modules/softcascade/perf/perf_cuda_softcascade.cpp

   1 #include "perf_precomp.hpp"
   2
   3 using std::tr1::get;
   4
   5 #define SC_PERF_TEST_P(fixture, name, params)  \
   6     class fixture##_##name : public fixture {\
   7      public:\
   8       fixture##_##name() {}\
   9      protected:\
  10         virtual void __cpu();\
  11         virtual void __gpu();\
  12       virtual void PerfTestBody();\
  13     };\
  14     TEST_P(fixture##_##name, name /*perf*/){ RunPerfTestBody(); }\
  15     INSTANTIATE_TEST_CASE_P(/*none*/, fixture##_##name, params);\
  16     void fixture##_##name::PerfTestBody() { if (PERF_RUN_CUDA()) __gpu(); else __cpu(); }
  17
  18 #define RUN_CPU(fixture, name)\
  19     void fixture##_##name::__cpu()
  20
  21 #define RUN_CUDA(fixture, name)\
  22     void fixture##_##name::__gpu()
  23
  24 #define NO_CPU(fixture, name)\
  25 void fixture##_##name::__cpu() { FAIL() << "No such CPU implementation analogy";}
  26
  27 namespace {
  28     struct DetectionLess
  29     {
  30         bool operator()(const cv::softcascade::Detection& a,
  31             const cv::softcascade::Detection& b) const
  32         {
  33             if (a.x != b.x)      return a.x < b.x;
  34             else if (a.y != b.y) return a.y < b.y;
  35             else if (a.w != b.w) return a.w < b.w;
  36             else return a.h < b.h;
  37         }
  38     };
  39
  40     cv::Mat sortDetections(cv::cuda::GpuMat& objects)
  41     {
  42         cv::Mat detections(objects);
  43
  44         typedef cv::softcascade::Detection Detection;
  45         Detection* begin = (Detection*)(detections.ptr<char>(0));
  46         Detection* end = (Detection*)(detections.ptr<char>(0) + detections.cols);
  47         std::sort(begin, end, DetectionLess());
  48
  49         return detections;
  50     }
  51 }
  52
  53
  54 typedef std::tr1::tuple<std::string, std::string> fixture_t;
  55 typedef perf::TestBaseWithParam<fixture_t> SCascadeTest;
  56
  57 SC_PERF_TEST_P(SCascadeTest, detect,
  58     testing::Combine(
  59         testing::Values(std::string("cv/cascadeandhog/cascades/inria_caltech-17.01.2013.xml"),
  60                         std::string("cv/cascadeandhog/cascades/sc_cvpr_2012_to_opencv_new_format.xml")),
  61         testing::Values(std::string("cv/cascadeandhog/images/image_00000000_0.png"))))
  62
  63 RUN_CUDA(SCascadeTest, detect)
  64 {
  65     cv::Mat cpu = cv::imread(getDataPath(get<1>(GetParam())));;
  66     ASSERT_FALSE(cpu.empty());
  67     cv::cuda::GpuMat colored(cpu);
  68
  69     cv::softcascade::SCascade cascade;
  70
  71     cv::FileStorage fs(getDataPath(get<0>(GetParam())), cv::FileStorage::READ);
  72     ASSERT_TRUE(fs.isOpened());
  73
  74     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
  75
  76     cv::cuda::GpuMat objectBoxes(1, 10000 * sizeof(cv::softcascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1);
  77     rois.setTo(1);
  78
  79     cascade.detect(colored, rois, objectBoxes);
  80
  81     TEST_CYCLE()
  82     {
  83         cascade.detect(colored, rois, objectBoxes);
  84     }
  85
  86     SANITY_CHECK(sortDetections(objectBoxes));
  87 }
  88
  89 NO_CPU(SCascadeTest, detect)
  90
  91 static cv::Rect getFromTable(int idx)
  92 {
  93     static const cv::Rect rois[] =
  94     {
  95         cv::Rect( 65 * 4,  20 * 4,  35 * 4, 80 * 4),
  96         cv::Rect( 95 * 4,  35 * 4,  45 * 4, 40 * 4),
  97         cv::Rect( 45 * 4,  35 * 4,  45 * 4, 40 * 4),
  98         cv::Rect( 25 * 4,  27 * 4,  50 * 4, 45 * 4),
  99         cv::Rect(100 * 4,  50 * 4,  45 * 4, 40 * 4),
 100
 101         cv::Rect( 60 * 4,  30 * 4,  45 * 4, 40 * 4),
 102         cv::Rect( 40 * 4,  55 * 4,  50 * 4, 40 * 4),
 103         cv::Rect( 48 * 4,  37 * 4,  72 * 4, 80 * 4),
 104         cv::Rect( 48 * 4,  32 * 4,  85 * 4, 58 * 4),
 105         cv::Rect( 48 * 4,   0 * 4,  32 * 4, 27 * 4)
 106     };
 107
 108     return rois[idx];
 109 }
 110
 111 typedef std::tr1::tuple<std::string, std::string, int> roi_fixture_t;
 112 typedef perf::TestBaseWithParam<roi_fixture_t> SCascadeTestRoi;
 113
 114 SC_PERF_TEST_P(SCascadeTestRoi, detectInRoi,
 115     testing::Combine(
