ts_perf.cpp

#include "precomp.hpp"

#ifdef HAVE_CUDA
#include "opencv2/core/gpumat.hpp"
#endif

#ifdef ANDROID
# include <sys/time.h>
#endif

using namespace perf;

int64 TestBase::timeLimitDefault = 0;
unsigned int TestBase::iterationsLimitDefault = (unsigned int)(-1);
int64 TestBase::_timeadjustment = 0;

const std::string command_line_keys =
    "{   perf_max_outliers   |8        |percent of allowed outliers}"
    "{   perf_min_samples    |10       |minimal required numer of samples}"
    "{   perf_force_samples  |100      |force set maximum number of samples for all tests}"
    "{   perf_seed           |809564   |seed for random numbers generator}"
    "{   perf_threads        |-1       |the number of worker threads, if parallel execution is enabled}"
    "{   perf_write_sanity   |         |create new records for sanity checks}"
    "{   perf_verify_sanity  |         |fail tests having no regression data for sanity checks}"
#ifdef ANDROID
    "{   perf_time_limit     |6.0      |default time limit for a single test (in seconds)}"
    "{   perf_affinity_mask  |0        |set affinity mask for the main thread}"
    "{   perf_log_power_checkpoints  | |additional xml logging for power measurement}"
#else
    "{   perf_time_limit     |3.0      |default time limit for a single test (in seconds)}"
#endif
    "{   perf_max_deviation  |1.0      |}"
    "{   help h              |         |print help info}"
#ifdef HAVE_CUDA
    "{   perf_run_cpu        |false    |run GPU performance tests for analogical CPU functions}"
    "{   perf_cuda_device    |0        |run GPU test suite onto specific CUDA capable device}"
    "{   perf_cuda_info_only |false    |print an information about system and an available CUDA devices and then exit.}"
#endif
;

static double       param_max_outliers;
static double       param_max_deviation;
static unsigned int param_min_samples;
static unsigned int param_force_samples;
static uint64       param_seed;
static double       param_time_limit;
static int          param_threads;
static bool         param_write_sanity;
static bool         param_verify_sanity;
#ifdef HAVE_CUDA
static bool         param_run_cpu;
static int          param_cuda_device;
#endif


#ifdef ANDROID
static int          param_affinity_mask;
static bool         log_power_checkpoints;

#include <sys/syscall.h>
#include <pthread.h>
static void setCurrentThreadAffinityMask(int mask)
{
    pid_t pid=gettid();
    int syscallres=syscall(__NR_sched_setaffinity, pid, sizeof(mask), &mask);
    if (syscallres)
    {
        int err=errno;
        err=err;//to avoid warnings about unused variables
        LOGE("Error in the syscall setaffinity: mask=%d=0x%x err=%d=0x%x", mask, mask, err, err);
    }
}
#endif

#ifdef HAVE_CUDA
# include <opencv2/core/gpumat.hpp>
#endif

namespace {

class PerfEnvironment: public ::testing::Environment
{
public:
    void TearDown()
    {
        cv::setNumThreads(-1);
    }
};

} // namespace

static void randu(cv::Mat& m)
{
    const int bigValue = 0x00000FFF;
    if (m.depth() < CV_32F)
    {
        int minmax[] = {0, 256};
        cv::Mat mr = cv::Mat(m.rows, (int)(m.cols * m.elemSize()), CV_8U, m.ptr(), m.step[0]);
        cv::randu(mr, cv::Mat(1, 1, CV_32S, minmax), cv::Mat(1, 1, CV_32S, minmax + 1));
    }
    else if (m.depth() == CV_32F)
    {
        //float minmax[] = {-FLT_MAX, FLT_MAX};
        float minmax[] = {-bigValue, bigValue};
        cv::Mat mr = m.reshape(1);
        cv::randu(mr, cv::Mat(1, 1, CV_32F, minmax), cv::Mat(1, 1, CV_32F, minmax + 1));
    }
    else
    {
        //double minmax[] = {-DBL_MAX, DBL_MAX};
        double minmax[] = {-bigValue, bigValue};
        cv::Mat mr = m.reshape(1);
        cv::randu(mr, cv::Mat(1, 1, CV_64F, minmax), cv::Mat(1, 1, CV_64F, minmax + 1));
    }
}

/*****************************************************************************************\
*                       inner exception class for early termination
\*****************************************************************************************/

class PerfEarlyExitException: public cv::Exception {};

/*****************************************************************************************\
*                                   ::perf::Regression
\*****************************************************************************************/

Regression& Regression::instance()
{
    static Regression single;
    return single;
}

Regression& Regression::add(TestBase* test, const std::string& name, cv::InputArray array, double eps, ERROR_TYPE err)
{
    if(test) test->verified = true;
    return instance()(name, array, eps, err);
}

Regression& Regression::addKeypoints(TestBase* test, const std::string& name, const std::vector<cv::KeyPoint>& array, double eps, ERROR_TYPE err)
{
    int len = (int)array.size();
    cv::Mat pt      (len, 1, CV_32FC2, (void*)&array[0].pt,       sizeof(cv::KeyPoint));
    cv::Mat size    (len, 1, CV_32FC1, (void*)&array[0].size,     sizeof(cv::KeyPoint));
    cv::Mat angle   (len, 1, CV_32FC1, (void*)&array[0].angle,    sizeof(cv::KeyPoint));
    cv::Mat response(len, 1, CV_32FC1, (void*)&array[0].response, sizeof(cv::KeyPoint));
    cv::Mat octave  (len, 1, CV_32SC1, (void*)&array[0].octave,   sizeof(cv::KeyPoint));
    cv::Mat class_id(len, 1, CV_32SC1, (void*)&array[0].class_id, sizeof(cv::KeyPoint));

    return Regression::add(test, name + "-pt",       pt,       eps, ERROR_ABSOLUTE)
                                (name + "-size",     size,     eps, ERROR_ABSOLUTE)
                                (name + "-angle",    angle,    eps, ERROR_ABSOLUTE)
                                (name + "-response", response, eps, err)
                                (name + "-octave",   octave,   eps, ERROR_ABSOLUTE)
                                (name + "-class_id", class_id, eps, ERROR_ABSOLUTE);
}

