add scope time logger w/ ACF_LOG_GPU_TIME option (hack benchmark in u…

…nit test)

bin/build-xcode.sh

@@ -1,14 +1,17 @@
 #!/bin/bash
 
 TOOLCHAIN=xcode
-CONFIG=MinSizeRel
+CONFIG=Release
 
 ARGS=(
     --verbose
     --toolchain ${TOOLCHAIN}
     --config ${CONFIG}
     --fwd HUNTER_CONFIGURATION_TYPES=${CONFIG}
+    ACF_USE_DRISHTI_CACHE=OFF
+    ACF_BUILD_TESTS=ON
     --jobs 8
 )
 
+#build.py ${ARGS[@]} --install --reconfig ${*}
 build.py ${ARGS[@]} --install ${*}

src/lib/acf/ut/test-acf.cpp

@@ -32,7 +32,6 @@ int gauze_main(int argc, char** argv)
 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 #define ACF_TEST_DISPLAY_OUTPUT 0
-#define ACF_TEST_WARM_UP_GPU 0 // for timing only
 
 // clang-format off
 #if defined(ACF_DO_GPU)
@@ -48,6 +47,10 @@ int gauze_main(int argc, char** argv)
 // clang-format on
 #include "io/cereal_pba.h"
 
+#include "util/ScopeTimeLogger.h"
+
+#define ACF_LOG_GPU_TIME 0
+
 // http://uscilab.github.io/cereal/serialization_archives.html
 #include <cereal/archives/portable_binary.hpp>
 #include <cereal/types/vector.hpp>
@@ -231,35 +234,32 @@ class ACFTest : public ::testing::Test
         static const bool doGrayscale = false;
         ogles_gpgpu::Size2d inputSize(image.cols, image.rows);
 
-        m_acf = std::make_shared<ogles_gpgpu::ACF>(nullptr, inputSize, sizes, ogles_gpgpu::ACF::kLUVM012345, doGrayscale, false);
+        m_acf = std::make_shared<ogles_gpgpu::ACF>(nullptr, inputSize, sizes, ogles_gpgpu::ACF::kM012345, doGrayscale, false);
         m_acf->setRotation(0);
+        m_acf->setDoLuvTransfer(false);
 
         cv::Mat input = image;
-#if ACF_TEST_WARM_UP_GPU
+
+#if ACF_LOG_GPU_TIME
+        // Warm up the GPU
         for (int i = 0; i < 10; i++)
         {
             (*m_acf)({ input.cols, input.rows }, input.ptr(), true, 0, DFLT_TEXTURE_FORMAT);
         }
+        m_acf->fill(Pgpu, Pcpu); // be sure to read and flush all commands            
 #endif
-
-        (*m_acf)({ input.cols, input.rows }, input.ptr(), true, 0, DFLT_TEXTURE_FORMAT);
-        m_acf->fill(Pgpu, Pcpu);
-
         {
-            // This code block is a place holder for 7 channel ACF output, which conveniently fits in 2 textures
-            // for faster transfers on low performing Android devices.  Currently there is no 7 channel ACF
-            // classifier/detector, so this is used as a place holder to illustrate the raw channel extraction
-            // and pyramid formatting until equivalent CPU formatting is in place.
-            auto acf7 = std::make_shared<ogles_gpgpu::ACF>(nullptr, inputSize, sizes, ogles_gpgpu::ACF::kM012345, doGrayscale, false);
-            (*acf7)({ input.cols, input.rows }, input.ptr(), true, 0, DFLT_TEXTURE_FORMAT);
-
-            acf::Detector::Pyramid Pgpu7;
-            acf7->fill(Pgpu7, Pcpu);
-
-            //cv::imshow("Pgpu7", draw(Pgpu7);
-            //cv::imshow("LM012345", acf7->getChannels());
-            //cv::imshow("LUVM012345", m_acf->getChannels());
-            //cv::waitKey(0);
+
+#if ACF_LOG_GPU_TIME
+            util::ScopeTimeLogger logger = [&](double elapsed)
+            {
+                std::cout << "ACF::fill(): " << input.cols << " " << elapsed << std::endl;
+            };
+#endif
+            // Fill in the pyramid:
+            (*m_acf)({ input.cols, input.rows }, input.ptr(), true, 0, DFLT_TEXTURE_FORMAT);
+            glFlush();
+            m_acf->fill(Pgpu, Pcpu);
         }
     }
 
@@ -450,7 +450,15 @@ TEST_F(ACFTest, ACFDetectionGPU10)
 
     std::vector<double> scores;
     std::vector<cv::Rect> objects;
-    (*m_detector)(Pgpu, objects);
+    {
+#if ACF_LOG_GPU_TIME
+        util::ScopeTimeLogger logger = [](double elapsed)
+        {
+            std::cout << "acf::Detector::operator():" << elapsed << std::endl;
+        };
+#endif
+        (*m_detector)(Pgpu, objects);
+    }
 
 #if ACF_TEST_DISPLAY_OUTPUT
     WaitKey waitKey;

src/lib/util/ScopeTimeLogger.h

@@ -0,0 +1,63 @@
+/*! -*-c++-*-
+  @file   timing.h
+  @author David Hirvonen
+  @brief  Scope based timing macro.
+
+  \copyright Copyright 2014-2017 Elucideye, Inc. All rights reserved.
+  \license{This project is released under the 3 Clause BSD License.}
+
+*/
+
+#ifndef __util_core_ScopeTimeLogger_h__
+#define __util_core_ScopeTimeLogger_h__
+
+#include "util/acf_util.h"
+
+#include <chrono>
+#include <functional>
+
+UTIL_NAMESPACE_BEGIN
+
+class ScopeTimeLogger
+{
+    using HighResolutionClock = std::chrono::high_resolution_clock;
+    using TimePoint = HighResolutionClock::time_point;
+
+public:
+    template <class Callable>
+    ScopeTimeLogger(Callable&& logger)
+        : m_logger(std::forward<Callable>(logger))
+    {
+        m_tic = HighResolutionClock::now();
+    }
+
+    ScopeTimeLogger(ScopeTimeLogger&& other)
+        : m_logger(std::move(other.m_logger))
+        , m_tic(std::move(other.m_tic))
+    {
+    }
+
+    ~ScopeTimeLogger()
+    {
+        auto now = HighResolutionClock::now();
+        m_logger(timeDifference(now, m_tic));
+    }
+
+    ScopeTimeLogger(const ScopeTimeLogger&) = delete;
+    void operator=(const ScopeTimeLogger&) = delete;
+
+    static double timeDifference(const TimePoint& a, const TimePoint& b)
+    {
+        return std::chrono::duration_cast<std::chrono::duration<double>>(a - b).count();
+    }
+
+    const TimePoint& getTime() const { return m_tic; }
+
+protected:
+    std::function<void(double)> m_logger;
+    TimePoint m_tic;
+};
+
+UTIL_NAMESPACE_END
+
+#endif // __util_core_ScopeTimeLogger_h__