Merge pull request opencv#8204 from terfendail:ovx_tlcontext

3rdparty/openvx/hal/openvx_hal.cpp

@@ -52,8 +52,17 @@ struct Tick
 
 inline ivx::Context& getOpenVXHALContext()
 {
-    // not thread safe
-    static ivx::Context instance = ivx::Context::create();
+#if __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1800)
+    //CXX11
+    static thread_local ivx::Context instance = ivx::Context::create();
+#else //__cplusplus >= 201103L || _MSC_VER >= 1800
+    //CXX98
+#ifdef WIN32
+        static __declspec(thread) ivx::Context instance = ivx::Context::create();
+#else
+        static __thread ivx::Context instance = ivx::Context::create();
+#endif
+#endif
     return instance;
 }
 

modules/core/include/opencv2/core/openvx/ovx_defs.hpp

@@ -20,6 +20,12 @@
 #define IVX_USE_OPENCV
 #include "ivx.hpp"
 
+namespace cv{
+namespace ovx{
+// Get common thread local OpenVX context
+CV_EXPORTS_W ivx::Context& getOpenVXContext();
+}}
+
 #define CV_OVX_RUN(condition, func, ...)          \
     if (cv::useOpenVX() && (condition) && func)   \
     {                                             \

modules/core/include/opencv2/core/utility.hpp

@@ -627,6 +627,9 @@ class CV_EXPORTS TLSDataContainer
     virtual void  deleteDataInstance(void* pData) const = 0;
 
     int key_;
+
+public:
+    void cleanup(); //! Release created TLS data container objects. It is similar to release() call, but it keeps TLS container valid.
 };
 
 // Main TLS data class
@@ -638,13 +641,15 @@ class TLSData : protected TLSDataContainer
     inline ~TLSData()       { release();            } // Release key and delete associated data
     inline T* get() const   { return (T*)getData(); } // Get data associated with key
 
-     // Get data from all threads
+    // Get data from all threads
     inline void gather(std::vector<T*> &data) const
     {
         std::vector<void*> &dataVoid = reinterpret_cast<std::vector<void*>&>(data);
         gatherData(dataVoid);
     }
 
+    inline void cleanup() { TLSDataContainer::cleanup(); }
+
 private:
     virtual void* createDataInstance() const {return new T;}                // Wrapper to allocate data by template
     virtual void  deleteDataInstance(void* pData) const {delete (T*)pData;} // Wrapper to release data by template

modules/core/src/convert.cpp

@@ -4673,7 +4673,7 @@ static bool _openvx_cvt(const T* src, size_t sstep,
 
     try
     {
-        Context context = Context::create();
+        Context context = ovx::getOpenVXContext();
 
         // Other conversions are marked as "experimental"
         if(context.vendorID() == VX_ID_KHRONOS &&
@@ -5406,7 +5406,7 @@ static bool openvx_LUT(Mat src, Mat dst, Mat _lut)
 
     try
     {
-        ivx::Context ctx = ivx::Context::create();
+        ivx::Context ctx = ovx::getOpenVXContext();
 
         ivx::Image
             ia = ivx::Image::createFromHandle(ctx, VX_DF_IMAGE_U8,

modules/core/src/ovx.cpp

@@ -14,6 +14,38 @@
 namespace cv
 {
 
+namespace ovx
+{
+#ifdef HAVE_OPENVX
+
+// Simple TLSData<ivx::Context> doesn't work, because default constructor doesn't create any OpenVX context.
+struct OpenVXTLSData
+{
+    OpenVXTLSData() : ctx(ivx::Context::create()) {}
+    ivx::Context ctx;
+};
+
+static TLSData<OpenVXTLSData>& getOpenVXTLSData()
+{
+    CV_SINGLETON_LAZY_INIT_REF(TLSData<OpenVXTLSData>, new TLSData<OpenVXTLSData>())
+}
+
+struct OpenVXCleanupFunctor
+{
+    ~OpenVXCleanupFunctor() { getOpenVXTLSData().cleanup(); }
+};
+static OpenVXCleanupFunctor g_openvx_cleanup_functor;
+
+ivx::Context& getOpenVXContext()
+{
+    return getOpenVXTLSData().get()->ctx;
+}
+
+#endif
+
+} // namespace
+
+
 bool haveOpenVX()
 {
 #ifdef HAVE_OPENVX
@@ -22,7 +54,7 @@ bool haveOpenVX()
     {
         try
         {
-        ivx::Context context = ivx::Context::create();
+        ivx::Context context = ovx::getOpenVXContext();
         vx_uint16 vComp = ivx::compiledWithVersion();
         vx_uint16 vCurr = context.version();
         g_haveOpenVX =

