IVGCVSW-6603 'Add a no fallback mode to the TfLite Delegate'

* Added disable-tflite-runtime-fallback option to armnn_delegate
* Updated armnn_delegate version

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I449b16404d3ffe98e6dac52a43e7c25225addd73

6 files changed, 238 insertions, 3 deletions

diff --git a/delegate/include/DelegateOptions.hpp b/delegate/include/DelegateOptions.hpp
index d789ea7285..2b0107e834 100644
--- a/delegate/include/DelegateOptions.hpp
+++ b/delegate/include/DelegateOptions.hpp
@@ -174,6 +174,12 @@ public:
      *                 This is an Experimental parameter that is incompatible with "infer-output-shape". \n
      *                 This parameter may be removed in a later update.
      *
+     *    Option key: "disable-tflite-runtime-fallback" \n
+     *    Possible values: ["true"/"false"] \n
+     *    Description: Disable TfLite Runtime fallback in the Arm NN TfLite delegate.
+     *                 An exception will be thrown if unsupported operators are encountered.
+     *                 This option is only for testing purposes.
+     *
      * @param[in]     option_keys     Delegate option names
      * @param[in]     options_values  Delegate option values
      * @param[in]     num_options     Number of delegate options
@@ -262,6 +268,15 @@ public:
         return m_RuntimeOptions;
     }
 
+    void DisableTfLiteRuntimeFallback(bool fallbackState)
+    {
+        m_DisableTfLiteRuntimeFallback = fallbackState;
+    }
+    bool TfLiteRuntimeFallbackDisabled()
+    {
+        return m_DisableTfLiteRuntimeFallback;
+    }
+
 private:
     /// Which backend to run Delegate on.
     /// Examples of possible values are: CpuRef, CpuAcc, GpuAcc.
@@ -295,6 +310,9 @@ private:
 
     /// If not empty then the optimized model will be serialized to a file with this file name in "dot" format.
     std::string m_SerializeToDot = "";
+
+    /// Option to disable TfLite Runtime fallback for unsupported operators.
+    bool m_DisableTfLiteRuntimeFallback = false;
 };
 
 } // namespace armnnDelegate
diff --git a/delegate/include/Version.hpp b/delegate/include/Version.hpp
index c14857e320..36d8fb54af 100644
--- a/delegate/include/Version.hpp
+++ b/delegate/include/Version.hpp
@@ -13,8 +13,8 @@ namespace armnnDelegate
 #define STRINGIFY_MACRO(s) #s
 
 // ArmNN Delegate version components
-#define DELEGATE_MAJOR_VERSION 26
-#define DELEGATE_MINOR_VERSION 1
+#define DELEGATE_MAJOR_VERSION 27
+#define DELEGATE_MINOR_VERSION 0
 #define DELEGATE_PATCH_VERSION 0
 
 /// DELEGATE_VERSION: "X.Y.Z"
diff --git a/delegate/src/DelegateOptions.cpp b/delegate/src/DelegateOptions.cpp
index f3e13c90c6..a55a579333 100644
--- a/delegate/src/DelegateOptions.cpp
+++ b/delegate/src/DelegateOptions.cpp
@@ -242,6 +242,12 @@ DelegateOptions::DelegateOptions(char const* const* options_keys,
         {
             this->SetSerializeToDot(options_values[i]);
         }
+
+            // Process disable-tflite-runtime-fallback
+        else if (std::string(options_keys[i]) == std::string("disable-tflite-runtime-fallback"))
+        {
+            this->DisableTfLiteRuntimeFallback(armnn::stringUtils::StringToBool(options_values[i]));
+        }
         else
         {
             throw armnn::Exception("Unknown option for the ArmNN Delegate given: " + std::string(options_keys[i]));
diff --git a/delegate/src/armnn_delegate.cpp b/delegate/src/armnn_delegate.cpp
index eac3862c18..c041dd1714 100644
--- a/delegate/src/armnn_delegate.cpp
+++ b/delegate/src/armnn_delegate.cpp
@@ -222,6 +222,15 @@ TfLiteIntArray* Delegate::IdentifyOperatorsToDelegate(TfLiteContext* tfLiteConte
                            *it);
     }
 
