IVGCVSW-5395 TfLiteDelegate: Implement the Softmax operators

Signed-off-by: James Ward <james.ward@arm.com>
Change-Id: I9f098c6b62ebb08e727aa8547e08bddc0b814705

4 files changed, 424 insertions, 9 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index 3c77dcf2f3..595784f37a 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -126,9 +126,11 @@ if(BUILD_UNIT_TESTS)
         src/test/QuantizationTestHelper.hpp
         src/test/ResizeTest.cpp
         src/test/ResizeTestHelper.hpp
+        src/test/SoftmaxTest.cpp
+        src/test/SoftmaxTestHelper.hpp
+        src/test/TestUtils.hpp
         src/test/TransposeTest.cpp
-        src/test/TransposeTestHelper.hpp
-        src/test/TestUtils.hpp)
+        src/test/TransposeTestHelper.hpp)
 
     add_executable(DelegateUnitTests ${armnnDelegate_unittest_sources})
     target_include_directories(DelegateUnitTests PRIVATE third-party)
diff --git a/delegate/src/Softmax.hpp b/delegate/src/Softmax.hpp
index ddadbc73c8..0de8e1438c 100644
--- a/delegate/src/Softmax.hpp
+++ b/delegate/src/Softmax.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>
@@ -15,19 +15,133 @@
 namespace armnnDelegate
 {
 
+TfLiteStatus ValidateSoftmaxOperator(DelegateData& delegateData,
+                                     TfLiteContext* tfLiteContext,
+                                     const armnn::TensorInfo& inputInfo,
+                                     const armnn::TensorInfo& outputTensorInfo,
+                                     const armnn::SoftmaxDescriptor& descriptor)
+{
+    bool isSupported = false;
+    FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                               tfLiteContext,
+                               IsSoftmaxSupported,
+                               delegateData.m_Backends,
+                               isSupported,
+                               inputInfo,
+                               outputTensorInfo,
+                               descriptor);
+    return isSupported ? kTfLiteOk : kTfLiteError;
+}
+
+
+TfLiteStatus ValidateLogSoftmaxOperator(DelegateData& delegateData,
+                                        TfLiteContext* tfLiteContext,
+                                        const armnn::TensorInfo& inputInfo,
+                                        const armnn::TensorInfo& outputTensorInfo,
+                                        const armnn::LogSoftmaxDescriptor& descriptor)
+{
+    bool isSupported = false;
+    FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                               tfLiteContext,
+                               IsLogSoftmaxSupported,
+                               delegateData.m_Backends,
+                               isSupported,
+                               inputInfo,
+                               outputTensorInfo,
+                               descriptor);
+    return isSupported ? kTfLiteOk : kTfLiteError;
+}
+
 TfLiteStatus VisitSoftmaxOperator(DelegateData& delegateData,
                                   TfLiteContext* tfLiteContext,
                                   TfLiteNode* tfLiteNode,
                                   int nodeIndex,
                                   int32_t softmaxOperatorCode)
 {
-    armnn::IgnoreUnused(delegateData,
-                        tfLiteContext,
-                        tfLiteNode,
-                        nodeIndex,
-                        softmaxOperatorCode);
+    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);
+
+
+    if (!delegateData.m_Network)
+    {
+        switch(softmaxOperatorCode)
+        {
+            case kTfLiteBuiltinSoftmax:
+            {
+                armnn::SoftmaxDescriptor descriptor;
+                auto* params = reinterpret_cast<TfLiteSoftmaxParams*>(tfLiteNode->builtin_data);
+                descriptor.m_Beta = params->beta;
+                return ValidateSoftmaxOperator(delegateData,
+                                               tfLiteContext,
+                                               inputTensorInfo,
+                                               outputTensorInfo,
+                                               descriptor);
+            }
+            case kTfLiteBuiltinLogSoftmax:
+            {
+                armnn::LogSoftmaxDescriptor descriptor;
+                return ValidateLogSoftmaxOperator(delegateData,
+                                                  tfLiteContext,
+                                                  inputTensorInfo,
+                                                  outputTensorInfo,
+                                                  descriptor);
+            }
+            default:
+                return kTfLiteError;
+        }
+    }
+
+    armnn::IConnectableLayer* softmaxLayer = nullptr;
+
+    switch(softmaxOperatorCode)
+    {
+        case kTfLiteBuiltinSoftmax:
+        {
+            armnn::SoftmaxDescriptor descriptor;
+            auto* params = reinterpret_cast<TfLiteSoftmaxParams*>(tfLiteNode->builtin_data);
+            descriptor.m_Beta = params->beta;
+            softmaxLayer = delegateData.m_Network->AddSoftmaxLayer(descriptor);
+            break;
+        }
+        case kTfLiteBuiltinLogSoftmax:
+        {
+            armnn::LogSoftmaxDescriptor descriptor;
+            softmaxLayer = delegateData.m_Network->AddLogSoftmaxLayer(descriptor);
+            break;
+        }
+        default:
+            return kTfLiteError;
+    }
+    ARMNN_ASSERT(softmaxLayer != nullptr);
+
+    armnn::IOutputSlot& outputSlot = softmaxLayer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
 
