IVGCVSW-5430 'Add CAST Operator Support to Delegate'

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I6da711950b8e7d3c0d5cbd443e91eb36700ac4c8

6 files changed, 332 insertions, 0 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index d72089ca85..fea07b37ec 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -118,6 +118,8 @@ if(BUILD_UNIT_TESTS)
         src/test/ArmnnDelegateTest.cpp
         src/test/BatchSpaceTest.cpp
         src/test/BatchSpaceTestHelper.hpp
+        src/test/CastTest.cpp
+        src/test/CastTestHelper.hpp
         src/test/ComparisonTest.cpp
         src/test/ComparisonTestHelper.hpp
         src/test/ControlTest.cpp
diff --git a/delegate/src/Redefine.hpp b/delegate/src/Redefine.hpp
index 3df26cacc3..766e600c7c 100644
--- a/delegate/src/Redefine.hpp
+++ b/delegate/src/Redefine.hpp
@@ -18,6 +18,64 @@
 namespace armnnDelegate
 {
 
+TfLiteStatus VisitCastOperator(DelegateData& delegateData,
+                               TfLiteContext* tfLiteContext,
+                               TfLiteNode* tfLiteNode,
+                               int nodeIndex,
+                               int32_t operatorCode)
+{
+    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 (!IsValid(tfLiteContext, tfLiteInputTensor, operatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
+    if (!IsValid(tfLiteContext, tfLiteOutputTensor, operatorCode, nodeIndex))
+    {
+        return kTfLiteError;
+    }
+
+    const armnn::TensorInfo& inputTensorInfo  = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
+    const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
+
+    bool isSupported = false;
+    auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported)
+    {
+        FORWARD_LAYER_SUPPORT_FUNC(__func__,
+                                   tfLiteContext,
+                                   IsCastSupported,
+                                   delegateData.m_Backends,
+                                   isSupported,
+                                   inputTensorInfo,
+                                   outInfo);
+    };
+
+    // If the m_Network is a nullptr, this signals that a prerequisite TfLite callback is required to clarify the
+    // support for the operator
+    // If supported, VisitCastOperator will be called again to add the layer to the network as seen further below
+    if (!delegateData.m_Network)
+    {
+        validateFunc(outputTensorInfo, isSupported);
+        return isSupported ? kTfLiteOk : kTfLiteError;
+    }
+
+    // Add a Cast layer
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddCastLayer();
+    ARMNN_ASSERT(layer != nullptr);
+
+    armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
+
+    // Connect
+    return Connect(layer, tfLiteNode, delegateData);
+}
+
+
 TfLiteStatus CreateOutputTensorShape(const armnn::TensorInfo& inputTensorInfo,
                                      const std::vector<int32_t>& targetShape,
                                      armnn::ReshapeDescriptor& reshapeDesc)
diff --git a/delegate/src/armnn_delegate.cpp b/delegate/src/armnn_delegate.cpp
index 2b07fc7098..4c625a9867 100644
--- a/delegate/src/armnn_delegate.cpp
+++ b/delegate/src/armnn_delegate.cpp
@@ -486,6 +486,12 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData,
                                                tfLiteNode,
                                                nodeIndex,
                                                kTfLiteBuiltinBatchToSpaceNd);
+        case kTfLiteBuiltinCast:
+            return VisitCastOperator(delegateData,
+                                     tfLiteContext,
+                                     tfLiteNode,
+                                     nodeIndex,
+                                     kTfLiteBuiltinCast);
         case kTfLiteBuiltinConcatenation:
             return VisitControlOperator(delegateData,
                                         tfLiteContext,
diff --git a/delegate/src/test/CastTest.cpp b/delegate/src/test/CastTest.cpp
new file mode 100644
index 0000000000..623c045247
--- /dev/null
+++ b/delegate/src/test/CastTest.cpp
@@ -0,0 +1,107 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "CastTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+void CastUint8ToFp32Test(std::vector<armnn::BackendId>& backends)
+{
+    std::vector<int32_t> inputShape  {1, 3, 2, 3};
+
+    std::vector<uint8_t> inputValues { 1, 3, 1, 3, 1, 3, 1, 3, 1,
+                                        3, 1, 3, 1, 2, 1, 3, 1, 3 };
+
+    std::vector<float> expectedOutputValues { 1.0f, 3.0f, 1.0f, 3.0f, 1.0f, 3.0f, 1.0f, 3.0f, 1.0f,
+                                              3.0f, 1.0f, 3.0f, 1.0f, 2.0f, 1.0f, 3.0f, 1.0f, 3.0f };
+
+    CastTest<uint8_t, float>(::tflite::TensorType_UINT8,
+                             ::tflite::TensorType_FLOAT32,
+                             backends,
+                             inputShape,
+                             inputValues,
+                             expectedOutputValues);
+}
+
+void CastInt32ToFp32Test(std::vector<armnn::BackendId>& backends)
+{
+    std::vector<int32_t> inputShape  {1, 3, 2, 3};
+
+    std::vector<int32_t> inputValues { -1, -3, -1, -3, -1, -3, -1, -3, 1,
+                                       3, 1, 3, 1, 2, 1, 3, 1, 3 };
+
+    std::vector<float> expectedOutputValues { -1.0f, -3.0f, -1.0f, -3.0f, -1.0f, -3.0f, -1.0f, -3.0f, 1.0f,
+                                              3.0f, 1.0f, 3.0f, 1.0f, 2.0f, 1.0f, 3.0f, 1.0f, 3.0f };
+
+    CastTest<int32_t, float>(::tflite::TensorType_INT32,
+                             ::tflite::TensorType_FLOAT32,
+                             backends,
+                             inputShape,
+                             inputValues,