 116         testing::Values(std::string("cv/cascadeandhog/cascades/inria_caltech-17.01.2013.xml"),
 117                         std::string("cv/cascadeandhog/cascades/sc_cvpr_2012_to_opencv_new_format.xml")),
 118         testing::Values(std::string("cv/cascadeandhog/images/image_00000000_0.png")),
 119         testing::Range(0, 5)))
 120
 121 RUN_CUDA(SCascadeTestRoi, detectInRoi)
 122 {
 123     cv::Mat cpu = cv::imread(getDataPath(get<1>(GetParam())));
 124     ASSERT_FALSE(cpu.empty());
 125     cv::cuda::GpuMat colored(cpu);
 126
 127     cv::softcascade::SCascade cascade;
 128
 129     cv::FileStorage fs(getDataPath(get<0>(GetParam())), cv::FileStorage::READ);
 130     ASSERT_TRUE(fs.isOpened());
 131
 132     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
 133
 134     cv::cuda::GpuMat objectBoxes(1, 16384 * 20, CV_8UC1), rois(colored.size(), CV_8UC1);
 135     rois.setTo(0);
 136
 137     int nroi = get<2>(GetParam());
 138     cv::RNG rng;
 139     for (int i = 0; i < nroi; ++i)
 140     {
 141         cv::Rect r = getFromTable(rng(10));
 142         cv::cuda::GpuMat sub(rois, r);
 143         sub.setTo(1);
 144     }
 145
 146     cascade.detect(colored, rois, objectBoxes);
 147
 148     TEST_CYCLE()
 149     {
 150         cascade.detect(colored, rois, objectBoxes);
 151     }
 152
 153     SANITY_CHECK(sortDetections(objectBoxes));
 154 }
 155
 156 NO_CPU(SCascadeTestRoi, detectInRoi)
 157
 158
 159 SC_PERF_TEST_P(SCascadeTestRoi, detectEachRoi,
 160     testing::Combine(
 161         testing::Values(std::string("cv/cascadeandhog/cascades/inria_caltech-17.01.2013.xml"),
 162                         std::string("cv/cascadeandhog/cascades/sc_cvpr_2012_to_opencv_new_format.xml")),
 163         testing::Values(std::string("cv/cascadeandhog/images/image_00000000_0.png")),
 164         testing::Range(0, 10)))
 165
 166 RUN_CUDA(SCascadeTestRoi, detectEachRoi)
 167 {
 168     cv::Mat cpu = cv::imread(getDataPath(get<1>(GetParam())));
 169     ASSERT_FALSE(cpu.empty());
 170     cv::cuda::GpuMat colored(cpu);
 171
 172     cv::softcascade::SCascade cascade;
 173
 174     cv::FileStorage fs(getDataPath(get<0>(GetParam())), cv::FileStorage::READ);
 175     ASSERT_TRUE(fs.isOpened());
 176
 177     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
 178
 179     cv::cuda::GpuMat objectBoxes(1, 16384 * 20, CV_8UC1), rois(colored.size(), CV_8UC1);
 180     rois.setTo(0);
 181
 182     int idx = get<2>(GetParam());
 183     cv::Rect r = getFromTable(idx);
 184     cv::cuda::GpuMat sub(rois, r);
 185     sub.setTo(1);
 186
 187     cascade.detect(colored, rois, objectBoxes);
 188
 189     TEST_CYCLE()
 190     {
 191         cascade.detect(colored, rois, objectBoxes);
 192     }
 193
 194     SANITY_CHECK(sortDetections(objectBoxes));
 195 }
 196
 197 NO_CPU(SCascadeTestRoi, detectEachRoi)
 198
 199 SC_PERF_TEST_P(SCascadeTest, detectStream,
 200     testing::Combine(
 201         testing::Values(std::string("cv/cascadeandhog/cascades/inria_caltech-17.01.2013.xml"),
 202                         std::string("cv/cascadeandhog/cascades/sc_cvpr_2012_to_opencv_new_format.xml")),
 203         testing::Values(std::string("cv/cascadeandhog/images/image_00000000_0.png"))))
 204
 205 RUN_CUDA(SCascadeTest, detectStream)
 206 {
 207     cv::Mat cpu = cv::imread(getDataPath(get<1>(GetParam())));
 208     ASSERT_FALSE(cpu.empty());
 209     cv::cuda::GpuMat colored(cpu);
 210
 211     cv::softcascade::SCascade cascade;
 212
 213     cv::FileStorage fs(getDataPath(get<0>(GetParam())), cv::FileStorage::READ);
 214     ASSERT_TRUE(fs.isOpened());
 215
 216     ASSERT_TRUE(cascade.load(fs.getFirstTopLevelNode()));
 217
 218     cv::cuda::GpuMat objectBoxes(1, 10000 * sizeof(cv::softcascade::Detection), CV_8UC1), rois(colored.size(), CV_8UC1);
 219     rois.setTo(1);
 220
 221     cv::cuda::Stream s;
 222
 223     cascade.detect(colored, rois, objectBoxes, s);
 224
 225     TEST_CYCLE()
 226     {
 227         cascade.detect(colored, rois, objectBoxes, s);
 228     }
 229
 230     s.waitForCompletion();
 231     SANITY_CHECK(sortDetections(objectBoxes));
 232 }
 233
 234 NO_CPU(SCascadeTest, detectStream)
 235
 236 #undef SC_PERF_TEST_P