Regression& Regression::addMatches(TestBase* test, const std::string& name, const std::vector<cv::DMatch>& array, double eps, ERROR_TYPE err)
{
    int len = (int)array.size();
    cv::Mat queryIdx(len, 1, CV_32SC1, (void*)&array[0].queryIdx, sizeof(cv::DMatch));
    cv::Mat trainIdx(len, 1, CV_32SC1, (void*)&array[0].trainIdx, sizeof(cv::DMatch));
    cv::Mat imgIdx  (len, 1, CV_32SC1, (void*)&array[0].imgIdx,   sizeof(cv::DMatch));
    cv::Mat distance(len, 1, CV_32FC1, (void*)&array[0].distance, sizeof(cv::DMatch));

    return Regression::add(test, name + "-queryIdx", queryIdx, DBL_EPSILON, ERROR_ABSOLUTE)
                                (name + "-trainIdx", trainIdx, DBL_EPSILON, ERROR_ABSOLUTE)
                                (name + "-imgIdx",   imgIdx,   DBL_EPSILON, ERROR_ABSOLUTE)
                                (name + "-distance", distance, eps, err);
}

void Regression::Init(const std::string& testSuitName, const std::string& ext)
{
    instance().init(testSuitName, ext);
}

void Regression::init(const std::string& testSuitName, const std::string& ext)
{
    if (!storageInPath.empty())
    {
        LOGE("Subsequent initialisation of Regression utility is not allowed.");
        return;
    }

    const char *data_path_dir = getenv("OPENCV_TEST_DATA_PATH");
    const char *path_separator = "/";

    if (data_path_dir)
    {
        int len = (int)strlen(data_path_dir)-1;
        if (len < 0) len = 0;
        std::string path_base = (data_path_dir[0] == 0 ? std::string(".") : std::string(data_path_dir))
                + (data_path_dir[len] == '/' || data_path_dir[len] == '\\' ? "" : path_separator)
                + "perf"
                + path_separator;

        storageInPath = path_base + testSuitName + ext;
        storageOutPath = path_base + testSuitName;
    }
    else
    {
        storageInPath = testSuitName + ext;
        storageOutPath = testSuitName;
    }

    suiteName = testSuitName;

    try
    {
        if (storageIn.open(storageInPath, cv::FileStorage::READ))
        {
            rootIn = storageIn.root();
            if (storageInPath.length() > 3 && storageInPath.substr(storageInPath.length()-3) == ".gz")
                storageOutPath += "_new";
            storageOutPath += ext;
        }
    }
    catch(cv::Exception&)
    {
        LOGE("Failed to open sanity data for reading: %s", storageInPath.c_str());
    }

    if(!storageIn.isOpened())
        storageOutPath = storageInPath;
}

Regression::Regression() : regRNG(cv::getTickCount())//this rng should be really random
{
}

Regression::~Regression()
{
    if (storageIn.isOpened())
        storageIn.release();
    if (storageOut.isOpened())
    {
        if (!currentTestNodeName.empty())
            storageOut << "}";
        storageOut.release();
    }
}

cv::FileStorage& Regression::write()
{
    if (!storageOut.isOpened() && !storageOutPath.empty())
    {
        int mode = (storageIn.isOpened() && storageInPath == storageOutPath)
                ? cv::FileStorage::APPEND : cv::FileStorage::WRITE;
        storageOut.open(storageOutPath, mode);
        if (!storageOut.isOpened())
        {
            LOGE("Could not open \"%s\" file for writing", storageOutPath.c_str());
            storageOutPath.clear();
        }
        else if (mode == cv::FileStorage::WRITE && !rootIn.empty())
        {
            //TODO: write content of rootIn node into the storageOut
        }
    }
    return storageOut;
}

std::string Regression::getCurrentTestNodeName()
{
    const ::testing::TestInfo* const test_info =
      ::testing::UnitTest::GetInstance()->current_test_info();

    if (test_info == 0)
        return "undefined";

    std::string nodename = std::string(test_info->test_case_name()) + "--" + test_info->name();
    size_t idx = nodename.find_first_of('/');
    if (idx != std::string::npos)
        nodename.erase(idx);

    const char* type_param = test_info->type_param();
    if (type_param != 0)
        (nodename += "--") += type_param;

    const char* value_param = test_info->value_param();
    if (value_param != 0)
        (nodename += "--") += value_param;

    for(size_t i = 0; i < nodename.length(); ++i)
        if (!isalnum(nodename[i]) && '_' != nodename[i])
            nodename[i] = '-';

    return nodename;
}

bool Regression::isVector(cv::InputArray a)
{
    return a.kind() == cv::_InputArray::STD_VECTOR_MAT || a.kind() == cv::_InputArray::STD_VECTOR_VECTOR;
}

double Regression::getElem(cv::Mat& m, int y, int x, int cn)
{
    switch (m.depth())
    {
    case CV_8U: return *(m.ptr<unsigned char>(y, x) + cn);
    case CV_8S: return *(m.ptr<signed char>(y, x) + cn);
    case CV_16U: return *(m.ptr<unsigned short>(y, x) + cn);
    case CV_16S: return *(m.ptr<signed short>(y, x) + cn);
    case CV_32S: return *(m.ptr<signed int>(y, x) + cn);
    case CV_32F: return *(m.ptr<float>(y, x) + cn);
    case CV_64F: return *(m.ptr<double>(y, x) + cn);
    default: return 0;
    }
}

void Regression::write(cv::Mat m)
{
    if (!m.empty() && m.dims < 2) return;

    double min, max;
    cv::minMaxIdx(m, &min, &max);
    write() << "min" << min << "max" << max;

    write() << "last" << "{" << "x" << m.size.p[1] - 1 << "y" << m.size.p[0] - 1
        << "val" << getElem(m, m.size.p[0] - 1, m.size.p[1] - 1, m.channels() - 1) << "}";

