benchmark.cpp

#include "benchmark.h"

static void write_compression_report(const std::vector<BaseBenchmarkRecord> &results, const std::string &reportFile)
{
    std::ofstream csvFile = std::ofstream(reportFile, std::ios::out);
    always_assert(csvFile.is_open());

    csvFile << std::fixed << std::setprecision(5);
    // CSV header.
    csvFile << "filename;subblock;pixel;width;height;compression;level;originalSize;compressedSize;compressedZ;compressionRatio;compressionRatioZ;compressionTime;compressionTimeZ" << std::endl;
    const char sep = ';';

    for (const BaseBenchmarkRecord &record : results)
    {
        record.write_to_stream(csvFile, sep);
    }
}

static void write_diff_report(const std::vector<DiffBenchmarkRecord> &results, const std::string &reportFile)
{
    std::ofstream csvFile = std::ofstream(reportFile, std::ios::out | std::ios::app);
    always_assert(csvFile.is_open());

    csvFile << std::fixed << std::setprecision(5);
    // CSV header.
    csvFile << "filename;channel;subblock;refSubblock;pixel;width;height;compression;level;diffType;bitsUsed;originalSize;compressedSize;compressedZ;compressionRatio;compressionRatioZ;compressionTime;compressionTimeZ" << std::endl;
    const char sep = ';';

    for (const DiffBenchmarkRecord &record : results)
    {
        record.write_to_stream(csvFile, sep);
    }
}

static void compression_thread_work(const ByteArray &data, CompressionMethod method, int compressionLevel, CompressionResult &result, const char *info)
{
    auto compResult = test_compression_method(data, method, compressionLevel);
    result = compResult;
    printf("Completed: %s\n", info);
}

static std::vector<BaseBenchmarkRecord> benchmark_continuos_compression_one_level(const ByteArray &data, const ByteArray &zOrderedData, int compressionLevel)
{
    // CompressionMethod_GZIP
    CompressionResult gzipResult = {};
    CompressionResult gzipZResult = {};

    // CompressionMethod_LZMA
    CompressionResult lzmaResult = {};
    CompressionResult lzmaZResult = {};

    // CompressionMethod_BZIP2
    CompressionResult bzip2Result = {};
    CompressionResult bzip2ZResult = {};

    std::vector<std::thread> workers;
    workers.resize(6);
    // We know that lzma is slowest, let's run all three in threads.
    workers[0] = std::thread(compression_thread_work, std::ref(data), CompressionMethod_GZIP, compressionLevel, std::ref(gzipResult), "Gzip normal order");
    workers[1] = std::thread(compression_thread_work, std::ref(zOrderedData), CompressionMethod_GZIP, compressionLevel, std::ref(gzipZResult), "Gzip Z order");

    workers[2] = std::thread(compression_thread_work, std::ref(data), CompressionMethod_LZMA, compressionLevel, std::ref(lzmaResult), "LZMA normal order");
    workers[3] = std::thread(compression_thread_work, std::ref(zOrderedData), CompressionMethod_LZMA, compressionLevel, std::ref(lzmaZResult), "LZMA Z order");

    workers[4] = std::thread(compression_thread_work, std::ref(data), CompressionMethod_BZIP2, compressionLevel, std::ref(bzip2Result), "Bzip2 normal order");
    workers[5] = std::thread(compression_thread_work, std::ref(zOrderedData), CompressionMethod_BZIP2, compressionLevel, std::ref(bzip2ZResult), "Bzip2 Z order");

    for (size_t i = 0; i < workers.size(); i++)
    {
        workers[i].join();
    }
    printf("All threads completed.\n");

    BaseBenchmarkRecord gzipRecord(gzipResult, gzipZResult);
    gzipRecord.compressionLevel = compressionLevel;
    gzipRecord.compressionMethod = GZIP_NAME;

