| // |
| // Copyright © 2023 Arm Ltd and Contributors. All rights reserved. |
| // SPDX-License-Identifier: MIT |
| // |
| |
| #include "FileComparisonExecutor.hpp" |
| #include <NetworkExecutionUtils/NetworkExecutionUtils.hpp> |
| #include <algorithm> |
| #include <ghc/filesystem.hpp> |
| #include <iterator> |
| |
| using namespace armnn; |
| |
| /** |
| * Given a buffer in the expected format. Extract from it the tensor name, tensor type as strings and return an |
| * index pointing to the start of the data section. |
| * |
| * @param buffer data to be parsed. |
| * @param tensorName the name of the tensor extracted from the header. |
| * @param tensorType the type of the tensor extracted from the header. |
| * @return index pointing to the start of the data in the buffer. |
| */ |
| unsigned int ExtractHeader(const std::vector<char>& buffer, std::string& tensorName, DataType& tensorType) |
| { |
| auto isColon = [](char c) { return c == ':'; }; |
| auto isComma = [](char c) { return c == ','; }; |
| |
| // Find the "," separator marks the end of the tensor name. |
| auto firstComma = std::find_if(buffer.begin(), buffer.end(), isComma); |
| if (firstComma == buffer.end()) |
| { |
| throw ParseException("Unable to read tensor name from file."); |
| } |
| tensorName.assign(buffer.begin(), firstComma); |
| |
| // The next colon marks the end of the data type string. |
| auto endOfHeader = std::find_if(firstComma, buffer.end(), isColon); |
| if (firstComma == buffer.end()) |
| { |
| throw ParseException("Unable to read tensor type from file."); |
| } |
| std::string type(++firstComma, endOfHeader); |
| // Remove any leading or trailing whitespace. |
| type.erase(remove_if(type.begin(), type.end(), isspace), type.end()); |
| if (type == "Float16") |
| { |
| tensorType = DataType::Float16; |
| } |
| else if (type == "Float32") |
| { |
| tensorType = DataType::Float32; |
| } |
| else if (type == "QAsymmU8") |
| { |
| tensorType = DataType::QAsymmU8; |
| } |
| else if (type == "Signed32") |
| { |
| tensorType = DataType::Signed32; |
| } |
| else if (type == "Boolean") |
| { |
| tensorType = DataType::Boolean; |
| } |
| else if (type == "QSymmS16") |
| { |
| tensorType = DataType::QSymmS16; |
| } |
| else if (type == "QSymmS8") |
| { |
| tensorType = DataType::QSymmS8; |
| } |
| else if (type == "QAsymmS8") |
| { |
| tensorType = DataType::QAsymmS8; |
| } |
| else if (type == "BFloat16") |
| { |
| tensorType = DataType::BFloat16; |
| } |
| else if (type == "Signed64") |
| { |
| tensorType = DataType::Signed64; |
| } |
| else |
| { |
| throw ParseException("Invalid data type in header."); |
| } |
| // Remember to move the iterator past the colon. |
| endOfHeader++; |
| // Breaking this into two parts to avoid an awkward gcc compiler problem. |
| auto dataStart = endOfHeader - buffer.begin(); |
| return static_cast<unsigned int>(dataStart); |
| } |
| |
| /** |
| * Extract the data from the file and return as a typed vector of elements. |
| * |
| * @param buffer data to be parsed. |
| * @param dataStart Index into the vector where the tensor data starts. |
| * @param tensorType the type of the tensor extracted from the header. |
| */ |
| template <typename T> |
| void ReadData(const std::vector<char>& buffer, |
| const unsigned int dataStart, |
| const DataType& tensorType, |
| std::vector<T>& results) |
| { |
| unsigned int index = dataStart; |
| while (index < buffer.size()) |
| { |
| std::string elementString; |
| // Extract into a string until the next space. |
| while (index < buffer.size() && buffer[index] != ' ') |
| { |
| elementString.push_back(buffer[index]); |
| index++; |
| } |
| if (!elementString.empty()) |
| { |
| switch (tensorType) |
| { |