    int x, y, cn;
    x = regRNG.uniform(0, m.size.p[1]);
    y = regRNG.uniform(0, m.size.p[0]);
    cn = regRNG.uniform(0, m.channels());
    write() << "rng1" << "{" << "x" << x << "y" << y;
    if(cn > 0) write() << "cn" << cn;
    write() << "val" << getElem(m, y, x, cn) << "}";

    x = regRNG.uniform(0, m.size.p[1]);
    y = regRNG.uniform(0, m.size.p[0]);
    cn = regRNG.uniform(0, m.channels());
    write() << "rng2" << "{" << "x" << x << "y" << y;
    if (cn > 0) write() << "cn" << cn;
    write() << "val" << getElem(m, y, x, cn) << "}";
}

static double evalEps(double expected, double actual, double _eps, ERROR_TYPE err)
{
    if (err == ERROR_ABSOLUTE)
        return _eps;
    else if (err == ERROR_RELATIVE)
        return std::max(std::abs(expected), std::abs(actual)) * _eps;
    return 0;
}

void Regression::verify(cv::FileNode node, cv::Mat actual, double _eps, std::string argname, ERROR_TYPE err)
{
    if (!actual.empty() && actual.dims < 2) return;

    double actual_min, actual_max;
    cv::minMaxIdx(actual, &actual_min, &actual_max);

    double expect_min = (double)node["min"];
    double eps = evalEps(expect_min, actual_min, _eps, err);
    ASSERT_NEAR(expect_min, actual_min, eps)
            << argname << " has unexpected minimal value" << std::endl;

    double expect_max = (double)node["max"];
    eps = evalEps(expect_max, actual_max, _eps, err);
    ASSERT_NEAR(expect_max, actual_max, eps)
            << argname << " has unexpected maximal value" << std::endl;

    cv::FileNode last = node["last"];
    double actual_last = getElem(actual, actual.size.p[0] - 1, actual.size.p[1] - 1, actual.channels() - 1);
    int expect_cols = (int)last["x"] + 1;
    int expect_rows = (int)last["y"] + 1;
    ASSERT_EQ(expect_cols, actual.size.p[1])
            << argname << " has unexpected number of columns" << std::endl;
    ASSERT_EQ(expect_rows, actual.size.p[0])
            << argname << " has unexpected number of rows" << std::endl;

    double expect_last = (double)last["val"];
    eps = evalEps(expect_last, actual_last, _eps, err);
    ASSERT_NEAR(expect_last, actual_last, eps)
            << argname << " has unexpected value of the last element" << std::endl;

    cv::FileNode rng1 = node["rng1"];
    int x1 = rng1["x"];
    int y1 = rng1["y"];
    int cn1 = rng1["cn"];

    double expect_rng1 = (double)rng1["val"];
    // it is safe to use x1 and y1 without checks here because we have already
    // verified that mat size is the same as recorded
    double actual_rng1 = getElem(actual, y1, x1, cn1);

    eps = evalEps(expect_rng1, actual_rng1, _eps, err);
    ASSERT_NEAR(expect_rng1, actual_rng1, eps)
            << argname << " has unexpected value of the ["<< x1 << ":" << y1 << ":" << cn1 <<"] element" << std::endl;

    cv::FileNode rng2 = node["rng2"];
    int x2 = rng2["x"];
    int y2 = rng2["y"];
    int cn2 = rng2["cn"];

    double expect_rng2 = (double)rng2["val"];
    double actual_rng2 = getElem(actual, y2, x2, cn2);

    eps = evalEps(expect_rng2, actual_rng2, _eps, err);
    ASSERT_NEAR(expect_rng2, actual_rng2, eps)
            << argname << " has unexpected value of the ["<< x2 << ":" << y2 << ":" << cn2 <<"] element" << std::endl;
}

void Regression::write(cv::InputArray array)
{
    write() << "kind" << array.kind();
    write() << "type" << array.type();
    if (isVector(array))
    {
        int total = (int)array.total();
        int idx = regRNG.uniform(0, total);
        write() << "len" << total;
        write() << "idx" << idx;

        cv::Mat m = array.getMat(idx);

        if (m.total() * m.channels() < 26) //5x5 or smaller
            write() << "val" << m;
        else
            write(m);
    }
    else
    {
        if (array.total() * array.channels() < 26) //5x5 or smaller
            write() << "val" << array.getMat();
        else
            write(array.getMat());
    }
}

static int countViolations(const cv::Mat& expected, const cv::Mat& actual, const cv::Mat& diff, double eps, double* max_violation = 0, double* max_allowed = 0)
{
    cv::Mat diff64f;
    diff.reshape(1).convertTo(diff64f, CV_64F);

    cv::Mat expected_abs = cv::abs(expected.reshape(1));
    cv::Mat actual_abs = cv::abs(actual.reshape(1));
    cv::Mat maximum, mask;
    cv::max(expected_abs, actual_abs, maximum);
    cv::multiply(maximum, cv::Vec<double, 1>(eps), maximum, CV_64F);
    cv::compare(diff64f, maximum, mask, cv::CMP_GT);

    int v = cv::countNonZero(mask);

    if (v > 0 && max_violation != 0 && max_allowed != 0)
    {
        int loc[10];
        cv::minMaxIdx(maximum, 0, max_allowed, 0, loc, mask);
        *max_violation = diff64f.at<double>(loc[1], loc[0]);
    }

    return v;
}

void Regression::verify(cv::FileNode node, cv::InputArray array, double eps, ERROR_TYPE err)
{
    int expected_kind = (int)node["kind"];
    int expected_type = (int)node["type"];
    ASSERT_EQ(expected_kind, array.kind()) << "  Argument \"" << node.name() << "\" has unexpected kind";
    ASSERT_EQ(expected_type, array.type()) << "  Argument \"" << node.name() << "\" has unexpected type";