    BaseBenchmarkRecord lzmaRecord(lzmaResult, lzmaZResult);
    lzmaRecord.compressionLevel = compressionLevel;
    lzmaRecord.compressionMethod = LZMA_NAME;

    BaseBenchmarkRecord bzip2Record(bzip2Result, bzip2ZResult);
    bzip2Record.compressionLevel = compressionLevel;
    bzip2Record.compressionMethod = BZIP2_NAME;

    std::vector<BaseBenchmarkRecord> results;
    results.resize(3);
    results[0] = gzipRecord;
    results[1] = lzmaRecord;
    results[2] = bzip2Record;
    return results;
}

void benchmark_continuos_compression(CziFile &cziFile, const std::string &reportFile, bool verbose, int level)
{
    always_assert(cziFile.subBlockDirectory.entries.size() > 0);
    const int minCompressionLevel = 1;
    const int maxCompressionLevel = 9;

    std::string fName = fs_wrapper::get_filename(cziFile.fileName);
    auto entry = cziFile.subBlockDirectory.entries[0];

    ByteArray data = cziFile.get_continuous_image_data(false);
    ByteArray zOrderData = cziFile.get_continuous_image_data(true);

    std::vector<BaseBenchmarkRecord> results;
    if (level != -1)
    {
        results = benchmark_continuos_compression_one_level(data, zOrderData, level);
        for (auto &&result : results)
        {
            result.fileName = fName.c_str();
            result.subblockId = 999;
            result.pixelType = cziFile.pixel_type_str(entry.pixelType);
            result.width = 0;
            result.height = 0;
        }

        write_compression_report(results, reportFile);
    }
    else
    {

        int levelDone = 0;
#pragma omp parallel for
        for (int compressionLevel = minCompressionLevel; compressionLevel <= maxCompressionLevel; compressionLevel++)
        {
            auto levelResults = benchmark_continuos_compression_one_level(data, zOrderData, compressionLevel);
#pragma omp critical
            {
                results.insert(results.end(), levelResults.begin(), levelResults.end());
                printf("\rFinished compression level %i/%i of normal order.", ++levelDone, maxCompressionLevel);
                fflush(stdout);
            }
        }
        if (verbose)
            printf("\n");
    }

    for (auto &&result : results)
    {
        result.fileName = fName.c_str();
        result.subblockId = 999;
        result.pixelType = cziFile.pixel_type_str(entry.pixelType);
        result.width = 0;
        result.height = 0;
    }

    if (verbose)
        printf("\nWriting report file...\n");

    write_compression_report(results, reportFile);

    printf("\nFinished benchmark, results are written in: %s\n", reportFile.c_str());
}

void benchmark_compression(CziFile &cziFile, const std::string &reportFile, bool verbose, int level)
{
    std::vector<BaseBenchmarkRecord> benchmarkResults;
    const int minCompressionLevel = 1;
    const int maxCompressionLevel = 9;
    std::string fName = fs_wrapper::get_filename(cziFile.fileName);

    int sbCount = (int)cziFile.subBlockDirectory.entries.size();
    for (size_t subblockId = 0; subblockId < cziFile.subBlockDirectory.entries.size(); subblockId++)
    {

        DirectoryEntryDV subblock = cziFile.subBlockDirectory.entries[subblockId];

        const char *pt = cziFile.pixel_type_str(subblock.pixelType);

        ByteArray data = cziFile.get_image_data(subblockId, false);
        ByteArray dataInZOrder = cziFile.get_image_data(subblockId, true);

        if (verbose)
        {
            printf("\rProcessing Subblock %i/%i", (int)subblockId + 1, sbCount);
            fflush(stdout);
        }
        int fromCL = (level == -1) ? minCompressionLevel : level;
        int toCL = (level == -1) ? maxCompressionLevel : level;
#pragma omp parallel for
        for (int compressionLevel = fromCL; compressionLevel <= toCL; compressionLevel++)
        {

