Extract CCTags from multiple images in parallel

src/applications/detection/CmdLine.cpp

@@ -28,6 +28,7 @@ static const struct option longopts[] =
     {"sync",       no_argument,       0, 0xd0 },
     {"debug-dir",  required_argument, 0, 0xd1 },
     {"use-cuda",   no_argument,       0, 0xd2 },
+    {"parallel",   required_argument, 0, 0xd3 },
 #endif
     {0,0,0,0},
 };
@@ -42,6 +43,7 @@ CmdLine::CmdLine( )
     , _switchSync( false )
     , _debugDir( "" )
     , _useCuda( false )
+    , _parallel( 1 )
 #endif
 { }
 
@@ -72,6 +74,7 @@ bool CmdLine::parse( int argc, char* argv[] )
       case 0xd0 : _switchSync        = true;   break;
       case 0xd1 : _debugDir          = optarg; break;
       case 0xd2 : _useCuda           = true;   break;
+      case 0xd3 : _parallel          = strtol( optarg, NULL, 0 );   break;
 #endif
       default : break;
     }
@@ -111,6 +114,7 @@ void CmdLine::usage( const char* const argv0 )
           "           [--sync]\n"
           "           [--debug-dir <debugdir>]\n"
           "           [--use-cuda]\n"
+          "           [--parallel <n>]\n"
           "\n"
           "    <imgpath>  - path to an image (JPG, PNG) or video\n"
           "    <nbrings>  - number of rings of the CCTags to detect\n"
@@ -120,6 +124,7 @@ void CmdLine::usage( const char* const argv0 )
           "    --sync     - CUDA debug option, run all CUDA ops synchronously\n"
           "    <debugdir> - path storing image to debug intermediate GPU results\n"
           "    --use-cuda - select GPU code instead of CPU code\n"
+          "    --parallel - use <n> CUDA pipes concurrently (default 1)\n"
           "\n" << std::endl;
 }
 

src/applications/detection/CmdLine.hpp

@@ -23,6 +23,7 @@ class CmdLine
     bool        _switchSync;
     std::string _debugDir;
     bool        _useCuda;
+    int         _parallel;
 #endif
 
     CmdLine( );

src/applications/detection/main.cpp

@@ -41,6 +41,8 @@
 #include <fstream>
 #include <exception>
 
+#include <tbb/tbb.h>
+
 #define PRINT_TO_CERR
 
 using namespace cctag;
@@ -49,7 +51,13 @@ using boost::timer;
 using namespace boost::gil;
 namespace bfs = boost::filesystem;
 
-void detection(std::size_t frameId, const cv::Mat & src, const cctag::Parameters & params, const cctag::CCTagMarkersBank & bank, std::ostream & output, std::string debugFileName = "")
+void detection( std::size_t frameId,
+                int         pipeId,
+                const cv::Mat & src,
+                const cctag::Parameters & params,
+                const cctag::CCTagMarkersBank & bank,
+                std::ostream & output,
+                std::string debugFileName = "")
 {
 
     if (debugFileName == "") {
@@ -67,7 +75,7 @@ void detection(std::size_t frameId, const cv::Mat & src, const cctag::Parameters
     static cctag::logtime::Mgmt* durations = 0;
 
     //Call the main CCTag detection function
-    cctagDetection( markers, frameId , src, params, bank, true, durations );
+    cctagDetection( markers, pipeId, frameId , src, params, bank, true, durations );
 
     if( durations ) {
         durations->print( std::cerr );
@@ -208,10 +216,11 @@ int main(int argc, char** argv)
     cv::Mat graySrc;
     cv::cvtColor( src, graySrc, CV_BGR2GRAY );
 