    cv::FileNode valnode = node["val"];
    if (isVector(array))
    {
        int expected_length = (int)node["len"];
        ASSERT_EQ(expected_length, (int)array.total()) << "  Vector \"" << node.name() << "\" has unexpected length";
        int idx = node["idx"];

        cv::Mat actual = array.getMat(idx);

        if (valnode.isNone())
        {
            ASSERT_LE((size_t)26, actual.total() * (size_t)actual.channels())
                    << "  \"" << node.name() << "[" <<  idx << "]\" has unexpected number of elements";
            verify(node, actual, eps, cv::format("%s[%d]", node.name().c_str(), idx), err);
        }
        else
        {
            cv::Mat expected;
            valnode >> expected;

            if(expected.empty())
            {
                ASSERT_TRUE(actual.empty())
                    << "  expected empty " << node.name() << "[" <<  idx<< "]";
            }
            else
            {
                ASSERT_EQ(expected.size(), actual.size())
                        << "  " << node.name() << "[" <<  idx<< "] has unexpected size";

                cv::Mat diff;
                cv::absdiff(expected, actual, diff);

                if (err == ERROR_ABSOLUTE)
                {
                    if (!cv::checkRange(diff, true, 0, 0, eps))
                    {
                        if(expected.total() * expected.channels() < 12)
                            std::cout << " Expected: " << std::endl << expected << std::endl << " Actual:" << std::endl << actual << std::endl;

                        double max;
                        cv::minMaxIdx(diff.reshape(1), 0, &max);

                        FAIL() << "  Absolute difference (=" << max << ") between argument \""
                               << node.name() << "[" <<  idx << "]\" and expected value is greater than " << eps;
                    }
                }
                else if (err == ERROR_RELATIVE)
                {
                    double maxv, maxa;
                    int violations = countViolations(expected, actual, diff, eps, &maxv, &maxa);
                    if (violations > 0)
                    {
                        FAIL() << "  Relative difference (" << maxv << " of " << maxa << " allowed) between argument \""
                               << node.name() << "[" <<  idx << "]\" and expected value is greater than " << eps << " in " << violations << " points";
                    }
                }
            }
        }
    }
    else
    {
        if (valnode.isNone())
        {
            ASSERT_LE((size_t)26, array.total() * (size_t)array.channels())
                    << "  Argument \"" << node.name() << "\" has unexpected number of elements";
            verify(node, array.getMat(), eps, "Argument \"" + node.name() + "\"", err);
        }
        else
        {
            cv::Mat expected;
            valnode >> expected;
            cv::Mat actual = array.getMat();

            if(expected.empty())
            {
                ASSERT_TRUE(actual.empty())
                    << "  expected empty " << node.name();
            }
            else
            {
                ASSERT_EQ(expected.size(), actual.size())
                        << "  Argument \"" << node.name() << "\" has unexpected size";

                cv::Mat diff;
                cv::absdiff(expected, actual, diff);

                if (err == ERROR_ABSOLUTE)
                {
                    if (!cv::checkRange(diff, true, 0, 0, eps))
                    {
                        if(expected.total() * expected.channels() < 12)
                            std::cout << " Expected: " << std::endl << expected << std::endl << " Actual:" << std::endl << actual << std::endl;

                        double max;
                        cv::minMaxIdx(diff.reshape(1), 0, &max);

                        FAIL() << "  Difference (=" << max << ") between argument1 \"" << node.name()
                               << "\" and expected value is greater than " << eps;
                    }
                }
                else if (err == ERROR_RELATIVE)
                {
                    double maxv, maxa;
                    int violations = countViolations(expected, actual, diff, eps, &maxv, &maxa);
                    if (violations > 0)
                    {
                        FAIL() << "  Relative difference (" << maxv << " of " << maxa << " allowed) between argument \"" << node.name()
                               << "\" and expected value is greater than " << eps << " in " << violations << " points";
                    }
                }
            }
        }
    }
}

Regression& Regression::operator() (const std::string& name, cv::InputArray array, double eps, ERROR_TYPE err)
{
    // exit if current test is already failed
    if(::testing::UnitTest::GetInstance()->current_test_info()->result()->Failed()) return *this;

    if(!array.empty() && array.depth() == CV_USRTYPE1)
    {
        ADD_FAILURE() << "  Can not check regression for CV_USRTYPE1 data type for " << name;
        return *this;
    }

    std::string nodename = getCurrentTestNodeName();

#ifdef HAVE_CUDA
    static const std::string prefix = (param_run_cpu)? "CPU_" : "GPU_";
    if(suiteName == "gpu")
        nodename = prefix + nodename;
#endif

    cv::FileNode n = rootIn[nodename];
    if(n.isNone())
    {
        if(param_write_sanity)
        {
            if (nodename != currentTestNodeName)
            {
                if (!currentTestNodeName.empty())
                    write() << "}";
                currentTestNodeName = nodename;

                write() << nodename << "{";
            }
            // TODO: verify that name is alphanumeric, current error message is useless
            write() << name << "{";
            write(array);
            write() << "}";
        }
        else if(param_verify_sanity)
        {
            ADD_FAILURE() << "  No regression data for " << name << " argument";
        }
    }
    else
    {
        cv::FileNode this_arg = n[name];
        if (!this_arg.isMap())
            ADD_FAILURE() << "  No regression data for " << name << " argument";
        else
            verify(this_arg, array, eps, err);
    }

    return *this;
}


/*****************************************************************************************\
*                                ::perf::performance_metrics
\*****************************************************************************************/
performance_metrics::performance_metrics()
{
    bytesIn = 0;
    bytesOut = 0;
    samples = 0;
    outliers = 0;
    gmean = 0;
    gstddev = 0;
    mean = 0;
    stddev = 0;
    median = 0;
    min = 0;
    frequency = 0;
    terminationReason = TERM_UNKNOWN;
}


/*****************************************************************************************\
*                                   ::perf::TestBase
\*****************************************************************************************/


void TestBase::Init(int argc, const char* const argv[])
{
    cv::CommandLineParser args(argc, argv, command_line_keys);
    if (args.has("help"))
    {
        args.printMessage();
        return;
    }