            // CompressionMethod_GZIP
            {
                CompressionResult gzipResult = test_compression_method(data, CompressionMethod_GZIP, compressionLevel);
                CompressionResult gzipZOrderedResult = test_compression_method(dataInZOrder, CompressionMethod_GZIP, compressionLevel);

                BaseBenchmarkRecord gzipRecord = BaseBenchmarkRecord(gzipResult, gzipZOrderedResult);
                gzipRecord.set_metadata(fName.c_str(), subblockId, pt, subblock.width, subblock.height, "GZIP", compressionLevel);

#pragma omp critical
                {
                    benchmarkResults.push_back(gzipRecord);
                }
            }

            // CompressionMethod_LZMA
            {

                CompressionResult lzmaResult = test_compression_method(data, CompressionMethod_LZMA, compressionLevel);
                CompressionResult lzmaZOrderedResult = test_compression_method(dataInZOrder, CompressionMethod_LZMA, compressionLevel);

                BaseBenchmarkRecord lzmaRecord = BaseBenchmarkRecord(lzmaResult, lzmaZOrderedResult);
                lzmaRecord.set_metadata(fName.c_str(), subblockId, pt, subblock.width, subblock.height, "LZMA2", compressionLevel);
#pragma omp critical
                {
                    benchmarkResults.push_back(lzmaRecord);
                }
            }

            // CompressionMethod_BZIP2
            {
                CompressionResult bzip2Result = test_compression_method(data, CompressionMethod_BZIP2, compressionLevel);
                CompressionResult bzip2ZOrderedResult = test_compression_method(dataInZOrder, CompressionMethod_BZIP2, compressionLevel);

                BaseBenchmarkRecord bzip2Record = BaseBenchmarkRecord(bzip2Result, bzip2ZOrderedResult);
                bzip2Record.set_metadata(fName.c_str(), subblockId, pt, subblock.width, subblock.height, "BZIP2", compressionLevel);
#pragma omp critical
                {
                    benchmarkResults.push_back(bzip2Record);
                }
            }
        }
    }

    if (verbose)
        printf("\nWriting report file...\n");
    write_compression_report(benchmarkResults, reportFile);

    printf("\nFinished benchmark, results are written in: %s\n", reportFile.c_str());
}

static void _diff_from_frame_gray16(const ByteArray &pfData, const ByteArray &cfData,
                                    const CompressionMethod cm, const int compLevel, CompressionResult &cr)
{

    always_assert(pfData.size() == cfData.size());

    std::vector<int> delta;

    {
        std::vector<ushort> pfUshortValues = bytes_to_ushort_array(pfData);
        std::vector<ushort> cfUshortValues = bytes_to_ushort_array(cfData);
        always_assert(pfUshortValues.size() == cfUshortValues.size());

        delta = vecUtil::diff_vectors<ushort, int>(pfUshortValues, cfUshortValues);
    }

    auto minMax = vecUtil::find_min_max(delta);
    long maxRequiredValue = (minMax.first < 0) ? (abs(minMax.first) + minMax.second) : (minMax.second);
    bool canBeMappedToUShort = maxRequiredValue < USHORT_MAX;

    always_assert(canBeMappedToUShort && "Value can not be mapped into ushort.");

    //ByteArray deltaBytes = int_array_to_bytes(delta);

    // Here we map integers into ushort.
    ByteArray ushortMappedDeltaBytes;

