updated performance sample

diff --git a/samples/gpu/performance/performance.cpp b/samples/gpu/performance/performance.cpp
index 133e72d..f827103 100644 (file)
--- a/samples/gpu/performance/performance.cpp
+++ b/samples/gpu/performance/performance.cpp
@@ -60,8 +60,11 @@ void TestSystem::finishCurrentSubtest()
         // There is no need to print subtest statistics\r
         return;\r
 \r
-    int cpu_time = static_cast<int>(cpu_elapsed_ / getTickFrequency() * 1000.0);\r
-    int gpu_time = static_cast<int>(gpu_elapsed_ / getTickFrequency() * 1000.0);\r
+    //int cpu_time = static_cast<int>(cpu_elapsed_ / getTickFrequency() * 1000.0);\r
+    //int gpu_time = static_cast<int>(gpu_elapsed_ / getTickFrequency() * 1000.0);\r
+\r
+    double cpu_time = cpu_elapsed_ / getTickFrequency() * 1000.0;\r
+    double gpu_time = gpu_elapsed_ / getTickFrequency() * 1000.0;\r
 \r
     double speedup = static_cast<double>(cpu_elapsed_) /\r
                      std::max((int64)1, gpu_elapsed_);\r
@@ -161,7 +164,8 @@ int main(int argc, const char* argv[])
        "{ f | filter  |       | filter for test }"\r
        "{ w | workdir |       | set working directory }"\r
        "{ l | list    | false | show all tests }"\r
-       "{ d  | device  |   0   | device id }";\r
+       "{ d | device  | 0     | device id }"\r
+       "{ i | iters   | 10    | iteration count }";\r
 \r
     CommandLineParser cmd(argc, argv, keys);\r
 \r
@@ -190,6 +194,7 @@ int main(int argc, const char* argv[])
     string filter = cmd.get<string>("filter");\r
     string workdir = cmd.get<string>("workdir");\r
     bool list = cmd.get<bool>("list");\r
+    int iters = cmd.get<int>("iters");\r
 \r
     if (!filter.empty())\r
         TestSystem::instance().setTestFilter(filter);\r
@@ -205,6 +210,8 @@ int main(int argc, const char* argv[])
     if (list)\r
         TestSystem::instance().setListMode(true);\r
 \r
+    TestSystem::instance().setIters(iters);\r
+\r
     TestSystem::instance().run();\r
 \r
     return 0;\r