    ::testing::AddGlobalTestEnvironment(new PerfEnvironment);

    param_max_outliers  = std::min(100., std::max(0., args.get<double>("perf_max_outliers")));
    param_min_samples   = std::max(1u, args.get<unsigned int>("perf_min_samples"));
    param_max_deviation = std::max(0., args.get<double>("perf_max_deviation"));
    param_seed          = args.get<unsigned int>("perf_seed");
    param_time_limit    = std::max(0., args.get<double>("perf_time_limit"));
    param_force_samples = args.get<unsigned int>("perf_force_samples");
    param_write_sanity  = args.has("perf_write_sanity");
    param_verify_sanity = args.has("perf_verify_sanity");
    param_threads  = args.get<int>("perf_threads");
#ifdef ANDROID
    param_affinity_mask   = args.get<int>("perf_affinity_mask");
    log_power_checkpoints = args.has("perf_log_power_checkpoints");
#endif

#ifdef HAVE_CUDA

    bool printOnly        = args.has("perf_cuda_info_only");

    if (printOnly)
        exit(0);

    param_run_cpu         = args.has("perf_run_cpu");
    param_cuda_device      = std::max(0, std::min(cv::gpu::getCudaEnabledDeviceCount(), args.get<int>("perf_cuda_device")));

    if (param_run_cpu)
        printf("[----------]\n[ GPU INFO ] \tRun test suite on CPU.\n[----------]\n"), fflush(stdout);
    else
    {
        cv::gpu::DeviceInfo info(param_cuda_device);
        if (!info.isCompatible())
        {
            printf("[----------]\n[ FAILURE  ] \tDevice %s is NOT compatible with current GPU module build.\n[----------]\n", info.name().c_str()), fflush(stdout);
            exit(-1);
        }

        cv::gpu::setDevice(param_cuda_device);

        printf("[----------]\n[ GPU INFO ] \tRun test suite on %s GPU.\n[----------]\n", info.name().c_str()), fflush(stdout);
    }
#endif

    if (!args.check())
    {
        args.printErrors();
        return;
    }

    timeLimitDefault = param_time_limit == 0.0 ? 1 : (int64)(param_time_limit * cv::getTickFrequency());
    iterationsLimitDefault = param_force_samples == 0 ? (unsigned)(-1) : param_force_samples;
    _timeadjustment = _calibrate();
}

int64 TestBase::_calibrate()
{
    class _helper : public ::perf::TestBase
    {
        public:
        performance_metrics& getMetrics() { return calcMetrics(); }
        virtual void TestBody() {}
        virtual void PerfTestBody()
        {
            //the whole system warmup
            SetUp();
            cv::Mat a(2048, 2048, CV_32S, cv::Scalar(1));
            cv::Mat b(2048, 2048, CV_32S, cv::Scalar(2));
            declare.time(30);
            double s = 0;
            for(declare.iterations(20); startTimer(), next(); stopTimer())
                s+=a.dot(b);
            declare.time(s);

            //self calibration
            SetUp();
            for(declare.iterations(1000); startTimer(), next(); stopTimer()){}
        }
    };

    _timeadjustment = 0;
    _helper h;
    h.PerfTestBody();
    double compensation = h.getMetrics().min;
    LOGD("Time compensation is %.0f", compensation);
    return (int64)compensation;
}

#ifdef _MSC_VER
# pragma warning(push)
# pragma warning(disable:4355)  // 'this' : used in base member initializer list
#endif
TestBase::TestBase(): declare(this)
{
}
#ifdef _MSC_VER
# pragma warning(pop)
#endif


void TestBase::declareArray(SizeVector& sizes, cv::InputOutputArray a, int wtype)
{
    if (!a.empty())
    {
        sizes.push_back(std::pair<int, cv::Size>(getSizeInBytes(a), getSize(a)));
        warmup(a, wtype);
    }
    else if (a.kind() != cv::_InputArray::NONE)
        ADD_FAILURE() << "  Uninitialized input/output parameters are not allowed for performance tests";
}

void TestBase::warmup(cv::InputOutputArray a, int wtype)
{
    if (a.empty()) return;
    if (a.kind() != cv::_InputArray::STD_VECTOR_MAT && a.kind() != cv::_InputArray::STD_VECTOR_VECTOR)
        warmup_impl(a.getMat(), wtype);
    else
    {
        size_t total = a.total();
        for (size_t i = 0; i < total; ++i)
            warmup_impl(a.getMat((int)i), wtype);
    }
}

int TestBase::getSizeInBytes(cv::InputArray a)
{
    if (a.empty()) return 0;
    int total = (int)a.total();
    if (a.kind() != cv::_InputArray::STD_VECTOR_MAT && a.kind() != cv::_InputArray::STD_VECTOR_VECTOR)
        return total * CV_ELEM_SIZE(a.type());

    int size = 0;
    for (int i = 0; i < total; ++i)
        size += (int)a.total(i) * CV_ELEM_SIZE(a.type(i));

    return size;
}

cv::Size TestBase::getSize(cv::InputArray a)
{
    if (a.kind() != cv::_InputArray::STD_VECTOR_MAT && a.kind() != cv::_InputArray::STD_VECTOR_VECTOR)
        return a.size();
    return cv::Size();
}

bool TestBase::next()
{
    bool has_next = ++currentIter < nIters && totalTime < timeLimit;
    cv::theRNG().state = param_seed; //this rng should generate same numbers for each run