    {
        TypeMapper<int, ushort> typeMapper;
        std::vector<ushort> ushortMappedDelta = typeMapper.map(delta);
        ushortMappedDeltaBytes = ushort_array_to_bytes(ushortMappedDelta);
    }

    cr = test_compression_method(ushortMappedDeltaBytes, cm, compLevel);
}

void value_diff_by_prev_frame_benchmark(CziFile &cziFile, const std::string &reportFile, bool verbose, int level, CompressionMethod cm)
{
    // NOTE: This benchmark works only for 16 bit pixels!
    printf("Compression method %s with compression level %i\n", compression_method_str(cm), level);
    std::string fName = fs_wrapper::get_filename(cziFile.fileName);
    auto framesByChannels = cziFile.get_subblocks_grouped_by_channels();

    uint iter = 0;
    uint iterCount = cziFile.subBlockDirectory.entryCount;
    std::vector<DiffBenchmarkRecord> benchmarkRecords;

    for (const std::pair<uint, std::vector<uint>> &channelGroup : framesByChannels)
    {
        printf_if(verbose, "Starting channel %u\n", channelGroup.first);
#pragma omp parallel for
        for (size_t i = 1; i < channelGroup.second.size(); i++)
        {
            DiffBenchmarkRecord record = {};

            uint prevFrameId = channelGroup.second[i - 1];
            uint currFrameId = channelGroup.second[i];

            DirectoryEntryDV prevEntry = cziFile.subBlockDirectory.entries[prevFrameId];
            DirectoryEntryDV currEntry = cziFile.subBlockDirectory.entries[currFrameId];
            always_assert(prevEntry.pixelType == PixelType_Gray16);
            always_assert(currEntry.pixelType == PixelType_Gray16);

            record.set_metadata(fName.c_str(), currFrameId, prevFrameId,
                                cziFile.pixel_type_str(currEntry.pixelType), currEntry.width, currEntry.height,
                                compression_method_str(cm), level, "PrevFrameDiff", 16, channelGroup.first);

            DimensionEntryDV1 prevDim = prevEntry.get_dimension(Dimension_Z);
            DimensionEntryDV1 currDim = currEntry.get_dimension(Dimension_Z);
            always_assert(!prevDim.isEmpty && !currDim.isEmpty);
            // We should always be going in direction of Z-stack.
            always_assert(prevDim.start < currDim.start);

            CompressionResult crN = {};
            CompressionResult crZ = {};

            // pf => previous frame, cf => current frame.
            size_t originalSize = 0;
            //std::thread(pfData,cfData, benchmarkRecord);
            {
                auto pfData = cziFile.get_image_data(prevFrameId, false);
                auto cfData = cziFile.get_image_data(currFrameId, false);
                _diff_from_frame_gray16(pfData, cfData, cm, level, crN);
                originalSize = pfData.size();
            }
            // Z order
            {
                auto pfDataZ = cziFile.get_image_data(prevFrameId, true);
                auto cfDataZ = cziFile.get_image_data(currFrameId, true);
                _diff_from_frame_gray16(pfDataZ, cfDataZ, cm, level, crZ);
                always_assert(pfDataZ.size() == originalSize);
            }

            record.originalSize = originalSize;
            record.compressedSize = crN.compressedSize;
            record.zOrderCompressedSize = crZ.compressedSize;
            record.compressionRatio = compression_ratio((float)originalSize, (float)crN.compressedSize);
            record.zOrderCompressionRatio = compression_ratio((float)originalSize, (float)crZ.compressedSize);
            record.compressionTime = 0;
            record.zOrderCompressionTime = 0;

#pragma omp critical
            {
                benchmarkRecords.push_back(record);

                printf("\rFinished %u/%u", ++iter, iterCount);
                fflush(stdout);
            }
        }
    }

    write_diff_report(benchmarkRecords, reportFile);
    printf("\rFinished %u/%u\n", iterCount, iterCount);
    printf("Report saved in %s\n", reportFile.c_str());
}

void bsdiff_by_prev_frame_benchmark(CziFile &cziFile, const std::string &reportFile, bool verbose)
{
    std::string fName = fs_wrapper::get_filename(cziFile.fileName);
    printf("Difference from previuos frame by bsdiff unix tool.\n");
    auto framesByChannels = cziFile.get_subblocks_grouped_by_channels();

    uint iter = 0;
    uint iterCount = cziFile.subBlockDirectory.entryCount;
    std::vector<DiffBenchmarkRecord> benchmarkRecords;

    for (const std::pair<uint, std::vector<uint>> &channelGroup : framesByChannels)
    {
        printf_if(verbose, "Starting channel %u\n", channelGroup.first);
        //#pragma omp parallel for
        for (size_t i = 1; i < channelGroup.second.size(); i++)
        {
            DiffBenchmarkRecord record = {};

            uint prevFrameId = channelGroup.second[i - 1];
            uint currFrameId = channelGroup.second[i];