+                             expectedOutputValues);
+}
+
+// CAST Test Suite
+TEST_SUITE("CAST_CpuRefTests")
+{
+
+TEST_CASE ("CAST_UINT8_TO_FP32_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
+    CastUint8ToFp32Test(backends);
+}
+
+TEST_CASE ("CAST_INT32_TO_FP32_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
+    CastInt32ToFp32Test(backends);
+}
+
+}
+
+TEST_SUITE("CAST_CpuAccTests")
+{
+
+TEST_CASE ("CAST_UINT8_TO_FP32_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::CpuAcc};
+    CastUint8ToFp32Test(backends);
+}
+
+TEST_CASE ("CAST_INT32_TO_FP32_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::CpuAcc};
+    CastInt32ToFp32Test(backends);
+}
+
+}
+
+TEST_SUITE("CAST_GpuAccTests")
+{
+
+TEST_CASE ("CAST_UINT8_TO_FP32_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
+    CastUint8ToFp32Test(backends);
+}
+
+TEST_CASE ("CAST_INT32_TO_FP32_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
+    CastInt32ToFp32Test(backends);
+}
+
+}
+// End of CAST Test Suite
+
+} // namespace armnnDelegate
\ No newline at end of file
diff --git a/delegate/src/test/CastTestHelper.hpp b/delegate/src/test/CastTestHelper.hpp
new file mode 100644
index 0000000000..6b1d5ee947
--- /dev/null
+++ b/delegate/src/test/CastTestHelper.hpp
@@ -0,0 +1,157 @@
+//
+// 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
+{
+std::vector<char> CreateCastTfLiteModel(tflite::TensorType inputTensorType,
+                                        tflite::TensorType outputTensorType,
+                                        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()),
+                              inputTensorType,
+                              0,
+                              flatBufferBuilder.CreateString("input"),
+                              quantizationParameters);
+    tensors[1] = CreateTensor(flatBufferBuilder,
+                              flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(),
+                                                                      tensorShape.size()),
+                              outputTensorType,
+                              0,
+                              flatBufferBuilder.CreateString("output"),
+                              quantizationParameters);
+
+    const std::vector<int32_t> operatorInputs({0});
+    const std::vector<int32_t> operatorOutputs({1});
+
+    flatbuffers::Offset<Operator> castOperator =
+        CreateOperator(flatBufferBuilder,
+                       0,
+                       flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                       flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
+                       BuiltinOptions_CastOptions,
+                       CreateCastOptions(flatBufferBuilder).Union());
+
+    flatbuffers::Offset<flatbuffers::String> modelDescription =
+        flatBufferBuilder.CreateString("ArmnnDelegate: CAST Operator Model");
+    flatbuffers::Offset<OperatorCode> operatorCode =
+        CreateOperatorCode(flatBufferBuilder, tflite::BuiltinOperator_CAST);
+
+    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(&castOperator, 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());
+}
+
+template<typename T, typename K>
+void CastTest(tflite::TensorType inputTensorType,
+              tflite::TensorType outputTensorType,
+              std::vector<armnn::BackendId>& backends,
+              std::vector<int32_t>& shape,
+              std::vector<T>& inputValues,
+              std::vector<K>& expectedOutputValues,
+              float quantScale = 1.0f,
+              int quantOffset = 0)
+{
+    using namespace tflite;
+    std::vector<char> modelBuffer = CreateCastTfLiteModel(inputTensorType,
+                                                          outputTensorType,
+                                                          shape,
+                                                          quantScale,
+                                                          quantOffset);
+
+    const Model* tfLiteModel = GetModel(modelBuffer.data());
+
+    // Create TfLite Interpreters
+    std::unique_ptr<Interpreter> armnnDelegate;
+    CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
+              (&armnnDelegate) == kTfLiteOk);
+    CHECK(armnnDelegate != nullptr);
+    CHECK(armnnDelegate->AllocateTensors() == kTfLiteOk);
+
+    std::unique_ptr<Interpreter> tfLiteDelegate;
+    CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
+              (&tfLiteDelegate) == kTfLiteOk);
+    CHECK(tfLiteDelegate != nullptr);
+    CHECK(tfLiteDelegate->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(armnnDelegate->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk);
+
+    // Set input data
+    armnnDelegate::FillInput<T>(tfLiteDelegate, 0, inputValues);
+    armnnDelegate::FillInput<T>(armnnDelegate, 0, inputValues);
+
+    // Run EnqueWorkload
+    CHECK(tfLiteDelegate->Invoke() == kTfLiteOk);
+    CHECK(armnnDelegate->Invoke() == kTfLiteOk);
+
+    // Compare output data
+    armnnDelegate::CompareOutputData<K>(tfLiteDelegate,
+                                        armnnDelegate,
+                                        shape,
+                                        expectedOutputValues,
+                                        0);
+
+    tfLiteDelegate.reset(nullptr);
+    armnnDelegate.reset(nullptr);
+}
+
+} // anonymous namespace
diff --git a/docs/01_03_delegate.dox b/docs/01_03_delegate.dox
index f6d8e7660d..a90883ba84 100644
--- a/docs/01_03_delegate.dox
+++ b/docs/01_03_delegate.dox
@@ -45,6 +45,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 - BATCH_TO_SPACE_ND
 
+- CAST
+
 - CONCATENATION, Supported Fused Activation: RELU , RELU6 , TANH, NONE
 
 - CONV_2D, Supported Fused Activation: RELU , RELU6 , TANH, NONE