#ifdef ANDROID
    if (log_power_checkpoints)
    {
        timeval tim;
        gettimeofday(&tim, NULL);
        unsigned long long t1 = tim.tv_sec * 1000LLU + (unsigned long long)(tim.tv_usec / 1000.f);

        if (currentIter == 1) RecordProperty("test_start", cv::format("%llu",t1).c_str());
        if (!has_next) RecordProperty("test_complete", cv::format("%llu",t1).c_str());
    }
#endif
    return has_next;
}

void TestBase::warmup_impl(cv::Mat m, int wtype)
{
    switch(wtype)
    {
    case WARMUP_READ:
        cv::sum(m.reshape(1));
        return;
    case WARMUP_WRITE:
        m.reshape(1).setTo(cv::Scalar::all(0));
        return;
    case WARMUP_RNG:
        randu(m);
        return;
    default:
        return;
    }
}

unsigned int TestBase::getTotalInputSize() const
{
    unsigned int res = 0;
    for (SizeVector::const_iterator i = inputData.begin(); i != inputData.end(); ++i)
        res += i->first;
    return res;
}

unsigned int TestBase::getTotalOutputSize() const
{
    unsigned int res = 0;
    for (SizeVector::const_iterator i = outputData.begin(); i != outputData.end(); ++i)
        res += i->first;
    return res;
}

void TestBase::startTimer()
{
    lastTime = cv::getTickCount();
}

void TestBase::stopTimer()
{
    int64 time = cv::getTickCount();
    if (lastTime == 0)
        ADD_FAILURE() << "  stopTimer() is called before startTimer()";
    lastTime = time - lastTime;
    totalTime += lastTime;
    lastTime -= _timeadjustment;
    if (lastTime < 0) lastTime = 0;
    times.push_back(lastTime);
    lastTime = 0;
}

performance_metrics& TestBase::calcMetrics()
{
    if ((metrics.samples == (unsigned int)currentIter) || times.size() == 0)
        return metrics;

    metrics.bytesIn = getTotalInputSize();
    metrics.bytesOut = getTotalOutputSize();
    metrics.frequency = cv::getTickFrequency();
    metrics.samples = (unsigned int)times.size();
    metrics.outliers = 0;

    if (metrics.terminationReason != performance_metrics::TERM_INTERRUPT && metrics.terminationReason != performance_metrics::TERM_EXCEPTION)
    {
        if (currentIter == nIters)
            metrics.terminationReason = performance_metrics::TERM_ITERATIONS;
        else if (totalTime >= timeLimit)
            metrics.terminationReason = performance_metrics::TERM_TIME;
        else
            metrics.terminationReason = performance_metrics::TERM_UNKNOWN;
    }

    std::sort(times.begin(), times.end());

    //estimate mean and stddev for log(time)
    double gmean = 0;
    double gstddev = 0;
    int n = 0;
    for(TimeVector::const_iterator i = times.begin(); i != times.end(); ++i)
    {
        double x = static_cast<double>(*i)/runsPerIteration;
        if (x < DBL_EPSILON) continue;
        double lx = log(x);

        ++n;
        double delta = lx - gmean;
        gmean += delta / n;
        gstddev += delta * (lx - gmean);
    }

    gstddev = n > 1 ? sqrt(gstddev / (n - 1)) : 0;

    TimeVector::const_iterator start = times.begin();
    TimeVector::const_iterator end = times.end();

    //filter outliers assuming log-normal distribution
    //http://stackoverflow.com/questions/1867426/modeling-distribution-of-performance-measurements
    int offset = 0;
    if (gstddev > DBL_EPSILON)
    {
        double minout = exp(gmean - 3 * gstddev) * runsPerIteration;
        double maxout = exp(gmean + 3 * gstddev) * runsPerIteration;
        while(*start < minout) ++start, ++metrics.outliers, ++offset;
        do --end, ++metrics.outliers; while(*end > maxout);
        ++end, --metrics.outliers;
    }

    metrics.min = static_cast<double>(*start)/runsPerIteration;
    //calc final metrics
    n = 0;
    gmean = 0;
    gstddev = 0;
    double mean = 0;
    double stddev = 0;
    int m = 0;
    for(; start != end; ++start)
    {
        double x = static_cast<double>(*start)/runsPerIteration;
        if (x > DBL_EPSILON)
        {
            double lx = log(x);
            ++m;
            double gdelta = lx - gmean;
            gmean += gdelta / m;
            gstddev += gdelta * (lx - gmean);
        }
        ++n;
        double delta = x - mean;
        mean += delta / n;
        stddev += delta * (x - mean);
    }

    metrics.mean = mean;
    metrics.gmean = exp(gmean);
    metrics.gstddev = m > 1 ? sqrt(gstddev / (m - 1)) : 0;
    metrics.stddev = n > 1 ? sqrt(stddev / (n - 1)) : 0;
    metrics.median = n % 2
            ? (double)times[offset + n / 2]
            : 0.5 * (times[offset + n / 2] + times[offset + n / 2 - 1]);

    metrics.median /= runsPerIteration;

    return metrics;
}

void TestBase::validateMetrics()
{
    performance_metrics& m = calcMetrics();

    if (HasFailure()) return;

    ASSERT_GE(m.samples, 1u)
      << "  No time measurements was performed.\nstartTimer() and stopTimer() commands are required for performance tests.";

    EXPECT_GE(m.samples, param_min_samples)
      << "  Only a few samples are collected.\nPlease increase number of iterations or/and time limit to get reliable performance measurements.";

    if (m.gstddev > DBL_EPSILON)
    {
        EXPECT_GT(/*m.gmean * */1., /*m.gmean * */ 2 * sinh(m.gstddev * param_max_deviation))
          << "  Test results are not reliable ((mean-sigma,mean+sigma) deviation interval is greater than measured time interval).";
    }

    EXPECT_LE(m.outliers, std::max((unsigned int)cvCeil(m.samples * param_max_outliers / 100.), 1u))
      << "  Test results are not reliable (too many outliers).";
}

void TestBase::reportMetrics(bool toJUnitXML)
{
    performance_metrics& m = calcMetrics();

    if (toJUnitXML)
    {
        RecordProperty("bytesIn", (int)m.bytesIn);
        RecordProperty("bytesOut", (int)m.bytesOut);
        RecordProperty("term", m.terminationReason);
        RecordProperty("samples", (int)m.samples);
        RecordProperty("outliers", (int)m.outliers);
        RecordProperty("frequency", cv::format("%.0f", m.frequency).c_str());
        RecordProperty("min", cv::format("%.0f", m.min).c_str());
        RecordProperty("median", cv::format("%.0f", m.median).c_str());
        RecordProperty("gmean", cv::format("%.0f", m.gmean).c_str());
        RecordProperty("gstddev", cv::format("%.6f", m.gstddev).c_str());
        RecordProperty("mean", cv::format("%.0f", m.mean).c_str());
        RecordProperty("stddev", cv::format("%.0f", m.stddev).c_str());
    }
    else
    {
        const ::testing::TestInfo* const test_info = ::testing::UnitTest::GetInstance()->current_test_info();
        const char* type_param = test_info->type_param();
        const char* value_param = test_info->value_param();