| case DataType::Float32: { |
| results.push_back(static_cast<T>(std::stof(elementString))); |
| break; |
| } |
| |
| case DataType::Signed32: { |
| results.push_back(static_cast<T>(std::stoi(elementString))); |
| break; |
| } |
| case DataType::QSymmS8: |
| case DataType::QAsymmS8: { |
| results.push_back(static_cast<T>(elementString[0])); |
| break; |
| } |
| case DataType::QAsymmU8: { |
| results.push_back(static_cast<T>(elementString[0])); |
| break; |
| } |
| case DataType::Float16: |
| case DataType::QSymmS16: |
| case DataType::BFloat16: |
| case DataType::Boolean: |
| case DataType::Signed64: |
| default: { |
| LogAndThrow("Unsupported DataType"); |
| } |
| } |
| // Finally, skip the space we know is there. |
| index++; |
| } |
| else |
| { |
| if (index < buffer.size()) |
| { |
| index++; |
| } |
| } |
| } |
| } |
| |
| /** |
| * Open the given file and read the data out of it to construct a Tensor. This could throw FileNotFoundException |
| * or InvalidArgumentException |
| * |
| * @param fileName the file to be read. |
| * @return a populated tensor. |
| */ |
| Tensor ReadTensorFromFile(const std::string fileName) |
| { |
| if (!ghc::filesystem::exists(fileName)) |
| { |
| throw FileNotFoundException("The file \"" + fileName + "\" could not be found."); |
| } |
| // The format we are reading in is based on NetworkExecutionUtils::WriteToFile. This could potentially |
| // be an enormous tensor. We'll limit what we can read in to 1Mb. |
| std::uintmax_t maxFileSize = 1048576; |
| std::uintmax_t fileSize = ghc::filesystem::file_size(fileName); |
| if (fileSize > maxFileSize) |
| { |
| throw InvalidArgumentException("The file \"" + fileName + "\" exceeds max size of 1 Mb."); |
| } |
| |
| // We'll read the entire file into one buffer. |
| std::ifstream file(fileName, std::ios::binary); |
| std::vector<char> buffer(fileSize); |
| if (file.read(buffer.data(), static_cast<std::streamsize>(fileSize))) |
| { |
| std::string tensorName; |
| DataType tensorType; |
| unsigned int tensorDataStart = ExtractHeader(buffer, tensorName, tensorType); |
| switch (tensorType) |
| { |
| case DataType::Float32: { |
| std::vector<float> floatVector; |
| ReadData(buffer, tensorDataStart, tensorType, floatVector); |
| TensorInfo info({ static_cast<unsigned int>(floatVector.size()), 1, 1, 1 }, DataType::Float32); |
| float* floats = new float[floatVector.size()]; |
| memcpy(floats, floatVector.data(), (floatVector.size() * sizeof(float))); |
| return Tensor(info, floats); |
| } |
| case DataType::Signed32: { |
| std::vector<int> intVector; |
| ReadData(buffer, tensorDataStart, tensorType, intVector); |
| TensorInfo info({ static_cast<unsigned int>(intVector.size()), 1, 1, 1 }, DataType::Signed32); |
| int* ints = new int[intVector.size()]; |
| memcpy(ints, intVector.data(), (intVector.size() * sizeof(float))); |
| return Tensor(info, ints); |
| } |
| case DataType::QSymmS8: { |
| std::vector<int8_t> intVector; |
| ReadData(buffer, tensorDataStart, tensorType, intVector); |
| TensorInfo info({ static_cast<unsigned int>(intVector.size()), 1, 1, 1 }, DataType::QSymmS8); |
| int8_t* ints = new int8_t[intVector.size()]; |
| memcpy(ints, intVector.data(), (intVector.size() * sizeof(float))); |
| return Tensor(info, ints); |
| } |
| case DataType::QAsymmS8: { |
| std::vector<int8_t> intVector; |
| ReadData(buffer, tensorDataStart, tensorType, intVector); |
| TensorInfo info({ static_cast<unsigned int>(intVector.size()), 1, 1, 1 }, DataType::QAsymmS8); |
| int8_t* ints = new int8_t[intVector.size()]; |
| memcpy(ints, intVector.data(), (intVector.size() * sizeof(float))); |
| return Tensor(info, ints); |
| } |
| case DataType::QAsymmU8: { |
| std::vector<uint8_t> intVector; |