-    return kTfLiteError;
+    // Connect
+    return Connect(softmaxLayer, tfLiteNode, delegateData);
 }
 
 } // namespace armnnDelegate
diff --git a/delegate/src/test/SoftmaxTest.cpp b/delegate/src/test/SoftmaxTest.cpp
new file mode 100644
index 0000000000..3aacfe0a04
--- /dev/null
+++ b/delegate/src/test/SoftmaxTest.cpp
@@ -0,0 +1,129 @@
+//
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "SoftmaxTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+/// Convenience function to run softmax and log-softmax test cases
+/// \param operatorCode tflite::BuiltinOperator_SOFTMAX or tflite::BuiltinOperator_LOG_SOFTMAX
+/// \param backends armnn backends to target
+/// \param beta multiplicative parameter to the softmax function
+/// \param expectedOutput to be checked against transformed input
+void SoftmaxTestCase(tflite::BuiltinOperator operatorCode,
+                     std::vector<armnn::BackendId> backends, float beta, std::vector<float> expectedOutput) {
+    std::vector<float> input = {
+        1.0, 2.5, 3.0, 4.5, 5.0,
+        -1.0, -2.5, -3.0, -4.5, -5.0};
+    std::vector<int32_t> shape = {2, 5};
+
+    SoftmaxTest(operatorCode,
+                tflite::TensorType_FLOAT32,
+                backends,
+                shape,
+                input,
+                expectedOutput,
+                beta);
+}
+
+TEST_SUITE ("Softmax_GpuAccTests")
+{
+
+TEST_CASE ("Softmax_Standard_Beta_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    std::vector<float> expectedOutput = {0.00994190481, 0.0445565246, 0.0734612942, 0.329230666, 0.542809606,
+                                         0.710742831, 0.158588171, 0.0961885825, 0.0214625746, 0.0130177103};
+    SoftmaxTestCase(tflite::BuiltinOperator_SOFTMAX, backends, 1, expectedOutput);
+}
+
+TEST_CASE ("Softmax_Different_Beta_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    std::vector<float> expectedOutput = {0.0946234912, 0.148399189, 0.172415257, 0.270400971, 0.314161092, 0.352414012,
+                                         0.224709094, 0.193408906, 0.123322964, 0.106145054};
+    SoftmaxTestCase(tflite::BuiltinOperator_SOFTMAX, backends, 0.3, expectedOutput);
+
+}
+
+TEST_CASE ("Log_Softmax_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
+    std::vector<float> expectedOutput =
+        {-4.61099672, -3.11099672, -2.61099672, -1.11099672, -0.610996664,
+         -0.341444582, -1.84144461, -2.34144449, -3.84144449, -4.34144449};
+    SoftmaxTestCase(tflite::BuiltinOperator_LOG_SOFTMAX, backends, 0, expectedOutput);
+}
+} // TEST_SUITE ("Softmax_GpuAccTests")
+
+TEST_SUITE ("Softmax_CpuAccTests")
+{
+
+TEST_CASE ("Softmax_Standard_Beta_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    std::vector<float> expectedOutput = {0.00994190481, 0.0445565246, 0.0734612942, 0.329230666, 0.542809606,
+                                         0.710742831, 0.158588171, 0.0961885825, 0.0214625746, 0.0130177103};
+    SoftmaxTestCase(tflite::BuiltinOperator_SOFTMAX, backends, 1, expectedOutput);
+}
+
+TEST_CASE ("Softmax_Different_Beta_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    std::vector<float> expectedOutput = {
+        0.0946234912, 0.148399189, 0.172415257, 0.270400971, 0.314161092,
+        0.352414012, 0.224709094, 0.193408906, 0.123322964, 0.106145054};
+    SoftmaxTestCase(tflite::BuiltinOperator_SOFTMAX, backends, 0.3, expectedOutput);
+}
+
+TEST_CASE ("Log_Softmax_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    std::vector<float> expectedOutput =
+        {-4.61099672, -3.11099672, -2.61099672, -1.11099672, -0.610996664,
+         -0.341444582, -1.84144461, -2.34144449, -3.84144449, -4.34144449};
+    SoftmaxTestCase(tflite::BuiltinOperator_LOG_SOFTMAX, backends, 0, expectedOutput);
+}
+} // TEST_SUITE ("Softmax_CpuAccTests")
+
+TEST_SUITE ("Softmax_CpuRefTests")
+{
+
+TEST_CASE ("Softmax_Standard_Beta_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    std::vector<float> expectedOutput = {
+        0.00994190481, 0.0445565246, 0.0734612942, 0.329230666, 0.542809606,
+        0.710742831, 0.158588171, 0.0961885825, 0.0214625746, 0.0130177103};
+    SoftmaxTestCase(tflite::BuiltinOperator_SOFTMAX, backends, 1, expectedOutput);
+}
+
+TEST_CASE ("Softmax_Different_Beta_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    std::vector<float> expectedOutput = {
+        0.0946234912, 0.148399189, 0.172415257, 0.270400971, 0.314161092,
+        0.352414012, 0.224709094, 0.193408906, 0.123322964, 0.106145054};
+    SoftmaxTestCase(tflite::BuiltinOperator_SOFTMAX, backends, 0.3, expectedOutput);
+}
+
+TEST_CASE ("Log_Softmax_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    std::vector<float> expectedOutput =
+        {-4.61099672, -3.11099672, -2.61099672, -1.11099672, -0.610996664,
+         -0.341444582, -1.84144461, -2.34144449, -3.84144449, -4.34144449};
+    SoftmaxTestCase(tflite::BuiltinOperator_LOG_SOFTMAX, backends, 0, expectedOutput);
+}
