IVGCVSW-5400 'TfLiteDelegate: FLOOR operator support'

* Added FLOOR operator support to Arm NN TfLiteDelegate

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I986ce8c5a825f509e0f8b3d257fd5b60834c322f

5 files changed, 285 insertions, 8 deletions

diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt
index 74390c8a93..11bff48a66 100644
--- a/delegate/CMakeLists.txt
+++ b/delegate/CMakeLists.txt
@@ -149,6 +149,8 @@ if(BUILD_UNIT_TESTS)
         src/test/ReshapeTest.cpp
         src/test/ResizeTest.cpp
         src/test/ResizeTestHelper.hpp
+        src/test/RoundTest.cpp
+        src/test/RoundTestHelper.hpp
         src/test/SoftmaxTest.cpp
         src/test/SoftmaxTestHelper.hpp
         src/test/SpaceDepthTest.cpp
diff --git a/delegate/src/Round.hpp b/delegate/src/Round.hpp
index 3335d0b337..1677607571 100644
--- a/delegate/src/Round.hpp
+++ b/delegate/src/Round.hpp
@@ -5,8 +5,6 @@
 
 #pragma once
 
-#include <armnn/utility/IgnoreUnused.hpp>
-
 #include <tensorflow/lite/builtin_ops.h>
 #include <tensorflow/lite/c/builtin_op_data.h>
 #include <tensorflow/lite/c/common.h>
@@ -21,13 +19,55 @@ TfLiteStatus VisitFloorOperator(DelegateData& delegateData,
                                 int nodeIndex,
                                 int32_t operatorCode)
 {
-    armnn::IgnoreUnused(delegateData,
-                        tfLiteContext,
-                        tfLiteNode,
-                        nodeIndex,
-                        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,
+                                   IsFloorSupported,
+                                   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, VisitFloorOperator 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 Floor layer
+    armnn::IConnectableLayer* layer = delegateData.m_Network->AddFloorLayer();
+    ARMNN_ASSERT(layer != nullptr);
+
+    armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
+    outputSlot.SetTensorInfo(outputTensorInfo);
 
-    return kTfLiteError;
+    // Connect
+    return Connect(layer, tfLiteNode, delegateData);
 }
 
 } // namespace armnnDelegate
diff --git a/delegate/src/test/RoundTest.cpp b/delegate/src/test/RoundTest.cpp
new file mode 100644
index 0000000000..9d323f3700
--- /dev/null
+++ b/delegate/src/test/RoundTest.cpp
@@ -0,0 +1,72 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "RoundTestHelper.hpp"
+
+#include <armnn_delegate.hpp>
+
+#include <flatbuffers/flatbuffers.h>
+#include <tensorflow/lite/schema/schema_generated.h>
+
+#include <doctest/doctest.h>
+
+namespace armnnDelegate
+{
+
+void FloorFp32Test(std::vector<armnn::BackendId>& backends)
+{
+    std::vector<int32_t> inputShape  {1, 3, 2, 3};
+    std::vector<int32_t> outputShape {1, 3, 2, 3};
+
+    std::vector<float> inputValues { -37.5f, -15.2f, -8.76f, -2.0f, -1.5f, -1.3f, -0.5f, -0.4f, 0.0f,
+                                     1.0f, 0.4f, 0.5f, 1.3f, 1.5f, 2.0f, 8.76f, 15.2f, 37.5f };
+
+    std::vector<float> expectedOutputValues { -38.0f, -16.0f, -9.0f, -2.0f, -2.0f, -2.0f, -1.0f, -1.0f, 0.0f,
+                                              1.0f, 0.0f, 0.0f, 1.0f, 1.0f, 2.0f, 8.0f, 15.0f, 37.0f };
+
+    RoundTest<float>(tflite::BuiltinOperator_FLOOR,
+                     ::tflite::TensorType_FLOAT32,
+                     backends,
+                     inputShape,
+                     inputValues,
+                     expectedOutputValues);
+}
+
+// FLOOR Test Suite
+TEST_SUITE("FLOOR_CpuRefTests")
+{
+
+TEST_CASE ("FLOOR_Fp32_CpuRef_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
+    FloorFp32Test(backends);
+}
+
+}
+
+TEST_SUITE("FLOOR_CpuAccTests")
+{
+
+TEST_CASE ("FLOOR_Fp32_CpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::CpuAcc};
+    FloorFp32Test(backends);
+}
+
+}
+
+TEST_SUITE("FLOOR_GpuAccTests")
+{
+
+TEST_CASE ("FLOOR_Fp32_GpuAcc_Test")
+{
+    std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
+    FloorFp32Test(backends);
+}
+
+}
+// End of FLOOR Test Suite
+
+} // namespace armnnDelegate
\ No newline at end of file
diff --git a/delegate/src/test/RoundTestHelper.hpp b/delegate/src/test/RoundTestHelper.hpp
new file mode 100644
index 0000000000..3a35ee0764
--- /dev/null
+++ b/delegate/src/test/RoundTestHelper.hpp
@@ -0,0 +1,161 @@
+//
+// 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> CreateRoundTfLiteModel(tflite::BuiltinOperator roundOperatorCode,
+                                         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> roundOperator;
+    flatbuffers::Offset<flatbuffers::String> modelDescription;
+    flatbuffers::Offset<OperatorCode> operatorCode;
+
+    switch (roundOperatorCode)
+    {
+        case tflite::BuiltinOperator_FLOOR:
+        default:
+            roundOperator =
+                CreateOperator(flatBufferBuilder,
+                               0,
+                               flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
+                               flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()));
+                modelDescription = flatBufferBuilder.CreateString("ArmnnDelegate: Floor Operator Model");
+                operatorCode = CreateOperatorCode(flatBufferBuilder, tflite::BuiltinOperator_FLOOR);
+            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(&roundOperator, 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>
+void RoundTest(tflite::BuiltinOperator roundOperatorCode,
+               tflite::TensorType tensorType,
+               std::vector<armnn::BackendId>& backends,
+               std::vector<int32_t>& shape,
+               std::vector<T>& inputValues,
+               std::vector<T>& expectedOutputValues,
+               float quantScale = 1.0f,
+               int quantOffset = 0)
+{
+    using namespace tflite;
+    std::vector<char> modelBuffer = CreateRoundTfLiteModel(roundOperatorCode,
+                                                           tensorType,
+                                                           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<T>(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 7c7763a923..73d869041e 100644
--- a/docs/01_03_delegate.dox
+++ b/docs/01_03_delegate.dox
@@ -63,6 +63,8 @@ The Arm NN SDK TensorFlow Lite delegate currently supports the following operato
 
 - FULLY_CONNECTED, Supported Fused Activation: RELU , RELU6 , TANH, NONE
 
+- FLOOR
+
 - GATHER
 
 - GREATER