            DirectoryEntryDV prevEntry = cziFile.subBlockDirectory.entries[prevFrameId];
            DirectoryEntryDV currEntry = cziFile.subBlockDirectory.entries[currFrameId];
            always_assert(prevEntry.pixelType == PixelType_Gray16);
            always_assert(currEntry.pixelType == PixelType_Gray16);

            record.set_metadata(fName.c_str(), currFrameId, prevFrameId,
                                cziFile.pixel_type_str(currEntry.pixelType), currEntry.width, currEntry.height,
                                "None", 0, "bsdiff_PrevFrameDiff", 16, channelGroup.first);

            DimensionEntryDV1 prevDim = prevEntry.get_dimension(Dimension_Z);
            DimensionEntryDV1 currDim = currEntry.get_dimension(Dimension_Z);
            always_assert(!prevDim.isEmpty && !currDim.isEmpty);
            // We should always be going in direction of Z-stack.
            always_assert(prevDim.start < currDim.start);

            // pf => previous frame, cf => current frame.
            size_t originalSize = 0;
            size_t patchSizeN = 0;
            size_t patchSizeZ = 0;
            {
                auto pfData = cziFile.get_image_data(prevFrameId, false);
                auto cfData = cziFile.get_image_data(currFrameId, false);
                originalSize = pfData.size();

                size_t maxPatchSize = lib_bsdiff::bsdiff_patchsize_max(pfData.size(), cfData.size());
                ByteArray patchBuffer;
                patchBuffer.resize(maxPatchSize);

                off_t actualBufferSize = lib_bsdiff::bsdiff(pfData.data(), pfData.size(), cfData.data(), cfData.size(), patchBuffer.data(), maxPatchSize);
                always_assert(actualBufferSize != -1 && "bsdiff error.");
                patchSizeN = actualBufferSize;
            }
            // Z order
            {
                auto pfDataZ = cziFile.get_image_data(prevFrameId, true);
                auto cfDataZ = cziFile.get_image_data(currFrameId, true);
                always_assert(pfDataZ.size() == originalSize);

                size_t maxPatchSize = lib_bsdiff::bsdiff_patchsize_max(pfDataZ.size(), cfDataZ.size());
                ByteArray patchBuffer;
                patchBuffer.resize(maxPatchSize);

                off_t actualBufferSize = lib_bsdiff::bsdiff(pfDataZ.data(), pfDataZ.size(), cfDataZ.data(), cfDataZ.size(), patchBuffer.data(), maxPatchSize);
                always_assert(actualBufferSize != -1 && "bsdiff error.");
                patchSizeZ = actualBufferSize;
            }

            record.originalSize = originalSize;
            record.compressedSize = patchSizeN;
            record.zOrderCompressedSize = patchSizeZ;
            record.compressionRatio = compression_ratio((float)originalSize, (float)patchSizeN);
            record.zOrderCompressionRatio = compression_ratio((float)originalSize, (float)patchSizeZ);
            record.compressionTime = 0;
            record.zOrderCompressionTime = 0;

#pragma omp critical
            {
                benchmarkRecords.push_back(record);

                printf("\rFinished %u/%u", ++iter, iterCount);
                fflush(stdout);
            }
        }
    }

    write_diff_report(benchmarkRecords, reportFile);
    printf("\rFinished %u/%u\n", iterCount, iterCount);
    printf("Report saved in %s\n", reportFile.c_str());
}