diff --git a/samples/gpu/performance/performance.h b/samples/gpu/performance/performance.h
index f82fc12..e686c7d 100644 (file)
--- a/samples/gpu/performance/performance.h
+++ b/samples/gpu/performance/performance.h
@@ -4,6 +4,7 @@
 #include <iostream>\r
 #include <cstdio>\r
 #include <vector>\r
+#include <numeric>\r
 #include <string>\r
 #include "opencv2/core/core.hpp"\r
 #include "opencv2/gpu/gpu.hpp"\r
@@ -40,6 +41,8 @@ public:
     void setTestFilter(const std::string& val) { test_filter_ = val; }\r
     const std::string& testFilter() const { return test_filter_; }\r
 \r
+    void setIters(int iters) { iters_ = iters; }\r
+\r
     void addInit(Runnable* init) { inits_.push_back(init); }\r
     void addTest(Runnable* test) { tests_.push_back(test); }\r
     void run();\r
@@ -53,20 +56,36 @@ public:
         return cur_subtest_description_;\r
     }\r
 \r
+    bool stop() const { return it_ >= iters_; }\r
+\r
     void cpuOn() { cpu_started_ = cv::getTickCount(); }\r
     void cpuOff() \r
     {\r
         int64 delta = cv::getTickCount() - cpu_started_;\r
-        cpu_elapsed_ += delta;\r
+        cpu_times_.push_back(delta);\r
+        ++it_;\r
+    }\r
+    void cpuComplete()\r
+    {\r
+        double delta_mean = std::accumulate(cpu_times_.begin(), cpu_times_.end(), 0.0) / iters_;\r
+        cpu_elapsed_ += delta_mean;\r
         cur_subtest_is_empty_ = false;\r
-    }  \r
+        it_ = 0;\r
+    }\r
 \r
     void gpuOn() { gpu_started_ = cv::getTickCount(); }\r
     void gpuOff() \r
     {\r
         int64 delta = cv::getTickCount() - gpu_started_;\r
-        gpu_elapsed_ += delta;\r
+        gpu_times_.push_back(delta);\r
+        ++it_;\r
+    }\r
+    void gpuComplete()\r
+    {\r
+        double delta_mean = std::accumulate(gpu_times_.begin(), gpu_times_.end(), 0.0) / iters_;\r
+        gpu_elapsed_ += delta_mean;\r
         cur_subtest_is_empty_ = false;\r
+        it_ = 0;\r
     }\r
 \r
     bool isListMode() const { return is_list_mode_; }\r
@@ -76,7 +95,13 @@ private:
     TestSystem(): cur_subtest_is_empty_(true), cpu_elapsed_(0),\r
                   gpu_elapsed_(0), speedup_total_(0.0),\r
                   num_subtests_called_(0),\r
-                  is_list_mode_(false) {}\r
+                  is_list_mode_(false) \r
+    {\r
+        iters_ = 10;\r
+        it_ = 0;\r
+        cpu_times_.reserve(iters_);\r
+        gpu_times_.reserve(iters_);\r
+    }\r
 \r
     void finishCurrentSubtest();\r
     void resetCurrentSubtest() \r
@@ -85,6 +110,9 @@ private:
         gpu_elapsed_ = 0;\r
         cur_subtest_description_.str("");\r
         cur_subtest_is_empty_ = true;\r
+        it_ = 0;\r
+        cpu_times_.clear();\r
+        gpu_times_.clear();\r
     }\r
 \r
     void printHeading();\r
@@ -107,6 +135,11 @@ private:
     int num_subtests_called_;\r
 \r
     bool is_list_mode_;\r
+\r
+    int iters_;\r
+    int it_;\r
+    std::vector<int64> cpu_times_;\r
+    std::vector<int64> gpu_times_;\r
 };\r
 \r
 \r
@@ -130,10 +163,12 @@ private:
     void name##_test::run()\r
 \r
 #define SUBTEST TestSystem::instance().startNewSubtest()\r
-#define CPU_ON TestSystem::instance().cpuOn()\r
-#define GPU_ON TestSystem::instance().gpuOn()\r
-#define CPU_OFF TestSystem::instance().cpuOff()\r
-#define GPU_OFF TestSystem::instance().gpuOff()\r
+\r
+#define CPU_ON while (!TestSystem::instance().stop()) { TestSystem::instance().cpuOn()\r
+#define CPU_OFF TestSystem::instance().cpuOff(); } TestSystem::instance().cpuComplete()\r
+\r
+#define GPU_ON while (!TestSystem::instance().stop()) { TestSystem::instance().gpuOn()\r
+#define GPU_OFF TestSystem::instance().gpuOff(); } TestSystem::instance().gpuComplete()\r
 \r
 // Generates matrix\r
 void gen(cv::Mat& mat, int rows, int cols, int type, cv::Scalar low, \r

diff --git a/samples/gpu/performance/tests.cpp b/samples/gpu/performance/tests.cpp
index 0c012fd..1188b95 100644 (file)
--- a/samples/gpu/performance/tests.cpp
+++ b/samples/gpu/performance/tests.cpp
@@ -767,6 +767,27 @@ TEST(threshold)
         gpu::threshold(d_src, d_dst, 50.0, 0.0, THRESH_BINARY);\r
         GPU_OFF;\r
     }\r
+\r
+    for (int size = 2000; size <= 4000; size += 1000)\r
+    {\r
+        SUBTEST << size << 'x' << size << ", 32FC1, THRESH_TRUNC [NPP]";\r
+\r
+        gen(src, size, size, CV_32FC1, 0, 100);\r
+\r
+        threshold(src, dst, 50.0, 0.0, THRESH_TRUNC);\r
+\r
+        CPU_ON; \r
+        threshold(src, dst, 50.0, 0.0, THRESH_TRUNC);\r
+        CPU_OFF;\r
+\r
+        d_src.upload(src);\r
+\r
+        gpu::threshold(d_src, d_dst, 50.0, 0.0, THRESH_TRUNC);\r
+\r
+        GPU_ON;\r
+        gpu::threshold(d_src, d_dst, 50.0, 0.0, THRESH_TRUNC);\r
+        GPU_OFF;\r
+    }\r
 }\r
 \r
 TEST(pow)\r