+    if (!unsupportedOperators.empty() && m_Options.TfLiteRuntimeFallbackDisabled())
+    {
+        std::stringstream exMessage;
+        exMessage << "TfLiteArmnnDelegate: There are unsupported operators in the model. ";
+        exMessage << "Not falling back to TfLite Runtime as fallback is disabled. ";
+        exMessage << "This should only be disabled under test conditions.";
+        throw armnn::Exception(exMessage.str());
+    }
+
     std::sort(&nodesToDelegate->data[0], &nodesToDelegate->data[nodesToDelegate->size]);
     return nodesToDelegate;
 }
diff --git a/delegate/src/test/DelegateOptionsTest.cpp b/delegate/src/test/DelegateOptionsTest.cpp
index c9f1530968..50d3f78563 100644
--- a/delegate/src/test/DelegateOptionsTest.cpp
+++ b/delegate/src/test/DelegateOptionsTest.cpp
@@ -150,6 +150,79 @@ TEST_CASE ("ArmnnDelegateOptimizerOptionsImport")
                                 delegateOptions);
 }
 
+TEST_CASE ("ArmnnDelegateStringParsingOptionDisableTfLiteRuntimeFallback")
+{
+    std::stringstream stringStream;
+    std::vector<std::string> keys   {  "backends", "debug-data", "disable-tflite-runtime-fallback"};
+    std::vector<std::string> values {    "CpuRef",          "1",                               "1"};
+
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    std::vector<int32_t> tensorShape { 1, 2, 2, 1 };
+    std::vector<float> inputData = { 0.1f, -2.1f, 3.0f, -4.6f };
+    std::vector<float> expectedResult = { 1.0f, -2.0f, 3.0f, -4.0f };
+
+    // Create options_keys and options_values char array
+    size_t num_options = keys.size();
+    std::unique_ptr<const char*> options_keys =
+            std::unique_ptr<const char*>(new const char*[num_options + 1]);
+    std::unique_ptr<const char*> options_values =
+            std::unique_ptr<const char*>(new const char*[num_options + 1]);
+    for (size_t i=0; i<num_options; ++i)
+    {
+        options_keys.get()[i]   = keys[i].c_str();
+        options_values.get()[i] = values[i].c_str();
+    }
+
+    StreamRedirector redirect(std::cout, stringStream.rdbuf());
+
+    armnnDelegate::DelegateOptions delegateOptions(options_keys.get(), options_values.get(), num_options, nullptr);
+    DelegateOptionNoFallbackTest<float>(::tflite::TensorType_FLOAT32,
+                                        backends,
+                                        tensorShape,
+                                        inputData,
+                                        expectedResult,
+                                        delegateOptions);
+    CHECK(stringStream.str().find("TfLiteArmnnDelegate: There are unsupported operators in the model")
+                                                                                                 != std::string::npos);
+}
+
+TEST_CASE ("ArmnnDelegateStringParsingOptionEnableTfLiteRuntimeFallback")
+{
+    std::stringstream stringStream;
+    std::vector<std::string> keys   {  "backends", "debug-data", "disable-tflite-runtime-fallback"};
+    std::vector<std::string> values {    "CpuRef",          "1",                               "0"};
+
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
+    std::vector<int32_t> tensorShape { 1, 2, 2, 1 };
+    std::vector<float> inputData = { 0.1f, -2.1f, 3.0f, -4.6f };
+    std::vector<float> expectedResult = { 1.0f, -2.0f, 3.0f, -4.0f };
+
+    // Create options_keys and options_values char array
+    size_t num_options = keys.size();
+    std::unique_ptr<const char*> options_keys =
+            std::unique_ptr<const char*>(new const char*[num_options + 1]);
+    std::unique_ptr<const char*> options_values =
+            std::unique_ptr<const char*>(new const char*[num_options + 1]);
+    for (size_t i=0; i<num_options; ++i)
+    {
+        options_keys.get()[i]   = keys[i].c_str();
+        options_values.get()[i] = values[i].c_str();
+    }
+
+    StreamRedirector redirect(std::cout, stringStream.rdbuf());
+
+    armnnDelegate::DelegateOptions delegateOptions(options_keys.get(), options_values.get(), num_options, nullptr);
+    DelegateOptionNoFallbackTest<float>(::tflite::TensorType_FLOAT32,
+                                        backends,
+                                        tensorShape,
+                                        inputData,
+                                        expectedResult,
+                                        delegateOptions);
+
+    CHECK(stringStream.str().find("TfLiteArmnnDelegate: There are unsupported operators in the model")
+                                                                                                 == std::string::npos);
+}
+
 }
 
 TEST_SUITE("DelegateOptions_CpuAccTests")
@@ -307,7 +380,6 @@ TEST_CASE ("ArmnnDelegateStringParsingOptionReduceFp32ToFp16")
     }
 }
 
