Multithreading template matching OpenCV

TestMatchThreads.cpp

#include <opencv2/opencv.hpp>
#include <chrono>
using namespace std;
typedef std::chrono::high_resolution_clock Clock;

void threadCallback(GpuMat &imS, GpuMat &imT, int count)
{
	std::cout << "Start Thread = " << std::this_thread::get_id() << std::endl;

	GpuMat imR = GpuMat(imS.size().height - imT.size().height + 1, 
		imS.size().width - imT.size().width + 1, CV_8UC1);
	Ptr<TemplateMatching> pMatcher = cuda::createTemplateMatching(CV_8U, CV_TM_CCORR_NORMED);

	cuda::Stream stream;
	for (int j = 0; j < count; j++) {
		pMatcher->match(imS, imT, imR, stream);
	}
	stream.waitForCompletion();

	std::cout << "Finished Thread = " << std::this_thread::get_id() << std::endl;
}

void TestMatchThreads()
{
	// Does threading improve perf for many matches?
	Mat im = cv::imread("D:/data/jayb1/000000/0/20190219223436749_jayb1_000000_0_16_0.tif");
	   
	GpuMat imT = GpuMat(100, 100, CV_8UC1);
	GpuMat imS = GpuMat(3840, 3840, CV_8UC1);
	GpuMat imR = GpuMat(3840, 3840, CV_8UC1);

	imS.upload(im);
	Rect rc = cv::Rect(cv::Point(0, 0), imT.size());
	Mat temp = im(rc);
	imT.upload(temp);

	std::vector< std::thread> threads;

	int N_THREADS = 3;
	int count = 10;

	auto t1 = Clock::now();

	for (int j = 0; j < N_THREADS; j++) {
		std::thread threadObj(threadCallback, imS, imT, count);
		threads.push_back(std::move(threadObj));
	}
	for (int j = 0; j < N_THREADS; j++) {
		threads[j].join();
	}

	auto t2 = Clock::now();
	std::cout << "elapsed: "
		<< std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1).count()
		<< " mS" << std::endl;
}

Above code gives the error:
OpenCV Error: Assertion failed (allocSize == allocations.back()) inanonymous-namespace'::MemoryStack::returnMemory, file C:\Users\derrickb\Documents\opencv\modules\core\src\cuda_stream.cpp, line 99
`
with 3 or more threads.