// // Copyright © 2020, 2023-2024 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // #pragma once #include "TestUtils.hpp" #include #include #include namespace { std::vector CreateQuantizationTfLiteModel(tflite::BuiltinOperator quantizationOperatorCode, tflite::TensorType inputTensorType, tflite::TensorType outputTensorType, const std::vector & inputTensorShape, const std::vector & outputTensorShape, float quantScale = 1.0f, int quantOffset = 0) { using namespace tflite; flatbuffers::FlatBufferBuilder flatBufferBuilder; std::vector> buffers; buffers.push_back(CreateBuffer(flatBufferBuilder)); buffers.push_back(CreateBuffer(flatBufferBuilder)); buffers.push_back(CreateBuffer(flatBufferBuilder)); auto quantizationParameters = CreateQuantizationParameters(flatBufferBuilder, 0, 0, flatBufferBuilder.CreateVector({ quantScale }), flatBufferBuilder.CreateVector({ quantOffset }), QuantizationDetails_CustomQuantization); std::array, 2> tensors; tensors[0] = CreateTensor(flatBufferBuilder, flatBufferBuilder.CreateVector(inputTensorShape.data(), inputTensorShape.size()), inputTensorType, 1, flatBufferBuilder.CreateString("input"), quantizationParameters); tensors[1] = CreateTensor(flatBufferBuilder, flatBufferBuilder.CreateVector(outputTensorShape.data(), outputTensorShape.size()), outputTensorType, 2, flatBufferBuilder.CreateString("output"), quantizationParameters); // create operator tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_NONE; flatbuffers::Offset 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 operatorInputs{0}; const std::vector operatorOutputs{1}; flatbuffers::Offset quantizationOperator = CreateOperator(flatBufferBuilder, 0, flatBufferBuilder.CreateVector(operatorInputs.data(), operatorInputs.size()), flatBufferBuilder.CreateVector(operatorOutputs.data(), operatorOutputs.size()), operatorBuiltinOptionsType, operatorBuiltinOptions); const std::vector subgraphInputs{0}; const std::vector subgraphOutputs{1}; flatbuffers::Offset subgraph = CreateSubGraph(flatBufferBuilder, flatBufferBuilder.CreateVector(tensors.data(), tensors.size()), flatBufferBuilder.CreateVector(subgraphInputs.data(), subgraphInputs.size()), flatBufferBuilder.CreateVector(subgraphOutputs.data(), subgraphOutputs.size()), flatBufferBuilder.CreateVector(&quantizationOperator, 1)); flatbuffers::Offset modelDescription = flatBufferBuilder.CreateString("ArmnnDelegate: Quantization Operator Model"); flatbuffers::Offset operatorCode = CreateOperatorCode(flatBufferBuilder, quantizationOperatorCode); flatbuffers::Offset 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(flatBufferBuilder.GetBufferPointer(), flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize()); } template void QuantizationTest(tflite::BuiltinOperator quantizeOperatorCode, tflite::TensorType inputTensorType, tflite::TensorType outputTensorType, std::vector& inputShape, std::vector& outputShape, std::vector& inputValues, std::vector& expectedOutputValues, const std::vector& backends = {}, float quantScale = 1.0f, int quantOffset = 0) { using namespace delegateTestInterpreter; std::vector 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 tfLiteOutputValues = tfLiteInterpreter.GetOutputResult(0); std::vector 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 armnnOutputValues = armnnInterpreter.GetOutputResult(0); std::vector armnnOutputShape = armnnInterpreter.GetOutputShape(0); armnnDelegate::CompareOutputData(tfLiteOutputValues, armnnOutputValues, expectedOutputValues); armnnDelegate::CompareOutputShape(tfLiteOutputShape, armnnOutputShape, outputShape); tfLiteInterpreter.Cleanup(); armnnInterpreter.Cleanup(); } } // anonymous namespace