+    const int pipeId = 0;
 #ifdef PRINT_TO_CERR
-    detection(0, graySrc, params, bank, std::cerr, myPath.stem().string());
+    detection(0, pipeId, graySrc, params, bank, std::cerr, myPath.stem().string());
 #else
-    detection(0, graySrc, params, bank, outputFile, myPath.stem().string());
+    detection(0, pipeId, graySrc, params, bank, outputFile, myPath.stem().string());
 #endif
 } else if (ext == ".avi" )
   {
@@ -259,10 +268,11 @@ int main(int argc, char** argv)
         //bitwise_not ( imgGray, imgGrayInverted );
 
         // Call the CCTag detection
+        const int pipeId = 0;
 #ifdef PRINT_TO_CERR
-        detection(frameId, *imgGray, params, bank, std::cerr, outFileName.str());
+        detection(frameId, pipeId, *imgGray, params, bank, std::cerr, outFileName.str());
 #else
-        detection(frameId, *imgGray, params, bank, outputFile, outFileName.str());
+        detection(frameId, pipeId, *imgGray, params, bank, outputFile, outFileName.str());
 #endif
         ++frameId; 
         if( frameId % 100 == 0 ) {
@@ -280,30 +290,38 @@ int main(int argc, char** argv)
 
     std::size_t frameId = 0;
 
+    std::map<int,bfs::path> files[2];
     for(const auto & fileInFolder : vFileInFolder) {
-      const std::string subExt(bfs::extension(fileInFolder));
+        files[frameId & 1].insert( std::pair<int,bfs::path>( frameId, fileInFolder ) );
+        frameId++;
+    }
+
+    tbb::parallel_for( 0, 2, [&](size_t fileListIdx) {
+      for(const auto & fileInFolder : files[fileListIdx]) {
+        const std::string subExt(bfs::extension(fileInFolder.second));
 
-      if ( (subExt == ".png") || (subExt == ".jpg") || (subExt == ".PNG") || (subExt == ".JPG") ) {
+        if ( (subExt == ".png") || (subExt == ".jpg") || (subExt == ".PNG") || (subExt == ".JPG") ) {
 
-        CCTAG_COUT( "Processing image " << fileInFolder.string() );
+          std::cerr << "Processing image " << fileInFolder.second.string() << std::endl;
 
-		cv::Mat src;
-    	src = cv::imread(fileInFolder.string());
+		  cv::Mat src;
+    	  src = cv::imread(fileInFolder.second.string());
 
-        cv::Mat imgGray;
-        cv::cvtColor( src, imgGray, CV_BGR2GRAY );
+          cv::Mat imgGray;
+          cv::cvtColor( src, imgGray, CV_BGR2GRAY );
 
-        // Call the CCTag detection
+          // Call the CCTag detection
+          int pipeId = ( fileInFolder.first & 1 );
 #ifdef PRINT_TO_CERR
-        detection(frameId, imgGray, params, bank, std::cerr, fileInFolder.stem().string());
+          detection(fileInFolder.first, pipeId, imgGray, params, bank, std::cerr, fileInFolder.second.stem().string());
 #else
-        detection(frameId, imgGray, params, bank, outputFile, fileInFolder.stem().string());
+          detection(fileInFolder.first, pipeId, imgGray, params, bank, outputFile, fileInFolder.second.stem().string());
 #endif
-++frameId;
+          std::cerr << "Done processing image " << fileInFolder.second.string() << std::endl;
+        }
       }
-    }
-  }else
-  {
+    } );
+  } else {
       throw std::logic_error("Unrecognized input.");
   }
   outputFile.close();

src/applications/regression/TestLog.cpp

@@ -26,7 +26,7 @@ FrameLog FrameLog::detect(size_t frame, const cv::Mat& src, const Parameters& pa
   CCTag::List markers;
 
   const auto t0 = high_resolution_clock::now();
-  cctagDetection(markers, frame, src, parameters, bank, true, nullptr);
+  cctagDetection(markers, 0, frame, src, parameters, bank, true, nullptr);
   const auto t1 = high_resolution_clock::now();
   const auto td = duration_cast<milliseconds>(t1 - t0).count() / 1000.f;
 

src/cctag/CCTag.cpp

@@ -110,14 +110,14 @@ void CCTag::scale(const float s)
 }
 
 #ifdef WITH_CUDA
-void CCTag::acquireNearbyPointMemory( )
+void CCTag::acquireNearbyPointMemory( int tagId )
 {
-    _cuda_result = popart::PinnedCounters::getPointPtr( __FILE__, __LINE__ );
+    _cuda_result = popart::PinnedCounters::getPointPtr( tagId, __FILE__, __LINE__ );
 }
 