+} // TEST_SUITE ("Softmax_CpuRefTests")
+} // namespace armnnDelegate
diff --git a/delegate/src/test/SoftmaxTestHelper.hpp b/delegate/src/test/SoftmaxTestHelper.hpp
new file mode 100644
index 0000000000..0474561a93
--- /dev/null
+++ b/delegate/src/test/SoftmaxTestHelper.hpp
@@ -0,0 +1,170 @@
+//
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn_delegate.hpp>
+#include <armnnUtils/FloatingPointComparison.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> CreateSoftmaxTfLiteModel(tflite::BuiltinOperator softmaxOperatorCode,
+                                           tflite::TensorType tensorType,
+                                           const std::vector <int32_t>& tensorShape,
+                                           float beta)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    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>(tensorShape.data(),
+                                                                      tensorShape.size()),
+                              tensorType,
+                              0);
+    tensors[1] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(),
+                                                                      tensorShape.size()),
+                              tensorType,
+                              0);
+
+    const std::vector<int32_t> operatorInputs({0});
+    const std::vector<int32_t> operatorOutputs({1});
+
+    flatbuffers::Offset<Operator> softmaxOperator;
+    flatbuffers::Offset<flatbuffers::String> modelDescription;
+    flatbuffers::Offset<OperatorCode> operatorCode;
+
+    switch (softmaxOperatorCode)
+    {
+        case tflite::BuiltinOperator_SOFTMAX:
+            softmaxOperator =
+                CreateOperator(flatBufferBuilder,
+                               0,
+                               flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                               flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                               BuiltinOptions_SoftmaxOptions,
+                               CreateSoftmaxOptions(flatBufferBuilder, beta).Union());
+                modelDescription = flatBufferBuilder.CreateString("ArmnnDelegate: Softmax Operator Model");
+                operatorCode = CreateOperatorCode(flatBufferBuilder,
+                                 tflite::BuiltinOperator_SOFTMAX);
+            break;
+        case tflite::BuiltinOperator_LOG_SOFTMAX:
+            softmaxOperator =
+                CreateOperator(flatBufferBuilder,
+                               0,
+                               flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                               flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                               BuiltinOptions_LogSoftmaxOptions,
+                               CreateLogSoftmaxOptions(flatBufferBuilder).Union());
+                flatBufferBuilder.CreateString("ArmnnDelegate: Log-Softmax Operator Model");
+            operatorCode = CreateOperatorCode(flatBufferBuilder,
+                                              tflite::BuiltinOperator_LOG_SOFTMAX);
+            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(&softmaxOperator, 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());
+}
+
+void SoftmaxTest(tflite::BuiltinOperator softmaxOperatorCode,
+                 tflite::TensorType tensorType,
+                 std::vector<armnn::BackendId>& backends,
+                 std::vector<int32_t>& shape,
+                 std::vector<float>& inputValues,
+                 std::vector<float>& expectedOutputValues,
+                 float beta = 0)
+{
+    using namespace tflite;
+    std::vector<char> modelBuffer = CreateSoftmaxTfLiteModel(softmaxOperatorCode,
+                                                             tensorType,
+                                                             shape,
+                                                             beta);
+
+    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 tfLiteInterpreterInputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateInputId);
+    for (unsigned int i = 0; i < inputValues.size(); ++i)
+    {
+        tfLiteInterpreterInputData[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 tfLiteInterpreterOutputId = tfLiteInterpreter->outputs()[0];
+    auto tfLiteInterpreterOutputData = tfLiteInterpreter->typed_tensor<float>(tfLiteInterpreterOutputId);
+    auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0];
+    auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateOutputId);
+
+    for (size_t i = 0; i < inputValues.size(); ++i)
+    {
+         CHECK(armnnUtils::within_percentage_tolerance(expectedOutputValues[i], armnnDelegateOutputData[i], 1e-5));
+         CHECK(armnnUtils::within_percentage_tolerance(tfLiteInterpreterOutputData[i],
+                                                       armnnDelegateOutputData[i], 1e-5));
+    }
+}
+
+} // anonymous namespace