| ReadData(buffer, tensorDataStart, tensorType, intVector); |
| TensorInfo info({ static_cast<unsigned int>(intVector.size()), 1, 1, 1 }, DataType::QAsymmU8); |
| uint8_t* ints = new uint8_t[intVector.size()]; |
| memcpy(ints, intVector.data(), (intVector.size() * sizeof(float))); |
| return Tensor(info, ints); |
| } |
| default: |
| throw InvalidArgumentException("The tensor data could not be read from \"" + fileName + "\""); |
| } |
| } |
| else |
| { |
| throw ParseException("Filed to read the contents of \"" + fileName + "\""); |
| } |
| |
| Tensor result; |
| return result; |
| } |
| |
| FileComparisonExecutor::FileComparisonExecutor(const ExecuteNetworkParams& params) |
| : m_Params(params) |
| {} |
| |
| std::vector<const void*> FileComparisonExecutor::Execute() |
| { |
| std::string filesToCompare = this->m_Params.m_ComparisonFile; |
| if (filesToCompare.empty()) |
| { |
| throw InvalidArgumentException("The file(s) to compare was not set."); |
| } |
| // filesToCompare is one or more files containing output tensors. Iterate and read in the tensors. |
| // We'll assume the string follows the same comma seperated format as write-outputs-to-file. |
| std::stringstream ss(filesToCompare); |
| std::vector<std::string> fileNames; |
| std::string errorString; |
| while (ss.good()) |
| { |
| std::string substr; |
| getline(ss, substr, ','); |
| // Check the file exist. |
| if (!ghc::filesystem::exists(substr)) |
| { |
| errorString += substr + " "; |
| } |
| else |
| { |
| fileNames.push_back(substr); |
| } |
| } |
| if (!errorString.empty()) |
| { |
| throw FileNotFoundException("The following file(s) to compare could not be found: " + errorString); |
| } |
| // Read in the tensors into m_OutputTensorsVec |
| OutputTensors outputs; |
| std::vector<const void*> results; |
| for (auto file : fileNames) |
| { |
| Tensor t = ReadTensorFromFile(file); |
| outputs.push_back({ 0, Tensor(t.GetInfo(), t.GetMemoryArea()) }); |
| results.push_back(t.GetMemoryArea()); |
| } |
| m_OutputTensorsVec.push_back(outputs); |
| return results; |
| } |
| |
| void FileComparisonExecutor::PrintNetworkInfo() |
| { |
| std::cout << "Not implemented in this class." << std::endl; |
| } |
| |
| void FileComparisonExecutor::CompareAndPrintResult(std::vector<const void*> otherOutput) |
| { |
| unsigned int index = 0; |
| std::string typeString; |
| for (const auto& outputTensors : m_OutputTensorsVec) |
| { |
| for (const auto& outputTensor : outputTensors) |
| { |
| size_t size = outputTensor.second.GetNumBytes(); |
| double result = ComputeByteLevelRMSE(outputTensor.second.GetMemoryArea(), otherOutput[index++], size); |
| std::cout << "Byte level root mean square error: " << result << "\n"; |
| } |
| } |
| } |
| |
| FileComparisonExecutor::~FileComparisonExecutor() |
| { |
| // If there are tensors defined in m_OutputTensorsVec we need to clean up their memory usage. |
| for (OutputTensors opTensor : m_OutputTensorsVec) |
| { |
| for (std::pair<LayerBindingId, class Tensor> pair : opTensor) |
| { |
| Tensor t = pair.second; |
| // Based on the tensor type and size recover the memory. |
| switch (t.GetDataType()) |
| { |
| case DataType::Float32: |
| delete[] static_cast<float*>(t.GetMemoryArea()); |
| break; |
| case DataType::Signed32: |
| delete[] static_cast<int*>(t.GetMemoryArea()); |
| break; |
| case DataType::QSymmS8: |
| delete[] static_cast<int8_t*>(t.GetMemoryArea()); |
| break; |
| case DataType::QAsymmS8: |
| delete[] static_cast<int8_t*>(t.GetMemoryArea()); |
| break; |
| case DataType::QAsymmU8: |
| delete[] static_cast<uint8_t*>(t.GetMemoryArea()); |
| break; |
| default: |
| std::cout << "The data type wasn't created in ReadTensorFromFile" << std::endl; |
| } |
| } |
| } |
| |
| } |