-void CCTag::releaseNearbyPointMemory( )
+void CCTag::releaseNearbyPointMemory( int tagId )
 {
-    popart::PinnedCounters::releaseAllPoints();
+    popart::PinnedCounters::releaseAllPoints( tagId );
 }
 #endif
 

src/cctag/CCTag.hpp

@@ -353,7 +353,7 @@ class CCTag : public ICCTag
    *  Instead, releaseNearbyPointMemory() invalidates all such
    *  pointers in the process.
    */
-  void acquireNearbyPointMemory( );
+  void acquireNearbyPointMemory( int pipeId );
 
   inline popart::NearbyPoint* getNearbyPointBuffer( ) {
     return _cuda_result;
@@ -363,7 +363,7 @@ class CCTag : public ICCTag
    *  Invalidates pointers in all objects and in all threads in
    *  this process.
    */
-  static void releaseNearbyPointMemory( );
+  static void releaseNearbyPointMemory( int pipeId );
 #endif
 
   void serialize(boost::archive::text_oarchive & ar, const unsigned int version);

src/cctag/Detection.cpp

@@ -775,7 +775,9 @@ popart::TagPipe* initCuda( int      pipeId,
  * @param[in] bank CCTag bank.
  * @param[in] No longer used.
  */
-void cctagDetection(CCTag::List& markers,
+void cctagDetection(
+        CCTag::List& markers,
+        int          pipeId,
         const std::size_t frame, 
         const cv::Mat & imgGraySrc,
         const Parameters & providedParams,
@@ -807,7 +809,7 @@ void cctagDetection(CCTag::List& markers,
     popart::TagPipe* pipe1 = 0;
 #ifdef WITH_CUDA
     if( params._useCuda ) {
-        pipe1 = initCuda( 0,
+        pipe1 = initCuda( pipeId,
                           imgGraySrc.size().width,
 	                      imgGraySrc.size().height,
 	                      params,
@@ -820,7 +822,7 @@ void cctagDetection(CCTag::List& markers,
         assert( imgGraySrc.type() == CV_8U );
         unsigned char* pix = imgGraySrc.data;
 
-        pipe1->load( pix );
+        pipe1->load( frame, pix );
 
         if( durations ) {
             cudaDeviceSynchronize();
@@ -870,7 +872,7 @@ void cctagDetection(CCTag::List& markers,
         }
 
         for( CCTag& tag : markers ) {
-            tag.acquireNearbyPointMemory( );
+            tag.acquireNearbyPointMemory( pipe1->getId() );
         }
     }
 #endif // WITH_CUDA
@@ -913,13 +915,12 @@ void cctagDetection(CCTag::List& markers,
         if( pipe1 && numTags > 0 ) {
             pipe1->uploadCuts( numTags, vSelectedCuts, params );
 
-            // cerr << __FILE__ << ":"<< __LINE__ << " WARNING: uncontrolled creation of CUDA streams!!!!" << endl;
-            pipe1->makeCudaStreams( numTags );
-
             tagIndex = 0;
             int debug_num_calls = 0;
             for( CCTag& cctag : markers ) {
-                if( detected[tagIndex] == status::id_reliable ) {
+                if( vSelectedCuts[tagIndex].size() <= 2 ) {
+                    detected[tagIndex] = status::no_selected_cuts;
+                } else if( detected[tagIndex] == status::id_reliable ) {
                     if( debug_num_calls >= numTags ) {
                         cerr << __FILE__ << ":" << __LINE__ << " center finding for more loops (" << debug_num_calls << ") than uploaded (" << numTags << ")?" << endl;
                     }
@@ -969,7 +970,7 @@ void cctagDetection(CCTag::List& markers,
     if( pipe1 ) {
         /* Releasing all points in all threads in the process.
          */
-        CCTag::releaseNearbyPointMemory();
+        CCTag::releaseNearbyPointMemory( pipe1->getId() );
     }
 #endif
 