#if defined(ANDROID) && defined(USE_ANDROID_LOGGING)
        LOGD("[ FAILED   ] %s.%s", test_info->test_case_name(), test_info->name());
#endif

        if (type_param)  LOGD("type      = %11s", type_param);
        if (value_param) LOGD("params    = %11s", value_param);

        switch (m.terminationReason)
        {
        case performance_metrics::TERM_ITERATIONS:
            LOGD("termination reason:  reached maximum number of iterations");
            break;
        case performance_metrics::TERM_TIME:
            LOGD("termination reason:  reached time limit");
            break;
        case performance_metrics::TERM_INTERRUPT:
            LOGD("termination reason:  aborted by the performance testing framework");
            break;
        case performance_metrics::TERM_EXCEPTION:
            LOGD("termination reason:  unhandled exception");
            break;
        case performance_metrics::TERM_UNKNOWN:
        default:
            LOGD("termination reason:  unknown");
            break;
        };

        LOGD("bytesIn   =%11lu", (unsigned long)m.bytesIn);
        LOGD("bytesOut  =%11lu", (unsigned long)m.bytesOut);
        if (nIters == (unsigned int)-1 || m.terminationReason == performance_metrics::TERM_ITERATIONS)
            LOGD("samples   =%11u",  m.samples);
        else
            LOGD("samples   =%11u of %u", m.samples, nIters);
        LOGD("outliers  =%11u", m.outliers);
        LOGD("frequency =%11.0f", m.frequency);
        if (m.samples > 0)
        {
            LOGD("min       =%11.0f = %.2fms", m.min, m.min * 1e3 / m.frequency);
            LOGD("median    =%11.0f = %.2fms", m.median, m.median * 1e3 / m.frequency);
            LOGD("gmean     =%11.0f = %.2fms", m.gmean, m.gmean * 1e3 / m.frequency);
            LOGD("gstddev   =%11.8f = %.2fms for 97%% dispersion interval", m.gstddev, m.gmean * 2 * sinh(m.gstddev * 3) * 1e3 / m.frequency);
            LOGD("mean      =%11.0f = %.2fms", m.mean, m.mean * 1e3 / m.frequency);
            LOGD("stddev    =%11.0f = %.2fms", m.stddev, m.stddev * 1e3 / m.frequency);
        }
    }
}

void TestBase::SetUp()
{
    cv::theRNG().state = param_seed; // this rng should generate same numbers for each run

    if (param_threads >= 0)
        cv::setNumThreads(param_threads);

#ifdef ANDROID
    if (param_affinity_mask)
        setCurrentThreadAffinityMask(param_affinity_mask);
#endif

    verified = false;
    lastTime = 0;
    totalTime = 0;
    runsPerIteration = 1;
    nIters = iterationsLimitDefault;
    currentIter = (unsigned int)-1;
    timeLimit = timeLimitDefault;
    times.clear();
}

void TestBase::TearDown()
{
    if (!HasFailure() && !verified)
        ADD_FAILURE() << "The test has no sanity checks. There should be at least one check at the end of performance test.";

    validateMetrics();
    if (HasFailure())
        reportMetrics(false);
    else
    {
        const ::testing::TestInfo* const test_info = ::testing::UnitTest::GetInstance()->current_test_info();
        const char* type_param = test_info->type_param();
        const char* value_param = test_info->value_param();
        if (value_param) printf("[ VALUE    ] \t%s\n", value_param), fflush(stdout);
        if (type_param)  printf("[ TYPE     ] \t%s\n", type_param), fflush(stdout);
        reportMetrics(true);
    }
}

std::string TestBase::getDataPath(const std::string& relativePath)
{
    if (relativePath.empty())
    {
        ADD_FAILURE() << "  Bad path to test resource";
        throw PerfEarlyExitException();
    }

    const char *data_path_dir = getenv("OPENCV_TEST_DATA_PATH");
    const char *path_separator = "/";

    std::string path;
    if (data_path_dir)
    {
        int len = (int)strlen(data_path_dir) - 1;
        if (len < 0) len = 0;
        path = (data_path_dir[0] == 0 ? std::string(".") : std::string(data_path_dir))
                + (data_path_dir[len] == '/' || data_path_dir[len] == '\\' ? "" : path_separator);
    }
    else
    {
        path = ".";
        path += path_separator;
    }

    if (relativePath[0] == '/' || relativePath[0] == '\\')
        path += relativePath.substr(1);
    else
        path += relativePath;

    FILE* fp = fopen(path.c_str(), "r");
    if (fp)
        fclose(fp);
    else
    {
        ADD_FAILURE() << "  Requested file \"" << path << "\" does not exist.";
        throw PerfEarlyExitException();
    }
    return path;
}

void TestBase::RunPerfTestBody()
{
    try
    {
        this->PerfTestBody();
    }
    catch(PerfEarlyExitException)
    {
        metrics.terminationReason = performance_metrics::TERM_INTERRUPT;
        return;//no additional failure logging
    }
    catch(cv::Exception e)
    {
        metrics.terminationReason = performance_metrics::TERM_EXCEPTION;
        #ifdef HAVE_CUDA
            if (e.code == CV_GpuApiCallError)
                cv::gpu::resetDevice();
        #endif
        FAIL() << "Expected: PerfTestBody() doesn't throw an exception.\n  Actual: it throws cv::Exception:\n  " << e.what();
    }
    catch(std::exception e)
    {
        metrics.terminationReason = performance_metrics::TERM_EXCEPTION;
        FAIL() << "Expected: PerfTestBody() doesn't throw an exception.\n  Actual: it throws std::exception:\n  " << e.what();
    }
    catch(...)
    {
        metrics.terminationReason = performance_metrics::TERM_EXCEPTION;
        FAIL() << "Expected: PerfTestBody() doesn't throw an exception.\n  Actual: it throws...";
    }
}

