Some stream fixes

src/cuda.h

@@ -26,6 +26,7 @@ int *cuda_make_int_array(size_t n);
 void cuda_push_array(float *x_gpu, float *x, size_t n);
 void cuda_pull_array(float *x_gpu, float *x, size_t n);
 void cuda_set_device(int n);
+int cuda_get_device();
 void cuda_free(float *x_gpu);
 void cuda_random(float *x_gpu, size_t n);
 float cuda_compare(float *x_gpu, float *x, size_t n, char *s);

src/network_kernels.cu

@@ -51,6 +51,7 @@ void forward_network_gpu(network net, network_state state)
             fill_ongpu(l.outputs * l.batch, 0, l.delta_gpu, 1);
         }
         l.forward_gpu(l, state);
+		cudaStreamSynchronize(get_cuda_stream());
         state.input = l.output_gpu;
     }
 }
@@ -392,6 +393,7 @@ float *get_network_output_gpu(network net)
 
 float *network_predict_gpu(network net, float *input)
 {
+	if (net.gpu_index != cuda_get_device())
 		cuda_set_device(net.gpu_index);
     int size = get_network_input_size(net) * net.batch;
     network_state state;

src/yolo_console_dll.cpp

@@ -158,7 +158,6 @@ int main(int argc, char *argv[])
 						det_image = detector.mat_to_image_resize(cur_frame);
 						result_vec = thread_result_vec;
 						result_vec = detector.tracking(result_vec);	// comment it - if track_id is not required
-
 #ifdef TRACK_OPTFLOW
 						// track optical flow
 						if (track_optflow_queue.size() > 0) {
@@ -189,7 +188,7 @@ int main(int argc, char *argv[])
 								//std::vector<bbox_t> result;
 								auto result = detector.detect_resized(*current_image, frame_size, 0.24, false);	// true
 								//Sleep(200);
-								Sleep(50);
+								//Sleep(50);
 								++fps_det_counter;
 								std::unique_lock<std::mutex> lock(mtx);
 								thread_result_vec = result;

src/yolo_v2_class.cpp

@@ -34,7 +34,7 @@ struct detector_gpu_t{
 };
 
 
-YOLODLL_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id)
+YOLODLL_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id) : cur_gpu_id(gpu_id)
 {
 	int old_gpu_index;
 #ifdef GPU
@@ -178,6 +178,7 @@ YOLODLL_API std::vector<bbox_t> Detector::detect(image_t img, float thresh, bool
 	int old_gpu_index;
 #ifdef GPU
 	cudaGetDevice(&old_gpu_index);
+	if(cur_gpu_id != old_gpu_index)
 		cudaSetDevice(net.gpu_index);
 #endif
 	//std::cout << "net.gpu_index = " << net.gpu_index << std::endl;
@@ -242,6 +243,7 @@ YOLODLL_API std::vector<bbox_t> Detector::detect(image_t img, float thresh, bool
 		free(sized.data);
 
 #ifdef GPU
+	if (cur_gpu_id != old_gpu_index)
 		cudaSetDevice(old_gpu_index);
 #endif
 

src/yolo_v2_class.hpp

@@ -47,6 +47,7 @@ struct image_t {
 class Detector {
 	std::shared_ptr<void> detector_gpu_ptr;
 	std::deque<std::vector<bbox_t>> prev_bbox_vec_deque;
+	const int cur_gpu_id;
 public:
 	float nms = .4;
 
@@ -170,8 +171,8 @@ public:
 
 		sync_PyrLKOpticalFlow_gpu = cv::cuda::SparsePyrLKOpticalFlow::create();
 		sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(21, 21));	// 15, 21, 31
-		sync_PyrLKOpticalFlow_gpu->setMaxLevel(5);		// +- 50 ptx
-		sync_PyrLKOpticalFlow_gpu->setNumIters(2000);	// def: 30
+		sync_PyrLKOpticalFlow_gpu->setMaxLevel(3);		// +- 5 ptx
+		sync_PyrLKOpticalFlow_gpu->setNumIters(1000);	// def: 30
 
 		cv::cuda::setDevice(old_gpu_id);
 	}
@@ -190,10 +191,9 @@ public:
 	void update_tracking_flow(cv::Mat src_mat)
 	{
 		int const old_gpu_id = cv::cuda::getDevice();
+		if (old_gpu_id != gpu_id)
 			cv::cuda::setDevice(gpu_id);
 
-		//cv::cuda::Stream stream;
-
 		if (src_mat.channels() == 3) {
 			if (src_mat_gpu.cols == 0) {
 				src_mat_gpu = cv::cuda::GpuMat(src_mat.size(), src_mat.type());
@@ -203,6 +203,7 @@ public:
 			src_mat_gpu.upload(src_mat, stream);
 			cv::cuda::cvtColor(src_mat_gpu, src_grey_gpu, CV_BGR2GRAY, 0, stream);
 		}
+		if (old_gpu_id != gpu_id)
 			cv::cuda::setDevice(old_gpu_id);
 	}
 
@@ -215,19 +216,18 @@ public:
 		}
 
 		int const old_gpu_id = cv::cuda::getDevice();
+		if(old_gpu_id != gpu_id)
 			cv::cuda::setDevice(gpu_id);
 
-		//cv::cuda::Stream stream;
-
 		if (dst_mat_gpu.cols == 0) {
 			dst_mat_gpu = cv::cuda::GpuMat(dst_mat.size(), dst_mat.type());
 			dst_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1);
 			tmp_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1);
 		}
 
-
 		dst_mat_gpu.upload(dst_mat, stream);
 
+
 		cv::cuda::cvtColor(dst_mat_gpu, dst_grey_gpu, CV_BGR2GRAY, 0, stream);
 
 		if (src_grey_gpu.rows != dst_grey_gpu.rows || src_grey_gpu.cols != dst_grey_gpu.cols) {
@@ -237,6 +237,8 @@ public:
 			return cur_bbox_vec;
 		}
 
+		//return cur_bbox_vec;
+
 		cv::Mat prev_pts, prev_pts_flow_cpu, cur_pts_flow_cpu;
 
 		for (auto &i : cur_bbox_vec) {
@@ -298,6 +300,7 @@ public:
 					}
 		}
 
+		if (old_gpu_id != gpu_id)
 			cv::cuda::setDevice(old_gpu_id);
 
 		return result_bbox_vec;