delegate/test/QuantizationTestHelper.hpp

//
// Copyright © 2020, 2023-2024 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#pragma once

#include "TestUtils.hpp"

#include <armnn_delegate.hpp>
#include <DelegateTestInterpreter.hpp>

#include <tensorflow/lite/version.h>

namespace
{

std::vector<char> CreateQuantizationTfLiteModel(tflite::BuiltinOperator quantizationOperatorCode,
                                                tflite::TensorType inputTensorType,
                                                tflite::TensorType outputTensorType,
                                                const std::vector <int32_t>& inputTensorShape,
                                                const std::vector <int32_t>& outputTensorShape,
                                                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));
    buffers.push_back(CreateBuffer(flatBufferBuilder));
    buffers.push_back(CreateBuffer(flatBufferBuilder));


    auto quantizationParameters =
            CreateQuantizationParameters(flatBufferBuilder,
                                         0,
                                         0,
                                         flatBufferBuilder.CreateVector<float>({ quantScale }),
                                         flatBufferBuilder.CreateVector<int64_t>({ quantOffset }),
                                         QuantizationDetails_CustomQuantization);

    std::array<flatbuffers::Offset<Tensor>, 2> tensors;
    tensors[0] = CreateTensor(flatBufferBuilder,
                              flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(),
                                                                      inputTensorShape.size()),
                              inputTensorType,
                              1,
                              flatBufferBuilder.CreateString("input"),
                              quantizationParameters);
    tensors[1] = CreateTensor(flatBufferBuilder,
                              flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(),
                                                                      outputTensorShape.size()),
                              outputTensorType,
                              2,
                              flatBufferBuilder.CreateString("output"),
                              quantizationParameters);

    // create operator
    tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_NONE;
    flatbuffers::Offset<void> operatorBuiltinOptions = 0;
    switch (quantizationOperatorCode)
    {
        case BuiltinOperator_QUANTIZE:
        {
            operatorBuiltinOptionsType = BuiltinOptions_QuantizeOptions;
            operatorBuiltinOptions = CreateQuantizeOptions(flatBufferBuilder).Union();
            break;
        }
        case BuiltinOperator_DEQUANTIZE:
        {
            operatorBuiltinOptionsType = BuiltinOptions_DequantizeOptions;
            operatorBuiltinOptions = CreateDequantizeOptions(flatBufferBuilder).Union();
            break;
        }
        default:
            break;
    }

    const std::vector<int32_t> operatorInputs{0};
    const std::vector<int32_t> operatorOutputs{1};
    flatbuffers::Offset <Operator> quantizationOperator =
            CreateOperator(flatBufferBuilder,
                           0,
                           flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
                           flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
                           operatorBuiltinOptionsType,
                           operatorBuiltinOptions);

    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(&quantizationOperator, 1));

    flatbuffers::Offset <flatbuffers::String> modelDescription =
            flatBufferBuilder.CreateString("ArmnnDelegate: Quantization Operator Model");
    flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, quantizationOperatorCode);

    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, armnnDelegate::FILE_IDENTIFIER);

    return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
                             flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
}

template <typename InputT, typename OutputT>
void QuantizationTest(tflite::BuiltinOperator quantizeOperatorCode,
                      tflite::TensorType inputTensorType,
                      tflite::TensorType outputTensorType,
                      std::vector<int32_t>& inputShape,
                      std::vector<int32_t>& outputShape,
                      std::vector<InputT>&  inputValues,
                      std::vector<OutputT>& expectedOutputValues,
                      const std::vector<armnn::BackendId>& backends = {},
                      float quantScale = 1.0f,
                      int quantOffset  = 0)
{
    using namespace delegateTestInterpreter;
    std::vector<char> modelBuffer = CreateQuantizationTfLiteModel(quantizeOperatorCode,
                                                                  inputTensorType,
                                                                  outputTensorType,
                                                                  inputShape,
                                                                  outputShape,
                                                                  quantScale,
                                                                  quantOffset);

    // Setup interpreter with just TFLite Runtime.
    auto tfLiteInterpreter = DelegateTestInterpreter(modelBuffer);
    CHECK(tfLiteInterpreter.AllocateTensors() == kTfLiteOk);
    CHECK(tfLiteInterpreter.FillInputTensor(inputValues, 0) == kTfLiteOk);
    CHECK(tfLiteInterpreter.Invoke() == kTfLiteOk);
    std::vector<OutputT> tfLiteOutputValues = tfLiteInterpreter.GetOutputResult<OutputT>(0);
    std::vector<int32_t> tfLiteOutputShape  = tfLiteInterpreter.GetOutputShape(0);

    // Setup interpreter with Arm NN Delegate applied.
    auto armnnInterpreter = DelegateTestInterpreter(modelBuffer, CaptureAvailableBackends(backends));
    CHECK(armnnInterpreter.AllocateTensors() == kTfLiteOk);
    CHECK(armnnInterpreter.FillInputTensor(inputValues, 0) == kTfLiteOk);
    CHECK(armnnInterpreter.Invoke() == kTfLiteOk);
    std::vector<OutputT> armnnOutputValues = armnnInterpreter.GetOutputResult<OutputT>(0);
    std::vector<int32_t> armnnOutputShape  = armnnInterpreter.GetOutputShape(0);

    armnnDelegate::CompareOutputData<OutputT>(tfLiteOutputValues, armnnOutputValues, expectedOutputValues);
    armnnDelegate::CompareOutputShape(tfLiteOutputShape, armnnOutputShape, outputShape);

    tfLiteInterpreter.Cleanup();
    armnnInterpreter.Cleanup();
}

} // anonymous namespace