/*****************************************************************************************\
*                          ::perf::TestBase::_declareHelper
\*****************************************************************************************/
TestBase::_declareHelper& TestBase::_declareHelper::iterations(unsigned int n)
{
    test->times.clear();
    test->times.reserve(n);
    test->nIters = std::min(n, TestBase::iterationsLimitDefault);
    test->currentIter = (unsigned int)-1;
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::time(double timeLimitSecs)
{
    test->times.clear();
    test->currentIter = (unsigned int)-1;
    test->timeLimit = (int64)(timeLimitSecs * cv::getTickFrequency());
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::tbb_threads(int n)
{
    cv::setNumThreads(n);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::runs(unsigned int runsNumber)
{
    test->runsPerIteration = runsNumber;
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::in(cv::InputOutputArray a1, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->inputData, a1, wtype);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::in(cv::InputOutputArray a1, cv::InputOutputArray a2, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->inputData, a1, wtype);
    TestBase::declareArray(test->inputData, a2, wtype);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::in(cv::InputOutputArray a1, cv::InputOutputArray a2, cv::InputOutputArray a3, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->inputData, a1, wtype);
    TestBase::declareArray(test->inputData, a2, wtype);
    TestBase::declareArray(test->inputData, a3, wtype);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::in(cv::InputOutputArray a1, cv::InputOutputArray a2, cv::InputOutputArray a3, cv::InputOutputArray a4, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->inputData, a1, wtype);
    TestBase::declareArray(test->inputData, a2, wtype);
    TestBase::declareArray(test->inputData, a3, wtype);
    TestBase::declareArray(test->inputData, a4, wtype);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::out(cv::InputOutputArray a1, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->outputData, a1, wtype);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::out(cv::InputOutputArray a1, cv::InputOutputArray a2, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->outputData, a1, wtype);
    TestBase::declareArray(test->outputData, a2, wtype);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::out(cv::InputOutputArray a1, cv::InputOutputArray a2, cv::InputOutputArray a3, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->outputData, a1, wtype);
    TestBase::declareArray(test->outputData, a2, wtype);
    TestBase::declareArray(test->outputData, a3, wtype);
    return *this;
}

TestBase::_declareHelper& TestBase::_declareHelper::out(cv::InputOutputArray a1, cv::InputOutputArray a2, cv::InputOutputArray a3, cv::InputOutputArray a4, int wtype)
{
    if (!test->times.empty()) return *this;
    TestBase::declareArray(test->outputData, a1, wtype);
    TestBase::declareArray(test->outputData, a2, wtype);
    TestBase::declareArray(test->outputData, a3, wtype);
    TestBase::declareArray(test->outputData, a4, wtype);
    return *this;
}

TestBase::_declareHelper::_declareHelper(TestBase* t) : test(t)
{
}

/*****************************************************************************************\
*                                  miscellaneous
\*****************************************************************************************/

namespace {
struct KeypointComparator
{
    std::vector<cv::KeyPoint>& pts_;
    comparators::KeypointGreater cmp;

    KeypointComparator(std::vector<cv::KeyPoint>& pts) : pts_(pts), cmp() {}

    bool operator()(int idx1, int idx2) const
    {
        return cmp(pts_[idx1], pts_[idx2]);
    }
private:
    const KeypointComparator& operator=(const KeypointComparator&); // quiet MSVC
};
}//namespace

void perf::sort(std::vector<cv::KeyPoint>& pts, cv::InputOutputArray descriptors)
{
    cv::Mat desc = descriptors.getMat();

    CV_Assert(pts.size() == (size_t)desc.rows);
    cv::AutoBuffer<int> idxs(desc.rows);

    for (int i = 0; i < desc.rows; ++i)
        idxs[i] = i;

    std::sort((int*)idxs, (int*)idxs + desc.rows, KeypointComparator(pts));

    std::vector<cv::KeyPoint> spts(pts.size());
    cv::Mat sdesc(desc.size(), desc.type());

    for(int j = 0; j < desc.rows; ++j)
    {
        spts[j] = pts[idxs[j]];
        cv::Mat row = sdesc.row(j);
        desc.row(idxs[j]).copyTo(row);
    }

    spts.swap(pts);
    sdesc.copyTo(desc);
}

/*****************************************************************************************\
*                                  ::perf::GpuPerf
\*****************************************************************************************/
#ifdef HAVE_CUDA
bool perf::GpuPerf::targetDevice()
{
    return !param_run_cpu;
}
#endif

/*****************************************************************************************\
*                                  ::perf::PrintTo
\*****************************************************************************************/
namespace perf
{

void PrintTo(const MatType& t, ::std::ostream* os)
{
    switch( CV_MAT_DEPTH((int)t) )
    {
        case CV_8U:  *os << "8U";  break;
        case CV_8S:  *os << "8S";  break;
        case CV_16U: *os << "16U"; break;
        case CV_16S: *os << "16S"; break;
        case CV_32S: *os << "32S"; break;
        case CV_32F: *os << "32F"; break;
        case CV_64F: *os << "64F"; break;
        case CV_USRTYPE1: *os << "USRTYPE1"; break;
        default: *os << "INVALID_TYPE"; break;
    }
    *os << 'C' << CV_MAT_CN((int)t);
}

} //namespace perf

/*****************************************************************************************\
*                                  ::cv::PrintTo
\*****************************************************************************************/
namespace cv {

void PrintTo(const Size& sz, ::std::ostream* os)
{
    *os << /*"Size:" << */sz.width << "x" << sz.height;
}

}  // namespace cv