modules/core/src/stat.cpp

@@ -1665,7 +1665,7 @@ namespace cv
 
         try
         {
-            ivx::Context ctx = ivx::Context::create();
+            ivx::Context ctx = ovx::getOpenVXContext();
 #ifndef VX_VERSION_1_1
             if (ctx.vendorID() == VX_ID_KHRONOS)
                 return false; // Do not use OpenVX meanStdDev estimation for sample 1.0.1 implementation due to lack of accuracy
@@ -2312,7 +2312,7 @@ static bool openvx_minMaxIdx(Mat &src, double* minVal, double* maxVal, int* minI
 
     try
     {
-        ivx::Context ctx = ivx::Context::create();
+        ivx::Context ctx = ovx::getOpenVXContext();
         ivx::Image
             ia = ivx::Image::createFromHandle(ctx, stype == CV_8UC1 ? VX_DF_IMAGE_U8 : VX_DF_IMAGE_S16,
                 ivx::Image::createAddressing(cols, rows, stype == CV_8UC1 ? 1 : 2, (vx_int32)(src.step[0])), src.ptr());

modules/core/src/system.cpp

@@ -1086,7 +1086,7 @@ class TlsStorage
     }
 
     // Release TLS storage index and pass associated data to caller
-    void releaseSlot(size_t slotIdx, std::vector<void*> &dataVec)
+    void releaseSlot(size_t slotIdx, std::vector<void*> &dataVec, bool keepSlot = false)
     {
         AutoLock guard(mtxGlobalAccess);
         CV_Assert(tlsSlots.size() > slotIdx);
@@ -1099,12 +1099,13 @@ class TlsStorage
                 if (thread_slots.size() > slotIdx && thread_slots[slotIdx])
                 {
                     dataVec.push_back(thread_slots[slotIdx]);
-                    threads[i]->slots[slotIdx] = 0;
+                    thread_slots[slotIdx] = NULL;
                 }
             }
         }
 
-        tlsSlots[slotIdx] = 0;
+        if (!keepSlot)
+            tlsSlots[slotIdx] = 0;
     }
 
     // Get data by TLS storage index
@@ -1196,9 +1197,18 @@ void TLSDataContainer::release()
     std::vector<void*> data;
     data.reserve(32);
     getTlsStorage().releaseSlot(key_, data); // Release key and get stored data for proper destruction
+    key_ = -1;
+    for(size_t i = 0; i < data.size(); i++)  // Delete all associated data
+        deleteDataInstance(data[i]);
+}
+
+void TLSDataContainer::cleanup()
+{
+    std::vector<void*> data;
+    data.reserve(32);
+    getTlsStorage().releaseSlot(key_, data, true); // Extract stored data with removal from TLS tables
     for(size_t i = 0; i < data.size(); i++)  // Delete all associated data
         deleteDataInstance(data[i]);
-    key_ = -1;
 }
 
 void* TLSDataContainer::getData() const

modules/features2d/src/fast.cpp

@@ -354,7 +354,7 @@ static bool openvx_FAST(InputArray _img, std::vector<KeyPoint>& keypoints,
 
     try
     {
-        Context context = Context::create();
+        Context context = ovx::getOpenVXContext();
         Image img = Image::createFromHandle(context, Image::matTypeToFormat(imgMat.type()),
                                             Image::createAddressing(imgMat), (void*)imgMat.data);
         ivx::Scalar threshold = ivx::Scalar::create<VX_TYPE_FLOAT32>(context, _threshold);

modules/imgproc/src/accum.cpp

@@ -1958,7 +1958,7 @@ static bool openvx_accumulate(InputArray _src, InputOutputArray _dst, InputArray
 
     try
     {
-        ivx::Context context = ivx::Context::create();
+        ivx::Context context = ovx::getOpenVXContext();
         ivx::Image srcImage = ivx::Image::createFromHandle(context, ivx::Image::matTypeToFormat(srcMat.type()),
                                                            ivx::Image::createAddressing(srcMat), srcMat.data);
         ivx::Image dstImage = ivx::Image::createFromHandle(context, ivx::Image::matTypeToFormat(dstMat.type()),

modules/imgproc/src/canny.cpp

@@ -782,7 +782,7 @@ static bool openvx_canny(const Mat& src, Mat& dst, int loVal, int hiVal, int kSi
 {
     using namespace ivx;
 
