+    ActivationTest(tflite::BuiltinOperator_LOGISTIC,
+                   backends,
+                   inputData,
+                   outputExpectedData);
+}
+
+
+void ActivationTanHTest(std::vector<armnn::BackendId>& backends)
+{
+    std::vector<float> inputData = {
+            -0.1f, -0.2f, -0.3f, -0.4f,
+            0.1f,  0.2f,  0.3f,  0.4f,
+            -1.0f, -2.0f, -3.0f, -4.0f,
+            1.0f,  2.0f,  3.0f,  4.0f
+    };
+
+    // Calculate output values for input.
+    auto f = [](float value)
+    {
+        return tanhf(value);
+    };
+    std::vector<float> outputExpectedData(inputData.size());
+    std::transform(inputData.begin(), inputData.end(), outputExpectedData.begin(), f);
+
+    ActivationTest(tflite::BuiltinOperator_TANH,
+                   backends,
+                   inputData,
+                   outputExpectedData);
+}
+
+TEST_SUITE("Activation_CpuRefTests")
+{
+
+TEST_CASE ("Activation_ReLu_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    ActivationReLuTest(backends);
+}
+
+TEST_CASE ("Activation_Bounded_Relu6_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    ActivationBoundedReluTest(backends);
+}
+
+TEST_CASE ("Activation_Sigmoid_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    ActivationSigmoidTest(backends);
+}
+
+
+TEST_CASE ("Activation_TanH_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    ActivationTanHTest(backends);
+}
+
+}
+
+TEST_SUITE("Activation_CpuAccTests")
+{
+
+TEST_CASE ("Activation_ReLu_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    ActivationReLuTest(backends);
+}
+
+TEST_CASE ("Activation_Bounded_Relu6_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    ActivationBoundedReluTest(backends);
+}
+
+TEST_CASE ("Activation_Sigmoid_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    ActivationSigmoidTest(backends);
+}
+
+
+TEST_CASE ("Activation_TanH_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    ActivationTanHTest(backends);
+}
+
+}
+
+TEST_SUITE("Activation_GpuAccTests")
+{
+
+TEST_CASE ("Activation_ReLu_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    ActivationReLuTest(backends);
+}
+
+TEST_CASE ("Activation_Bounded_Relu6_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    ActivationBoundedReluTest(backends);
+}
+
+TEST_CASE ("Activation_Sigmoid_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    ActivationSigmoidTest(backends);
+}
+
+
+TEST_CASE ("Activation_TanH_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    ActivationTanHTest(backends);
+}
+
+}
+
+} // namespace armnnDelegate
\ No newline at end of file
diff --git a/delegate/src/test/ActivationTestHelper.hpp b/delegate/src/test/ActivationTestHelper.hpp
new file mode 100644
index 0000000000..04da3ccc49
--- /dev/null
+++ b/delegate/src/test/ActivationTestHelper.hpp
@@ -0,0 +1,140 @@
+//
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#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> CreateActivationTfLiteModel(tflite::BuiltinOperator activationOperatorCode,
+                                              tflite::TensorType tensorType,
+                                              const std::vector <int32_t>& tensorShape)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers;
+    buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({}));
+
+    std::array<flatbuffers::Offset<Tensor>, 2> tensors;
+    tensors[0] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(), tensorShape.size()),
+                              tensorType);
+    tensors[1] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(), tensorShape.size()),
+                              tensorType);
+
+    // create operator
+    const std::vector<int> operatorInputs{{0}};
+    const std::vector<int> operatorOutputs{{1}};
+    flatbuffers::Offset <Operator> unaryOperator =
+        CreateOperator(flatBufferBuilder,
+                       0,
+                       flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()));
+
+    const std::vector<int> subgraphInputs{{0}};
+    const std::vector<int> subgraphOutputs{{1}};
+    flatbuffers::Offset <SubGraph> subgraph =
+        CreateSubGraph(flatBufferBuilder,
+                       flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(), subgraphOutputs.size()),
+                       flatBufferBuilder.CreateVector(&unaryOperator, 1));
+
+    flatbuffers::Offset <flatbuffers::String> modelDescription =
+        flatBufferBuilder.CreateString("ArmnnDelegate: Activation Operator Model");
+    flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, activationOperatorCode);
+
+    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());
+}
+
+void ActivationTest(tflite::BuiltinOperator activationOperatorCode,
+                    std::vector<armnn::BackendId>& backends,
+                    std::vector<float>& inputValues,
+                    std::vector<float>& expectedOutputValues)
+{
+    using namespace tflite;
+    const std::vector<int32_t> inputShape  { { 4, 1, 4} };
+    std::vector<char> modelBuffer = CreateActivationTfLiteModel(activationOperatorCode,
+                                                                      ::tflite::TensorType_FLOAT32,
+                                                                      inputShape);
+
+    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
+    auto tfLiteDelegateInputId = tfLiteInterpreter->inputs()[0];
+    auto tfLiteDelageInputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateInputId);
+    for (unsigned int i = 0; i < inputValues.size(); ++i)
+    {
+        tfLiteDelageInputData[i] = inputValues[i];
+    }
+
+    auto armnnDelegateInputId = armnnDelegateInterpreter->inputs()[0];
+    auto armnnDelegateInputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateInputId);
+    for (unsigned int i = 0; i < inputValues.size(); ++i)
+    {
+        armnnDelegateInputData[i] = inputValues[i];
+    }
+    // Run EnqueWorkload
+    CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
+
+    // Compare output data
+    auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[0];
+    auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateOutputId);
+    auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0];
+    auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateOutputId);
+    for (size_t i = 0; i < inputValues.size(); i++)
+    {
+        CHECK(expectedOutputValues[i] == doctest::Approx(armnnDelegateOutputData[i]));
+        CHECK(tfLiteDelageOutputData[i] == doctest::Approx(armnnDelegateOutputData[i]));
+    }
+}
+
+} // anonymous namespace
\ No newline at end of file