class ParallelLoopBodyWrapperContext
    {
    public:
        ParallelLoopBodyWrapperContext(const cv::ParallelLoopBody& _body, const cv::Range& _r, double _nstripes) :
            is_rng_used(false), hasException(false)
        {

            body = &_body;
            wholeRange = _r;
            double len = wholeRange.end - wholeRange.start;
            nstripes = cvRound(_nstripes <= 0 ? len : MIN(MAX(_nstripes, 1.), len));

            // propagate main thread state
            rng = cv::theRNG();

#ifdef OPENCV_TRACE
            traceRootRegion = CV_TRACE_NS::details::getCurrentRegion();
            traceRootContext = CV_TRACE_NS::details::getTraceManager().tls.get();
#endif

#ifdef ENABLE_INSTRUMENTATION
            pThreadRoot = cv::instr::getInstrumentTLSStruct().pCurrentNode;
#endif
        }
        void finalize()
        {
#ifdef ENABLE_INSTRUMENTATION
            for(size_t i = 0; i < pThreadRoot->m_childs.size(); i++)
                SyncNodes(pThreadRoot->m_childs[i]);
#endif
            if (is_rng_used)
            {
                // Some parallel backends execute nested jobs in the main thread,
                // so we need to restore initial RNG state here.
                cv::theRNG() = rng;
                // We can't properly update RNG state based on RNG usage in worker threads,
                // so lets just change main thread RNG state to the next value.
                // Note: this behaviour is not equal to single-threaded mode.
                cv::theRNG().next();
            }
#ifdef OPENCV_TRACE
            if (traceRootRegion)
                CV_TRACE_NS::details::parallelForFinalize(*traceRootRegion);
#endif

            if (hasException)
            {
#if CV__EXCEPTION_PTR
                std::rethrow_exception(pException);
#else
                CV_Error(Error::StsError, "Exception in parallel_for() body: " + exception_message);
#endif
            }
        }
        ~ParallelLoopBodyWrapperContext() {}

        const cv::ParallelLoopBody* body;
        cv::Range wholeRange;
        int nstripes;
        cv::RNG rng;
        mutable bool is_rng_used;
#ifdef OPENCV_TRACE
        CV_TRACE_NS::details::Region* traceRootRegion;
        CV_TRACE_NS::details::TraceManagerThreadLocal* traceRootContext;
#endif
#ifdef ENABLE_INSTRUMENTATION
        cv::instr::InstrNode *pThreadRoot;
#endif
        bool hasException;
#if CV__EXCEPTION_PTR
        std::exception_ptr pException;
#else
        cv::String exception_message;
#endif
#if CV__EXCEPTION_PTR
        void recordException()
#else
        void recordException(const cv::String& msg)
#endif
        {
            if (!hasException)
            {
                cv::AutoLock lock(cv::getInitializationMutex());
                if (!hasException)
                {
                    hasException = true;
#if CV__EXCEPTION_PTR
                    pException = std::current_exception();
#else
                    exception_message = msg;
#endif
                }
            }
        }
    private:
        ParallelLoopBodyWrapperContext(const ParallelLoopBodyWrapperContext&); // disabled
        ParallelLoopBodyWrapperContext& operator=(const ParallelLoopBodyWrapperContext&); // disabled
    };


    class ParallelLoopBodyWrapper : public cv::ParallelLoopBody
    {
    public:
        ParallelLoopBodyWrapper(ParallelLoopBodyWrapperContext& ctx_) :
            ctx(ctx_)
        {
        }
        ~ParallelLoopBodyWrapper()
        {
        }
        void operator()(const cv::Range& sr) const CV_OVERRIDE
        {
#ifdef OPENCV_TRACE
            // TODO CV_TRACE_NS::details::setCurrentRegion(rootRegion);
            if (ctx.traceRootRegion && ctx.traceRootContext)
                CV_TRACE_NS::details::parallelForSetRootRegion(*ctx.traceRootRegion, *ctx.traceRootContext);
            CV__TRACE_OPENCV_FUNCTION_NAME("parallel_for_body");
            if (ctx.traceRootRegion)
                CV_TRACE_NS::details::parallelForAttachNestedRegion(*ctx.traceRootRegion);
#endif

#ifdef ENABLE_INSTRUMENTATION
            {
                cv::instr::InstrTLSStruct *pInstrTLS = &cv::instr::getInstrumentTLSStruct();
                pInstrTLS->pCurrentNode = ctx.pThreadRoot; // Initialize TLS node for thread
            }
            CV_INSTRUMENT_REGION()
#endif

            // propagate main thread state
            cv::theRNG() = ctx.rng;

            cv::Range r;
            cv::Range wholeRange = ctx.wholeRange;
            int nstripes = ctx.nstripes;
            r.start = (int)(wholeRange.start +
                            ((uint64)sr.start*(wholeRange.end - wholeRange.start) + nstripes/2)/nstripes);
            r.end = sr.end >= nstripes ? wholeRange.end : (int)(wholeRange.start +
                            ((uint64)sr.end*(wholeRange.end - wholeRange.start) + nstripes/2)/nstripes);

#ifdef OPENCV_TRACE
            CV_TRACE_ARG_VALUE(range_start, "range.start", (int64)r.start);
            CV_TRACE_ARG_VALUE(range_end, "range.end", (int64)r.end);
#endif

            try
            {
                (*ctx.body)(r);
            }
#if CV__EXCEPTION_PTR
            catch (...)
            {
                ctx.recordException();
            }
#else
            catch (const cv::Exception& e)
            {
                ctx.recordException(e.what());
            }
            catch (const std::exception& e)
            {
                ctx.recordException(e.what());
            }
            catch (...)
            {
                ctx.recordException("Unknown exception");
            }
#endif

            if (!ctx.is_rng_used && !(cv::theRNG() == ctx.rng))
                ctx.is_rng_used = true;
        }
        cv::Range stripeRange() const { return cv::Range(0, ctx.nstripes); }

    protected:
        ParallelLoopBodyWrapperContext& ctx;
    };