-
 }
 
 } // namespace armnnDelegate
diff --git a/delegate/src/test/DelegateOptionsTestHelper.hpp b/delegate/src/test/DelegateOptionsTestHelper.hpp
index 6e0cc3154c..87bf0d6c3d 100644
--- a/delegate/src/test/DelegateOptionsTestHelper.hpp
+++ b/delegate/src/test/DelegateOptionsTestHelper.hpp
@@ -148,6 +148,77 @@ std::vector<char> CreateAddDivTfLiteModel(tflite::TensorType tensorType,
                              flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
 }
 
+std::vector<char> CreateCeilTfLiteModel(tflite::TensorType tensorType,
+                                        const std::vector <int32_t>& tensorShape,
+                                        float quantScale = 1.0f,
+                                        int quantOffset = 0)
+{
+    using namespace tflite;
+    flatbuffers::FlatBufferBuilder flatBufferBuilder;
+
+    std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
+    buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
+
+    auto quantizationParameters =
+        CreateQuantizationParameters(flatBufferBuilder,
+                                     0,
+                                     0,
+                                     flatBufferBuilder.CreateVector<float>({quantScale}),
+                                     flatBufferBuilder.CreateVector<int64_t>({quantOffset}));
+
+    std::array<flatbuffers::Offset<Tensor>, 2> tensors;
+    tensors[0] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(),
+                                                                      tensorShape.size()),
+                              tensorType,
+                              0,
+                              flatBufferBuilder.CreateString("input"),
+                              quantizationParameters);
+    tensors[1] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(),
+                                                                      tensorShape.size()),
+                              tensorType,
+                              0,
+                              flatBufferBuilder.CreateString("output"),
+                              quantizationParameters);
+
+    const std::vector<int32_t> operatorInputs({0});
+    const std::vector<int32_t> operatorOutputs({1});
+
+    flatbuffers::Offset<Operator> ceilOperator =
+        CreateOperator(flatBufferBuilder,
+                       0,
+                       flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                       BuiltinOptions_NONE);
+
+    flatbuffers::Offset<flatbuffers::String> modelDescription =
+        flatBufferBuilder.CreateString("ArmnnDelegate: CEIL Operator Model");
+    flatbuffers::Offset<OperatorCode> operatorCode =
+        CreateOperatorCode(flatBufferBuilder, tflite::BuiltinOperator_CEIL);
+
+    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(&ceilOperator, 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 ReduceFp32ToBf16TestImpl()
 {
     using namespace tflite;
@@ -295,4 +366,63 @@ void DelegateOptionTest(tflite::TensorType tensorType,
     armnnDelegateInterpreter.reset(nullptr);
 }
 
+template <typename T>
+void DelegateOptionNoFallbackTest(tflite::TensorType tensorType,
+                                  const std::vector<armnn::BackendId>& backends,
+                                  std::vector<int32_t>& tensorShape,
+                                  std::vector<T>& inputValues,
+                                  std::vector<T>& expectedOutputValues,
+                                  const armnnDelegate::DelegateOptions& delegateOptions,
+                                  float quantScale = 1.0f,
+                                  int quantOffset  = 0)
+{
+    using namespace tflite;
+    std::vector<char> modelBuffer = CreateCeilTfLiteModel(tensorType,
+                                                          tensorShape,
+                                                          quantScale,
+                                                          quantOffset);
+
+    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
+    std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
+        theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
+                         armnnDelegate::TfLiteArmnnDelegateDelete);
+    CHECK(theArmnnDelegate != nullptr);
+    // Modify armnnDelegateInterpreter to use armnnDelegate
+    try
+    {
+        armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get());
+    }
+    catch (const armnn::Exception& e)
+    {
+        // Forward the exception message to std::cout
+        std::cout << e.what() << std::endl;
+    }
+
+    // Set input data
+    armnnDelegate::FillInput(tfLiteInterpreter, 0, inputValues);
+    armnnDelegate::FillInput(armnnDelegateInterpreter, 0, inputValues);
+
+    // Run EnqueueWorkload
+    CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
+    CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
+
+    armnnDelegate::CompareOutputData<T>(tfLiteInterpreter, armnnDelegateInterpreter, tensorShape, expectedOutputValues);
+
+    armnnDelegateInterpreter.reset(nullptr);
+}
+
 } // anonymous namespace
\ No newline at end of file