-    Context context = Context::create();
+    Context context = ovx::getOpenVXContext();
     try
     {
     Image _src = Image::createFromHandle(

modules/imgproc/src/deriv.cpp

@@ -236,7 +236,7 @@ namespace cv
 
         try
         {
-            ivx::Context ctx = ivx::Context::create();
+            ivx::Context ctx = ovx::getOpenVXContext();
             if ((vx_size)ksize > ctx.convolutionMaxDimension())
                 return false;
 

modules/imgproc/src/featureselect.cpp

@@ -286,7 +286,7 @@ static bool openvx_harris(Mat image, OutputArray _corners,
 
     try
     {
-        Context context = Context::create();
+        Context context = ovx::getOpenVXContext();
 
         Image ovxImage = Image::createFromHandle(context, Image::matTypeToFormat(image.type()),
                                                  Image::createAddressing(image), image.data);

modules/imgproc/src/histogram.cpp

@@ -1282,7 +1282,7 @@ namespace cv
 
         try
         {
-            ivx::Context ctx = ivx::Context::create();
+            ivx::Context ctx = ovx::getOpenVXContext();
 #if VX_VERSION <= VX_VERSION_1_0
             if (ctx.vendorID() == VX_ID_KHRONOS && (range % histSize))
                 return false;
@@ -3773,7 +3773,7 @@ static bool openvx_equalize_hist(Mat srcMat, Mat dstMat)
 
     try
     {
-        Context context = Context::create();
+        Context context = ovx::getOpenVXContext();
         Image srcImage = Image::createFromHandle(context, Image::matTypeToFormat(srcMat.type()),
                                                  Image::createAddressing(srcMat), srcMat.data);
         Image dstImage = Image::createFromHandle(context, Image::matTypeToFormat(dstMat.type()),

modules/imgproc/src/imgwarp.cpp

@@ -4795,7 +4795,7 @@ static bool openvx_remap(Mat src, Mat dst, Mat map1, Mat map2, int interpolation
 
     try
     {
-        ivx::Context ctx = ivx::Context::create();
+        ivx::Context ctx = ovx::getOpenVXContext();
 
         Mat a;
         if (dst.data != src.data)

modules/imgproc/src/pyramids.cpp

@@ -1290,7 +1290,7 @@ static bool openvx_pyrDown( InputArray _src, OutputArray _dst, const Size& _dsz,
 
     try
     {
-        Context context = Context::create();
+        Context context = ovx::getOpenVXContext();
         if(context.vendorID() == VX_ID_KHRONOS)
         {
             // This implementation performs floor-like rounding

modules/imgproc/src/smooth.cpp

@@ -1677,7 +1677,7 @@ namespace cv
 
         try
         {
-            ivx::Context ctx = ivx::Context::create();
+            ivx::Context ctx = ovx::getOpenVXContext();
             if ((vx_size)(ksize.width) > ctx.convolutionMaxDimension() || (vx_size)(ksize.height) > ctx.convolutionMaxDimension())
                 return false;
 
@@ -2239,7 +2239,7 @@ static bool openvx_gaussianBlur(InputArray _src, OutputArray _dst, Size ksize,
 
     try
     {
-        ivx::Context ctx = ivx::Context::create();
+        ivx::Context ctx = ovx::getOpenVXContext();
         if ((vx_size)(ksize.width) > ctx.convolutionMaxDimension() || (vx_size)(ksize.height) > ctx.convolutionMaxDimension())
             return false;
 
@@ -3361,7 +3361,7 @@ namespace cv
 
         try
         {
-            ivx::Context ctx = ivx::Context::create();
+            ivx::Context ctx = ovx::getOpenVXContext();
 #ifdef VX_VERSION_1_1
             if ((vx_size)ksize > ctx.nonlinearMaxDimension())
                 return false;

modules/imgproc/src/thresh.cpp

@@ -1301,7 +1301,7 @@ static bool openvx_threshold(Mat src, Mat dst, int thresh, int maxval, int type)
 
     try
     {
-        ivx::Context ctx = ivx::Context::create();
+        ivx::Context ctx = ovx::getOpenVXContext();
 
         ivx::Threshold thh = ivx::Threshold::createBinary(ctx, VX_TYPE_UINT8, thresh);
         thh.setValueTrue(trueVal);

modules/video/src/lkpyramid.cpp

@@ -1099,7 +1099,7 @@ namespace
 
         try
         {
-            Context context = Context::create();
+            Context context = ovx::getOpenVXContext();
 
             if(context.vendorID() == VX_ID_KHRONOS)
             {