src/cctag/Detection.hpp

@@ -33,6 +33,7 @@ class EdgePointImage;
  * @brief Perform the CCTag detection on a gray scale image. Cf. application/detection/main.cpp for example of usage.
  * 
  * @param[out] markers Detected markers. WARNING: only markers with status == 1 are valid ones. (status available via getStatus()) 
+ * @param[in] pipeId Choose one of up to 3 parallel CUDA pipes
  * @param[in] frame A frame number. Can be anything (e.g. 0).
  * @param[in] imgGraySrc Gray scale input image.
  * @param[in] providedParams Contains all the parameters.
@@ -41,6 +42,7 @@ class EdgePointImage;
  */
 void cctagDetection(
         CCTag::List& markers,
+        int          pipeId,
         const std::size_t frame,
         const cv::Mat & graySrc,
         const Parameters & params,

src/cctag/ICCTag.cpp

@@ -24,6 +24,7 @@ namespace cctag {
  * @brief Perform the CCTag detection on a gray scale image
  * 
  * @param[out] markers Detected markers. WARNING: only markers with status == 1 are valid ones. (status available via getStatus()) 
+ * @param[in] pipeId Choose between several CUDA pipeline instances
  * @param[in] frame A frame number. Can be anything (e.g. 0).
  * @param[in] graySrc Gray scale input image.
  * @param[in] nRings Number of CCTag rings.
@@ -33,6 +34,7 @@ namespace cctag {
  */
 void cctagDetection(
       boost::ptr_list<ICCTag> & markers,
+      int                       pipeId,
       const std::size_t frame,
       const cv::Mat & graySrc,
       const std::size_t nRings,
@@ -63,11 +65,12 @@ void cctagDetection(
     bank = CCTagMarkersBank(cctagBankFilename);
   }
 
-  cctagDetection(markers, frame, graySrc, params, durations, &bank);
+  cctagDetection(markers, pipeId, frame, graySrc, params, durations, &bank);
 }
 
 void cctagDetection(
       boost::ptr_list<ICCTag> & markers,
+      int                       pipeId,
       const std::size_t frame,
       const cv::Mat & graySrc,
       const cctag::Parameters & params,
@@ -79,10 +82,10 @@ void cctagDetection(
   if ( pBank == NULL)
   {
     CCTagMarkersBank bank(params._nCrowns);
-    cctag::cctagDetection(cctags, frame, graySrc, params, bank, false, durations);
+    cctag::cctagDetection(cctags, pipeId, frame, graySrc, params, bank, false, durations);
   }else
   {
-    cctag::cctagDetection(cctags, frame, graySrc, params, *pBank, false, durations);
+    cctag::cctagDetection(cctags, pipeId, frame, graySrc, params, *pBank, false, durations);
   }
 
   markers.clear();

src/cctag/ICCTag.hpp

@@ -63,6 +63,7 @@ class ICCTag
  */
 void cctagDetection(
       boost::ptr_list<ICCTag> & markers,
+      int                       pipeId,
       const std::size_t frame,
       const cv::Mat & graySrc,
       logtime::Mgmt* durations = 0,
@@ -72,6 +73,7 @@ void cctagDetection(
 
 void cctagDetection(
       boost::ptr_list<ICCTag> & markers,
+      int                       pipeId,
       const std::size_t frame,
       const cv::Mat & graySrc,
       const cctag::Parameters & params,